방재호
commited on
Commit
·
cd625c2
1
Parent(s):
da6034f
init
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- .DS_Store +0 -0
- CHANGELOG.md +507 -0
- CITATION.cff +7 -0
- CODEOWNERS +12 -0
- LICENSE.txt +663 -0
- README.md +177 -3
- __pycache__/launch.cpython-310.pyc +0 -0
- __pycache__/webui.cpython-310.pyc +0 -0
- cache.json +8 -0
- configs/alt-diffusion-inference.yaml +72 -0
- configs/instruct-pix2pix.yaml +98 -0
- configs/v1-inference.yaml +70 -0
- configs/v1-inpainting-inference.yaml +70 -0
- embeddings/Place Textual Inversion embeddings here.txt +0 -0
- environment-wsl2.yaml +11 -0
- extensions-builtin/.DS_Store +0 -0
- extensions-builtin/LDSR/__pycache__/ldsr_model_arch.cpython-310.pyc +0 -0
- extensions-builtin/LDSR/__pycache__/preload.cpython-310.pyc +0 -0
- extensions-builtin/LDSR/__pycache__/sd_hijack_autoencoder.cpython-310.pyc +0 -0
- extensions-builtin/LDSR/__pycache__/sd_hijack_ddpm_v1.cpython-310.pyc +0 -0
- extensions-builtin/LDSR/__pycache__/vqvae_quantize.cpython-310.pyc +0 -0
- extensions-builtin/LDSR/ldsr_model_arch.py +250 -0
- extensions-builtin/LDSR/preload.py +6 -0
- extensions-builtin/LDSR/scripts/__pycache__/ldsr_model.cpython-310.pyc +0 -0
- extensions-builtin/LDSR/scripts/ldsr_model.py +68 -0
- extensions-builtin/LDSR/sd_hijack_autoencoder.py +293 -0
- extensions-builtin/LDSR/sd_hijack_ddpm_v1.py +1443 -0
- extensions-builtin/LDSR/vqvae_quantize.py +147 -0
- extensions-builtin/Lora/__pycache__/extra_networks_lora.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/lora.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/lyco_helpers.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/network.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/network_full.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/network_hada.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/network_ia3.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/network_lokr.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/network_lora.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/networks.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/preload.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/ui_edit_user_metadata.cpython-310.pyc +0 -0
- extensions-builtin/Lora/__pycache__/ui_extra_networks_lora.cpython-310.pyc +0 -0
- extensions-builtin/Lora/extra_networks_lora.py +67 -0
- extensions-builtin/Lora/lora.py +9 -0
- extensions-builtin/Lora/lora_patches.py +31 -0
- extensions-builtin/Lora/lyco_helpers.py +21 -0
- extensions-builtin/Lora/network.py +158 -0
- extensions-builtin/Lora/network_full.py +27 -0
- extensions-builtin/Lora/network_hada.py +55 -0
- extensions-builtin/Lora/network_ia3.py +30 -0
- extensions-builtin/Lora/network_lokr.py +64 -0
.DS_Store
ADDED
Binary file (12.3 kB). View file
|
|
CHANGELOG.md
ADDED
@@ -0,0 +1,507 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## 1.6.0
|
2 |
+
|
3 |
+
### Features:
|
4 |
+
* refiner support [#12371](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12371)
|
5 |
+
* add NV option for Random number generator source setting, which allows to generate same pictures on CPU/AMD/Mac as on NVidia videocards
|
6 |
+
* add style editor dialog
|
7 |
+
* hires fix: add an option to use a different checkpoint for second pass ([#12181](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12181))
|
8 |
+
* option to keep multiple loaded models in memory ([#12227](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12227))
|
9 |
+
* new samplers: Restart, DPM++ 2M SDE Exponential, DPM++ 2M SDE Heun, DPM++ 2M SDE Heun Karras, DPM++ 2M SDE Heun Exponential, DPM++ 3M SDE, DPM++ 3M SDE Karras, DPM++ 3M SDE Exponential ([#12300](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12300), [#12519](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12519), [#12542](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12542))
|
10 |
+
* rework DDIM, PLMS, UniPC to use CFG denoiser same as in k-diffusion samplers:
|
11 |
+
* makes all of them work with img2img
|
12 |
+
* makes prompt composition posssible (AND)
|
13 |
+
* makes them available for SDXL
|
14 |
+
* always show extra networks tabs in the UI ([#11808](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11808))
|
15 |
+
* use less RAM when creating models ([#11958](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11958), [#12599](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12599))
|
16 |
+
* textual inversion inference support for SDXL
|
17 |
+
* extra networks UI: show metadata for SD checkpoints
|
18 |
+
* checkpoint merger: add metadata support
|
19 |
+
* prompt editing and attention: add support for whitespace after the number ([ red : green : 0.5 ]) (seed breaking change) ([#12177](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12177))
|
20 |
+
* VAE: allow selecting own VAE for each checkpoint (in user metadata editor)
|
21 |
+
* VAE: add selected VAE to infotext
|
22 |
+
* options in main UI: add own separate setting for txt2img and img2img, correctly read values from pasted infotext, add setting for column count ([#12551](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12551))
|
23 |
+
* add resize handle to txt2img and img2img tabs, allowing to change the amount of horizontable space given to generation parameters and resulting image gallery ([#12687](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12687), [#12723](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12723))
|
24 |
+
* change default behavior for batching cond/uncond -- now it's on by default, and is disabled by an UI setting (Optimizatios -> Batch cond/uncond) - if you are on lowvram/medvram and are getting OOM exceptions, you will need to enable it
|
25 |
+
* show current position in queue and make it so that requests are processed in the order of arrival ([#12707](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12707))
|
26 |
+
* add `--medvram-sdxl` flag that only enables `--medvram` for SDXL models
|
27 |
+
* prompt editing timeline has separate range for first pass and hires-fix pass (seed breaking change) ([#12457](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12457))
|
28 |
+
|
29 |
+
### Minor:
|
30 |
+
* img2img batch: RAM savings, VRAM savings, .tif, .tiff in img2img batch ([#12120](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12120), [#12514](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12514), [#12515](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12515))
|
31 |
+
* postprocessing/extras: RAM savings ([#12479](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12479))
|
32 |
+
* XYZ: in the axis labels, remove pathnames from model filenames
|
33 |
+
* XYZ: support hires sampler ([#12298](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12298))
|
34 |
+
* XYZ: new option: use text inputs instead of dropdowns ([#12491](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12491))
|
35 |
+
* add gradio version warning
|
36 |
+
* sort list of VAE checkpoints ([#12297](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12297))
|
37 |
+
* use transparent white for mask in inpainting, along with an option to select the color ([#12326](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12326))
|
38 |
+
* move some settings to their own section: img2img, VAE
|
39 |
+
* add checkbox to show/hide dirs for extra networks
|
40 |
+
* Add TAESD(or more) options for all the VAE encode/decode operation ([#12311](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12311))
|
41 |
+
* gradio theme cache, new gradio themes, along with explanation that the user can input his own values ([#12346](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12346), [#12355](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12355))
|
42 |
+
* sampler fixes/tweaks: s_tmax, s_churn, s_noise, s_tmax ([#12354](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12354), [#12356](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12356), [#12357](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12357), [#12358](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12358), [#12375](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12375), [#12521](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12521))
|
43 |
+
* update README.md with correct instructions for Linux installation ([#12352](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12352))
|
44 |
+
* option to not save incomplete images, on by default ([#12338](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12338))
|
45 |
+
* enable cond cache by default
|
46 |
+
* git autofix for repos that are corrupted ([#12230](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12230))
|
47 |
+
* allow to open images in new browser tab by middle mouse button ([#12379](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12379))
|
48 |
+
* automatically open webui in browser when running "locally" ([#12254](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12254))
|
49 |
+
* put commonly used samplers on top, make DPM++ 2M Karras the default choice
|
50 |
+
* zoom and pan: option to auto-expand a wide image, improved integration ([#12413](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12413), [#12727](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12727))
|
51 |
+
* option to cache Lora networks in memory
|
52 |
+
* rework hires fix UI to use accordion
|
53 |
+
* face restoration and tiling moved to settings - use "Options in main UI" setting if you want them back
|
54 |
+
* change quicksettings items to have variable width
|
55 |
+
* Lora: add Norm module, add support for bias ([#12503](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12503))
|
56 |
+
* Lora: output warnings in UI rather than fail for unfitting loras; switch to logging for error output in console
|
57 |
+
* support search and display of hashes for all extra network items ([#12510](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12510))
|
58 |
+
* add extra noise param for img2img operations ([#12564](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12564))
|
59 |
+
* support for Lora with bias ([#12584](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12584))
|
60 |
+
* make interrupt quicker ([#12634](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12634))
|
61 |
+
* configurable gallery height ([#12648](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12648))
|
62 |
+
* make results column sticky ([#12645](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12645))
|
63 |
+
* more hash filename patterns ([#12639](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12639))
|
64 |
+
* make image viewer actually fit the whole page ([#12635](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12635))
|
65 |
+
* make progress bar work independently from live preview display which results in it being updated a lot more often
|
66 |
+
* forbid Full live preview method for medvram and add a setting to undo the forbidding
|
67 |
+
* make it possible to localize tooltips and placeholders
|
68 |
+
* add option to align with sgm repo's sampling implementation ([#12818](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12818))
|
69 |
+
* Restore faces and Tiling generation parameters have been moved to settings out of main UI
|
70 |
+
* if you want to put them back into main UI, use `Options in main UI` setting on the UI page.
|
71 |
+
|
72 |
+
### Extensions and API:
|
73 |
+
* gradio 3.41.2
|
74 |
+
* also bump versions for packages: transformers, GitPython, accelerate, scikit-image, timm, tomesd
|
75 |
+
* support tooltip kwarg for gradio elements: gr.Textbox(label='hello', tooltip='world')
|
76 |
+
* properly clear the total console progressbar when using txt2img and img2img from API
|
77 |
+
* add cmd_arg --disable-extra-extensions and --disable-all-extensions ([#12294](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12294))
|
78 |
+
* shared.py and webui.py split into many files
|
79 |
+
* add --loglevel commandline argument for logging
|
80 |
+
* add a custom UI element that combines accordion and checkbox
|
81 |
+
* avoid importing gradio in tests because it spams warnings
|
82 |
+
* put infotext label for setting into OptionInfo definition rather than in a separate list
|
83 |
+
* make `StableDiffusionProcessingImg2Img.mask_blur` a property, make more inline with PIL `GaussianBlur` ([#12470](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12470))
|
84 |
+
* option to make scripts UI without gr.Group
|
85 |
+
* add a way for scripts to register a callback for before/after just a single component's creation
|
86 |
+
* use dataclass for StableDiffusionProcessing
|
87 |
+
* store patches for Lora in a specialized module instead of inside torch
|
88 |
+
* support http/https URLs in API ([#12663](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12663), [#12698](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12698))
|
89 |
+
* add extra noise callback ([#12616](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12616))
|
90 |
+
* dump current stack traces when exiting with SIGINT
|
91 |
+
* add type annotations for extra fields of shared.sd_model
|
92 |
+
|
93 |
+
### Bug Fixes:
|
94 |
+
* Don't crash if out of local storage quota for javascriot localStorage
|
95 |
+
* XYZ plot do not fail if an exception occurs
|
96 |
+
* fix missing TI hash in infotext if generation uses both negative and positive TI ([#12269](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12269))
|
97 |
+
* localization fixes ([#12307](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12307))
|
98 |
+
* fix sdxl model invalid configuration after the hijack
|
99 |
+
* correctly toggle extras checkbox for infotext paste ([#12304](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12304))
|
100 |
+
* open raw sysinfo link in new page ([#12318](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12318))
|
101 |
+
* prompt parser: Account for empty field in alternating words syntax ([#12319](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12319))
|
102 |
+
* add tab and carriage return to invalid filename chars ([#12327](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12327))
|
103 |
+
* fix api only Lora not working ([#12387](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12387))
|
104 |
+
* fix options in main UI misbehaving when there's just one element
|
105 |
+
* make it possible to use a sampler from infotext even if it's hidden in the dropdown
|
106 |
+
* fix styles missing from the prompt in infotext when making a grid of batch of multiplie images
|
107 |
+
* prevent bogus progress output in console when calculating hires fix dimensions
|
108 |
+
* fix --use-textbox-seed
|
109 |
+
* fix broken `Lora/Networks: use old method` option ([#12466](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12466))
|
110 |
+
* properly return `None` for VAE hash when using `--no-hashing` ([#12463](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12463))
|
111 |
+
* MPS/macOS fixes and optimizations ([#12526](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12526))
|
112 |
+
* add second_order to samplers that mistakenly didn't have it
|
113 |
+
* when refreshing cards in extra networks UI, do not discard user's custom resolution
|
114 |
+
* fix processing error that happens if batch_size is not a multiple of how many prompts/negative prompts there are ([#12509](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12509))
|
115 |
+
* fix inpaint upload for alpha masks ([#12588](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12588))
|
116 |
+
* fix exception when image sizes are not integers ([#12586](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12586))
|
117 |
+
* fix incorrect TAESD Latent scale ([#12596](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12596))
|
118 |
+
* auto add data-dir to gradio-allowed-path ([#12603](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12603))
|
119 |
+
* fix exception if extensuions dir is missing ([#12607](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12607))
|
120 |
+
* fix issues with api model-refresh and vae-refresh ([#12638](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12638))
|
121 |
+
* fix img2img background color for transparent images option not being used ([#12633](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12633))
|
122 |
+
* attempt to resolve NaN issue with unstable VAEs in fp32 mk2 ([#12630](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12630))
|
123 |
+
* implement missing undo hijack for SDXL
|
124 |
+
* fix xyz swap axes ([#12684](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12684))
|
125 |
+
* fix errors in backup/restore tab if any of config files are broken ([#12689](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12689))
|
126 |
+
* fix SD VAE switch error after model reuse ([#12685](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12685))
|
127 |
+
* fix trying to create images too large for the chosen format ([#12667](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12667))
|
128 |
+
* create Gradio temp directory if necessary ([#12717](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12717))
|
129 |
+
* prevent possible cache loss if exiting as it's being written by using an atomic operation to replace the cache with the new version
|
130 |
+
* set devices.dtype_unet correctly
|
131 |
+
* run RealESRGAN on GPU for non-CUDA devices ([#12737](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
|
132 |
+
* prevent extra network buttons being obscured by description for very small card sizes ([#12745](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12745))
|
133 |
+
* fix error that causes some extra networks to be disabled if both <lora:> and <lyco:> are present in the prompt
|
134 |
+
* fix defaults settings page breaking when any of main UI tabs are hidden
|
135 |
+
* fix incorrect save/display of new values in Defaults page in settings
|
136 |
+
* fix for Reload UI function: if you reload UI on one tab, other opened tabs will no longer stop working
|
137 |
+
* fix an error that prevents VAE being reloaded after an option change if a VAE near the checkpoint exists ([#12797](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
|
138 |
+
* hide broken image crop tool ([#12792](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
|
139 |
+
* don't show hidden samplers in dropdown for XYZ script ([#12780](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
|
140 |
+
* fix style editing dialog breaking if it's opened in both img2img and txt2img tabs
|
141 |
+
* fix a bug allowing users to bypass gradio and API authentication (reported by vysecurity)
|
142 |
+
* fix notification not playing when built-in webui tab is inactive ([#12834](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12834))
|
143 |
+
* honor `--skip-install` for extension installers ([#12832](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12832))
|
144 |
+
* don't print blank stdout in extension installers ([#12833](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12832), [#12855](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12855))
|
145 |
+
* do not change quicksettings dropdown option when value returned is `None` ([#12854](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12854))
|
146 |
+
* get progressbar to display correctly in extensions tab
|
147 |
+
|
148 |
+
|
149 |
+
## 1.5.2
|
150 |
+
|
151 |
+
### Bug Fixes:
|
152 |
+
* fix memory leak when generation fails
|
153 |
+
* update doggettx cross attention optimization to not use an unreasonable amount of memory in some edge cases -- suggestion by MorkTheOrk
|
154 |
+
|
155 |
+
|
156 |
+
## 1.5.1
|
157 |
+
|
158 |
+
### Minor:
|
159 |
+
* support parsing text encoder blocks in some new LoRAs
|
160 |
+
* delete scale checker script due to user demand
|
161 |
+
|
162 |
+
### Extensions and API:
|
163 |
+
* add postprocess_batch_list script callback
|
164 |
+
|
165 |
+
### Bug Fixes:
|
166 |
+
* fix TI training for SD1
|
167 |
+
* fix reload altclip model error
|
168 |
+
* prepend the pythonpath instead of overriding it
|
169 |
+
* fix typo in SD_WEBUI_RESTARTING
|
170 |
+
* if txt2img/img2img raises an exception, finally call state.end()
|
171 |
+
* fix composable diffusion weight parsing
|
172 |
+
* restyle Startup profile for black users
|
173 |
+
* fix webui not launching with --nowebui
|
174 |
+
* catch exception for non git extensions
|
175 |
+
* fix some options missing from /sdapi/v1/options
|
176 |
+
* fix for extension update status always saying "unknown"
|
177 |
+
* fix display of extra network cards that have `<>` in the name
|
178 |
+
* update lora extension to work with python 3.8
|
179 |
+
|
180 |
+
|
181 |
+
## 1.5.0
|
182 |
+
|
183 |
+
### Features:
|
184 |
+
* SD XL support
|
185 |
+
* user metadata system for custom networks
|
186 |
+
* extended Lora metadata editor: set activation text, default weight, view tags, training info
|
187 |
+
* Lora extension rework to include other types of networks (all that were previously handled by LyCORIS extension)
|
188 |
+
* show github stars for extenstions
|
189 |
+
* img2img batch mode can read extra stuff from png info
|
190 |
+
* img2img batch works with subdirectories
|
191 |
+
* hotkeys to move prompt elements: alt+left/right
|
192 |
+
* restyle time taken/VRAM display
|
193 |
+
* add textual inversion hashes to infotext
|
194 |
+
* optimization: cache git extension repo information
|
195 |
+
* move generate button next to the generated picture for mobile clients
|
196 |
+
* hide cards for networks of incompatible Stable Diffusion version in Lora extra networks interface
|
197 |
+
* skip installing packages with pip if they all are already installed - startup speedup of about 2 seconds
|
198 |
+
|
199 |
+
### Minor:
|
200 |
+
* checkbox to check/uncheck all extensions in the Installed tab
|
201 |
+
* add gradio user to infotext and to filename patterns
|
202 |
+
* allow gif for extra network previews
|
203 |
+
* add options to change colors in grid
|
204 |
+
* use natural sort for items in extra networks
|
205 |
+
* Mac: use empty_cache() from torch 2 to clear VRAM
|
206 |
+
* added automatic support for installing the right libraries for Navi3 (AMD)
|
207 |
+
* add option SWIN_torch_compile to accelerate SwinIR upscale
|
208 |
+
* suppress printing TI embedding info at start to console by default
|
209 |
+
* speedup extra networks listing
|
210 |
+
* added `[none]` filename token.
|
211 |
+
* removed thumbs extra networks view mode (use settings tab to change width/height/scale to get thumbs)
|
212 |
+
* add always_discard_next_to_last_sigma option to XYZ plot
|
213 |
+
* automatically switch to 32-bit float VAE if the generated picture has NaNs without the need for `--no-half-vae` commandline flag.
|
214 |
+
|
215 |
+
### Extensions and API:
|
216 |
+
* api endpoints: /sdapi/v1/server-kill, /sdapi/v1/server-restart, /sdapi/v1/server-stop
|
217 |
+
* allow Script to have custom metaclass
|
218 |
+
* add model exists status check /sdapi/v1/options
|
219 |
+
* rename --add-stop-route to --api-server-stop
|
220 |
+
* add `before_hr` script callback
|
221 |
+
* add callback `after_extra_networks_activate`
|
222 |
+
* disable rich exception output in console for API by default, use WEBUI_RICH_EXCEPTIONS env var to enable
|
223 |
+
* return http 404 when thumb file not found
|
224 |
+
* allow replacing extensions index with environment variable
|
225 |
+
|
226 |
+
### Bug Fixes:
|
227 |
+
* fix for catch errors when retrieving extension index #11290
|
228 |
+
* fix very slow loading speed of .safetensors files when reading from network drives
|
229 |
+
* API cache cleanup
|
230 |
+
* fix UnicodeEncodeError when writing to file CLIP Interrogator batch mode
|
231 |
+
* fix warning of 'has_mps' deprecated from PyTorch
|
232 |
+
* fix problem with extra network saving images as previews losing generation info
|
233 |
+
* fix throwing exception when trying to resize image with I;16 mode
|
234 |
+
* fix for #11534: canvas zoom and pan extension hijacking shortcut keys
|
235 |
+
* fixed launch script to be runnable from any directory
|
236 |
+
* don't add "Seed Resize: -1x-1" to API image metadata
|
237 |
+
* correctly remove end parenthesis with ctrl+up/down
|
238 |
+
* fixing --subpath on newer gradio version
|
239 |
+
* fix: check fill size none zero when resize (fixes #11425)
|
240 |
+
* use submit and blur for quick settings textbox
|
241 |
+
* save img2img batch with images.save_image()
|
242 |
+
* prevent running preload.py for disabled extensions
|
243 |
+
* fix: previously, model name was added together with directory name to infotext and to [model_name] filename pattern; directory name is now not included
|
244 |
+
|
245 |
+
|
246 |
+
## 1.4.1
|
247 |
+
|
248 |
+
### Bug Fixes:
|
249 |
+
* add queue lock for refresh-checkpoints
|
250 |
+
|
251 |
+
## 1.4.0
|
252 |
+
|
253 |
+
### Features:
|
254 |
+
* zoom controls for inpainting
|
255 |
+
* run basic torch calculation at startup in parallel to reduce the performance impact of first generation
|
256 |
+
* option to pad prompt/neg prompt to be same length
|
257 |
+
* remove taming_transformers dependency
|
258 |
+
* custom k-diffusion scheduler settings
|
259 |
+
* add an option to show selected settings in main txt2img/img2img UI
|
260 |
+
* sysinfo tab in settings
|
261 |
+
* infer styles from prompts when pasting params into the UI
|
262 |
+
* an option to control the behavior of the above
|
263 |
+
|
264 |
+
### Minor:
|
265 |
+
* bump Gradio to 3.32.0
|
266 |
+
* bump xformers to 0.0.20
|
267 |
+
* Add option to disable token counters
|
268 |
+
* tooltip fixes & optimizations
|
269 |
+
* make it possible to configure filename for the zip download
|
270 |
+
* `[vae_filename]` pattern for filenames
|
271 |
+
* Revert discarding penultimate sigma for DPM-Solver++(2M) SDE
|
272 |
+
* change UI reorder setting to multiselect
|
273 |
+
* read version info form CHANGELOG.md if git version info is not available
|
274 |
+
* link footer API to Wiki when API is not active
|
275 |
+
* persistent conds cache (opt-in optimization)
|
276 |
+
|
277 |
+
### Extensions:
|
278 |
+
* After installing extensions, webui properly restarts the process rather than reloads the UI
|
279 |
+
* Added VAE listing to web API. Via: /sdapi/v1/sd-vae
|
280 |
+
* custom unet support
|
281 |
+
* Add onAfterUiUpdate callback
|
282 |
+
* refactor EmbeddingDatabase.register_embedding() to allow unregistering
|
283 |
+
* add before_process callback for scripts
|
284 |
+
* add ability for alwayson scripts to specify section and let user reorder those sections
|
285 |
+
|
286 |
+
### Bug Fixes:
|
287 |
+
* Fix dragging text to prompt
|
288 |
+
* fix incorrect quoting for infotext values with colon in them
|
289 |
+
* fix "hires. fix" prompt sharing same labels with txt2img_prompt
|
290 |
+
* Fix s_min_uncond default type int
|
291 |
+
* Fix for #10643 (Inpainting mask sometimes not working)
|
292 |
+
* fix bad styling for thumbs view in extra networks #10639
|
293 |
+
* fix for empty list of optimizations #10605
|
294 |
+
* small fixes to prepare_tcmalloc for Debian/Ubuntu compatibility
|
295 |
+
* fix --ui-debug-mode exit
|
296 |
+
* patch GitPython to not use leaky persistent processes
|
297 |
+
* fix duplicate Cross attention optimization after UI reload
|
298 |
+
* torch.cuda.is_available() check for SdOptimizationXformers
|
299 |
+
* fix hires fix using wrong conds in second pass if using Loras.
|
300 |
+
* handle exception when parsing generation parameters from png info
|
301 |
+
* fix upcast attention dtype error
|
302 |
+
* forcing Torch Version to 1.13.1 for RX 5000 series GPUs
|
303 |
+
* split mask blur into X and Y components, patch Outpainting MK2 accordingly
|
304 |
+
* don't die when a LoRA is a broken symlink
|
305 |
+
* allow activation of Generate Forever during generation
|
306 |
+
|
307 |
+
|
308 |
+
## 1.3.2
|
309 |
+
|
310 |
+
### Bug Fixes:
|
311 |
+
* fix files served out of tmp directory even if they are saved to disk
|
312 |
+
* fix postprocessing overwriting parameters
|
313 |
+
|
314 |
+
## 1.3.1
|
315 |
+
|
316 |
+
### Features:
|
317 |
+
* revert default cross attention optimization to Doggettx
|
318 |
+
|
319 |
+
### Bug Fixes:
|
320 |
+
* fix bug: LoRA don't apply on dropdown list sd_lora
|
321 |
+
* fix png info always added even if setting is not enabled
|
322 |
+
* fix some fields not applying in xyz plot
|
323 |
+
* fix "hires. fix" prompt sharing same labels with txt2img_prompt
|
324 |
+
* fix lora hashes not being added properly to infotex if there is only one lora
|
325 |
+
* fix --use-cpu failing to work properly at startup
|
326 |
+
* make --disable-opt-split-attention command line option work again
|
327 |
+
|
328 |
+
## 1.3.0
|
329 |
+
|
330 |
+
### Features:
|
331 |
+
* add UI to edit defaults
|
332 |
+
* token merging (via dbolya/tomesd)
|
333 |
+
* settings tab rework: add a lot of additional explanations and links
|
334 |
+
* load extensions' Git metadata in parallel to loading the main program to save a ton of time during startup
|
335 |
+
* update extensions table: show branch, show date in separate column, and show version from tags if available
|
336 |
+
* TAESD - another option for cheap live previews
|
337 |
+
* allow choosing sampler and prompts for second pass of hires fix - hidden by default, enabled in settings
|
338 |
+
* calculate hashes for Lora
|
339 |
+
* add lora hashes to infotext
|
340 |
+
* when pasting infotext, use infotext's lora hashes to find local loras for `<lora:xxx:1>` entries whose hashes match loras the user has
|
341 |
+
* select cross attention optimization from UI
|
342 |
+
|
343 |
+
### Minor:
|
344 |
+
* bump Gradio to 3.31.0
|
345 |
+
* bump PyTorch to 2.0.1 for macOS and Linux AMD
|
346 |
+
* allow setting defaults for elements in extensions' tabs
|
347 |
+
* allow selecting file type for live previews
|
348 |
+
* show "Loading..." for extra networks when displaying for the first time
|
349 |
+
* suppress ENSD infotext for samplers that don't use it
|
350 |
+
* clientside optimizations
|
351 |
+
* add options to show/hide hidden files and dirs in extra networks, and to not list models/files in hidden directories
|
352 |
+
* allow whitespace in styles.csv
|
353 |
+
* add option to reorder tabs
|
354 |
+
* move some functionality (swap resolution and set seed to -1) to client
|
355 |
+
* option to specify editor height for img2img
|
356 |
+
* button to copy image resolution into img2img width/height sliders
|
357 |
+
* switch from pyngrok to ngrok-py
|
358 |
+
* lazy-load images in extra networks UI
|
359 |
+
* set "Navigate image viewer with gamepad" option to false by default, by request
|
360 |
+
* change upscalers to download models into user-specified directory (from commandline args) rather than the default models/<...>
|
361 |
+
* allow hiding buttons in ui-config.json
|
362 |
+
|
363 |
+
### Extensions:
|
364 |
+
* add /sdapi/v1/script-info api
|
365 |
+
* use Ruff to lint Python code
|
366 |
+
* use ESlint to lint Javascript code
|
367 |
+
* add/modify CFG callbacks for Self-Attention Guidance extension
|
368 |
+
* add command and endpoint for graceful server stopping
|
369 |
+
* add some locals (prompts/seeds/etc) from processing function into the Processing class as fields
|
370 |
+
* rework quoting for infotext items that have commas in them to use JSON (should be backwards compatible except for cases where it didn't work previously)
|
371 |
+
* add /sdapi/v1/refresh-loras api checkpoint post request
|
372 |
+
* tests overhaul
|
373 |
+
|
374 |
+
### Bug Fixes:
|
375 |
+
* fix an issue preventing the program from starting if the user specifies a bad Gradio theme
|
376 |
+
* fix broken prompts from file script
|
377 |
+
* fix symlink scanning for extra networks
|
378 |
+
* fix --data-dir ignored when launching via webui-user.bat COMMANDLINE_ARGS
|
379 |
+
* allow web UI to be ran fully offline
|
380 |
+
* fix inability to run with --freeze-settings
|
381 |
+
* fix inability to merge checkpoint without adding metadata
|
382 |
+
* fix extra networks' save preview image not adding infotext for jpeg/webm
|
383 |
+
* remove blinking effect from text in hires fix and scale resolution preview
|
384 |
+
* make links to `http://<...>.git` extensions work in the extension tab
|
385 |
+
* fix bug with webui hanging at startup due to hanging git process
|
386 |
+
|
387 |
+
|
388 |
+
## 1.2.1
|
389 |
+
|
390 |
+
### Features:
|
391 |
+
* add an option to always refer to LoRA by filenames
|
392 |
+
|
393 |
+
### Bug Fixes:
|
394 |
+
* never refer to LoRA by an alias if multiple LoRAs have same alias or the alias is called none
|
395 |
+
* fix upscalers disappearing after the user reloads UI
|
396 |
+
* allow bf16 in safe unpickler (resolves problems with loading some LoRAs)
|
397 |
+
* allow web UI to be ran fully offline
|
398 |
+
* fix localizations not working
|
399 |
+
* fix error for LoRAs: `'LatentDiffusion' object has no attribute 'lora_layer_mapping'`
|
400 |
+
|
401 |
+
## 1.2.0
|
402 |
+
|
403 |
+
### Features:
|
404 |
+
* do not wait for Stable Diffusion model to load at startup
|
405 |
+
* add filename patterns: `[denoising]`
|
406 |
+
* directory hiding for extra networks: dirs starting with `.` will hide their cards on extra network tabs unless specifically searched for
|
407 |
+
* LoRA: for the `<...>` text in prompt, use name of LoRA that is in the metdata of the file, if present, instead of filename (both can be used to activate LoRA)
|
408 |
+
* LoRA: read infotext params from kohya-ss's extension parameters if they are present and if his extension is not active
|
409 |
+
* LoRA: fix some LoRAs not working (ones that have 3x3 convolution layer)
|
410 |
+
* LoRA: add an option to use old method of applying LoRAs (producing same results as with kohya-ss)
|
411 |
+
* add version to infotext, footer and console output when starting
|
412 |
+
* add links to wiki for filename pattern settings
|
413 |
+
* add extended info for quicksettings setting and use multiselect input instead of a text field
|
414 |
+
|
415 |
+
### Minor:
|
416 |
+
* bump Gradio to 3.29.0
|
417 |
+
* bump PyTorch to 2.0.1
|
418 |
+
* `--subpath` option for gradio for use with reverse proxy
|
419 |
+
* Linux/macOS: use existing virtualenv if already active (the VIRTUAL_ENV environment variable)
|
420 |
+
* do not apply localizations if there are none (possible frontend optimization)
|
421 |
+
* add extra `None` option for VAE in XYZ plot
|
422 |
+
* print error to console when batch processing in img2img fails
|
423 |
+
* create HTML for extra network pages only on demand
|
424 |
+
* allow directories starting with `.` to still list their models for LoRA, checkpoints, etc
|
425 |
+
* put infotext options into their own category in settings tab
|
426 |
+
* do not show licenses page when user selects Show all pages in settings
|
427 |
+
|
428 |
+
### Extensions:
|
429 |
+
* tooltip localization support
|
430 |
+
* add API method to get LoRA models with prompt
|
431 |
+
|
432 |
+
### Bug Fixes:
|
433 |
+
* re-add `/docs` endpoint
|
434 |
+
* fix gamepad navigation
|
435 |
+
* make the lightbox fullscreen image function properly
|
436 |
+
* fix squished thumbnails in extras tab
|
437 |
+
* keep "search" filter for extra networks when user refreshes the tab (previously it showed everthing after you refreshed)
|
438 |
+
* fix webui showing the same image if you configure the generation to always save results into same file
|
439 |
+
* fix bug with upscalers not working properly
|
440 |
+
* fix MPS on PyTorch 2.0.1, Intel Macs
|
441 |
+
* make it so that custom context menu from contextMenu.js only disappears after user's click, ignoring non-user click events
|
442 |
+
* prevent Reload UI button/link from reloading the page when it's not yet ready
|
443 |
+
* fix prompts from file script failing to read contents from a drag/drop file
|
444 |
+
|
445 |
+
|
446 |
+
## 1.1.1
|
447 |
+
### Bug Fixes:
|
448 |
+
* fix an error that prevents running webui on PyTorch<2.0 without --disable-safe-unpickle
|
449 |
+
|
450 |
+
## 1.1.0
|
451 |
+
### Features:
|
452 |
+
* switch to PyTorch 2.0.0 (except for AMD GPUs)
|
453 |
+
* visual improvements to custom code scripts
|
454 |
+
* add filename patterns: `[clip_skip]`, `[hasprompt<>]`, `[batch_number]`, `[generation_number]`
|
455 |
+
* add support for saving init images in img2img, and record their hashes in infotext for reproducability
|
456 |
+
* automatically select current word when adjusting weight with ctrl+up/down
|
457 |
+
* add dropdowns for X/Y/Z plot
|
458 |
+
* add setting: Stable Diffusion/Random number generator source: makes it possible to make images generated from a given manual seed consistent across different GPUs
|
459 |
+
* support Gradio's theme API
|
460 |
+
* use TCMalloc on Linux by default; possible fix for memory leaks
|
461 |
+
* add optimization option to remove negative conditioning at low sigma values #9177
|
462 |
+
* embed model merge metadata in .safetensors file
|
463 |
+
* extension settings backup/restore feature #9169
|
464 |
+
* add "resize by" and "resize to" tabs to img2img
|
465 |
+
* add option "keep original size" to textual inversion images preprocess
|
466 |
+
* image viewer scrolling via analog stick
|
467 |
+
* button to restore the progress from session lost / tab reload
|
468 |
+
|
469 |
+
### Minor:
|
470 |
+
* bump Gradio to 3.28.1
|
471 |
+
* change "scale to" to sliders in Extras tab
|
472 |
+
* add labels to tool buttons to make it possible to hide them
|
473 |
+
* add tiled inference support for ScuNET
|
474 |
+
* add branch support for extension installation
|
475 |
+
* change Linux installation script to install into current directory rather than `/home/username`
|
476 |
+
* sort textual inversion embeddings by name (case-insensitive)
|
477 |
+
* allow styles.csv to be symlinked or mounted in docker
|
478 |
+
* remove the "do not add watermark to images" option
|
479 |
+
* make selected tab configurable with UI config
|
480 |
+
* make the extra networks UI fixed height and scrollable
|
481 |
+
* add `disable_tls_verify` arg for use with self-signed certs
|
482 |
+
|
483 |
+
### Extensions:
|
484 |
+
* add reload callback
|
485 |
+
* add `is_hr_pass` field for processing
|
486 |
+
|
487 |
+
### Bug Fixes:
|
488 |
+
* fix broken batch image processing on 'Extras/Batch Process' tab
|
489 |
+
* add "None" option to extra networks dropdowns
|
490 |
+
* fix FileExistsError for CLIP Interrogator
|
491 |
+
* fix /sdapi/v1/txt2img endpoint not working on Linux #9319
|
492 |
+
* fix disappearing live previews and progressbar during slow tasks
|
493 |
+
* fix fullscreen image view not working properly in some cases
|
494 |
+
* prevent alwayson_scripts args param resizing script_arg list when they are inserted in it
|
495 |
+
* fix prompt schedule for second order samplers
|
496 |
+
* fix image mask/composite for weird resolutions #9628
|
497 |
+
* use correct images for previews when using AND (see #9491)
|
498 |
+
* one broken image in img2img batch won't stop all processing
|
499 |
+
* fix image orientation bug in train/preprocess
|
500 |
+
* fix Ngrok recreating tunnels every reload
|
501 |
+
* fix `--realesrgan-models-path` and `--ldsr-models-path` not working
|
502 |
+
* fix `--skip-install` not working
|
503 |
+
* use SAMPLE file format in Outpainting Mk2 & Poorman
|
504 |
+
* do not fail all LoRAs if some have failed to load when making a picture
|
505 |
+
|
506 |
+
## 1.0.0
|
507 |
+
* everything
|
CITATION.cff
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
cff-version: 1.2.0
|
2 |
+
message: "If you use this software, please cite it as below."
|
3 |
+
authors:
|
4 |
+
- given-names: AUTOMATIC1111
|
5 |
+
title: "Stable Diffusion Web UI"
|
6 |
+
date-released: 2022-08-22
|
7 |
+
url: "https://github.com/AUTOMATIC1111/stable-diffusion-webui"
|
CODEOWNERS
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
* @AUTOMATIC1111
|
2 |
+
|
3 |
+
# if you were managing a localization and were removed from this file, this is because
|
4 |
+
# the intended way to do localizations now is via extensions. See:
|
5 |
+
# https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Developing-extensions
|
6 |
+
# Make a repo with your localization and since you are still listed as a collaborator
|
7 |
+
# you can add it to the wiki page yourself. This change is because some people complained
|
8 |
+
# the git commit log is cluttered with things unrelated to almost everyone and
|
9 |
+
# because I believe this is the best overall for the project to handle localizations almost
|
10 |
+
# entirely without my oversight.
|
11 |
+
|
12 |
+
|
LICENSE.txt
ADDED
@@ -0,0 +1,663 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
GNU AFFERO GENERAL PUBLIC LICENSE
|
2 |
+
Version 3, 19 November 2007
|
3 |
+
|
4 |
+
Copyright (c) 2023 AUTOMATIC1111
|
5 |
+
|
6 |
+
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
|
7 |
+
Everyone is permitted to copy and distribute verbatim copies
|
8 |
+
of this license document, but changing it is not allowed.
|
9 |
+
|
10 |
+
Preamble
|
11 |
+
|
12 |
+
The GNU Affero General Public License is a free, copyleft license for
|
13 |
+
software and other kinds of works, specifically designed to ensure
|
14 |
+
cooperation with the community in the case of network server software.
|
15 |
+
|
16 |
+
The licenses for most software and other practical works are designed
|
17 |
+
to take away your freedom to share and change the works. By contrast,
|
18 |
+
our General Public Licenses are intended to guarantee your freedom to
|
19 |
+
share and change all versions of a program--to make sure it remains free
|
20 |
+
software for all its users.
|
21 |
+
|
22 |
+
When we speak of free software, we are referring to freedom, not
|
23 |
+
price. Our General Public Licenses are designed to make sure that you
|
24 |
+
have the freedom to distribute copies of free software (and charge for
|
25 |
+
them if you wish), that you receive source code or can get it if you
|
26 |
+
want it, that you can change the software or use pieces of it in new
|
27 |
+
free programs, and that you know you can do these things.
|
28 |
+
|
29 |
+
Developers that use our General Public Licenses protect your rights
|
30 |
+
with two steps: (1) assert copyright on the software, and (2) offer
|
31 |
+
you this License which gives you legal permission to copy, distribute
|
32 |
+
and/or modify the software.
|
33 |
+
|
34 |
+
A secondary benefit of defending all users' freedom is that
|
35 |
+
improvements made in alternate versions of the program, if they
|
36 |
+
receive widespread use, become available for other developers to
|
37 |
+
incorporate. Many developers of free software are heartened and
|
38 |
+
encouraged by the resulting cooperation. However, in the case of
|
39 |
+
software used on network servers, this result may fail to come about.
|
40 |
+
The GNU General Public License permits making a modified version and
|
41 |
+
letting the public access it on a server without ever releasing its
|
42 |
+
source code to the public.
|
43 |
+
|
44 |
+
The GNU Affero General Public License is designed specifically to
|
45 |
+
ensure that, in such cases, the modified source code becomes available
|
46 |
+
to the community. It requires the operator of a network server to
|
47 |
+
provide the source code of the modified version running there to the
|
48 |
+
users of that server. Therefore, public use of a modified version, on
|
49 |
+
a publicly accessible server, gives the public access to the source
|
50 |
+
code of the modified version.
|
51 |
+
|
52 |
+
An older license, called the Affero General Public License and
|
53 |
+
published by Affero, was designed to accomplish similar goals. This is
|
54 |
+
a different license, not a version of the Affero GPL, but Affero has
|
55 |
+
released a new version of the Affero GPL which permits relicensing under
|
56 |
+
this license.
|
57 |
+
|
58 |
+
The precise terms and conditions for copying, distribution and
|
59 |
+
modification follow.
|
60 |
+
|
61 |
+
TERMS AND CONDITIONS
|
62 |
+
|
63 |
+
0. Definitions.
|
64 |
+
|
65 |
+
"This License" refers to version 3 of the GNU Affero General Public License.
|
66 |
+
|
67 |
+
"Copyright" also means copyright-like laws that apply to other kinds of
|
68 |
+
works, such as semiconductor masks.
|
69 |
+
|
70 |
+
"The Program" refers to any copyrightable work licensed under this
|
71 |
+
License. Each licensee is addressed as "you". "Licensees" and
|
72 |
+
"recipients" may be individuals or organizations.
|
73 |
+
|
74 |
+
To "modify" a work means to copy from or adapt all or part of the work
|
75 |
+
in a fashion requiring copyright permission, other than the making of an
|
76 |
+
exact copy. The resulting work is called a "modified version" of the
|
77 |
+
earlier work or a work "based on" the earlier work.
|
78 |
+
|
79 |
+
A "covered work" means either the unmodified Program or a work based
|
80 |
+
on the Program.
|
81 |
+
|
82 |
+
To "propagate" a work means to do anything with it that, without
|
83 |
+
permission, would make you directly or secondarily liable for
|
84 |
+
infringement under applicable copyright law, except executing it on a
|
85 |
+
computer or modifying a private copy. Propagation includes copying,
|
86 |
+
distribution (with or without modification), making available to the
|
87 |
+
public, and in some countries other activities as well.
|
88 |
+
|
89 |
+
To "convey" a work means any kind of propagation that enables other
|
90 |
+
parties to make or receive copies. Mere interaction with a user through
|
91 |
+
a computer network, with no transfer of a copy, is not conveying.
|
92 |
+
|
93 |
+
An interactive user interface displays "Appropriate Legal Notices"
|
94 |
+
to the extent that it includes a convenient and prominently visible
|
95 |
+
feature that (1) displays an appropriate copyright notice, and (2)
|
96 |
+
tells the user that there is no warranty for the work (except to the
|
97 |
+
extent that warranties are provided), that licensees may convey the
|
98 |
+
work under this License, and how to view a copy of this License. If
|
99 |
+
the interface presents a list of user commands or options, such as a
|
100 |
+
menu, a prominent item in the list meets this criterion.
|
101 |
+
|
102 |
+
1. Source Code.
|
103 |
+
|
104 |
+
The "source code" for a work means the preferred form of the work
|
105 |
+
for making modifications to it. "Object code" means any non-source
|
106 |
+
form of a work.
|
107 |
+
|
108 |
+
A "Standard Interface" means an interface that either is an official
|
109 |
+
standard defined by a recognized standards body, or, in the case of
|
110 |
+
interfaces specified for a particular programming language, one that
|
111 |
+
is widely used among developers working in that language.
|
112 |
+
|
113 |
+
The "System Libraries" of an executable work include anything, other
|
114 |
+
than the work as a whole, that (a) is included in the normal form of
|
115 |
+
packaging a Major Component, but which is not part of that Major
|
116 |
+
Component, and (b) serves only to enable use of the work with that
|
117 |
+
Major Component, or to implement a Standard Interface for which an
|
118 |
+
implementation is available to the public in source code form. A
|
119 |
+
"Major Component", in this context, means a major essential component
|
120 |
+
(kernel, window system, and so on) of the specific operating system
|
121 |
+
(if any) on which the executable work runs, or a compiler used to
|
122 |
+
produce the work, or an object code interpreter used to run it.
|
123 |
+
|
124 |
+
The "Corresponding Source" for a work in object code form means all
|
125 |
+
the source code needed to generate, install, and (for an executable
|
126 |
+
work) run the object code and to modify the work, including scripts to
|
127 |
+
control those activities. However, it does not include the work's
|
128 |
+
System Libraries, or general-purpose tools or generally available free
|
129 |
+
programs which are used unmodified in performing those activities but
|
130 |
+
which are not part of the work. For example, Corresponding Source
|
131 |
+
includes interface definition files associated with source files for
|
132 |
+
the work, and the source code for shared libraries and dynamically
|
133 |
+
linked subprograms that the work is specifically designed to require,
|
134 |
+
such as by intimate data communication or control flow between those
|
135 |
+
subprograms and other parts of the work.
|
136 |
+
|
137 |
+
The Corresponding Source need not include anything that users
|
138 |
+
can regenerate automatically from other parts of the Corresponding
|
139 |
+
Source.
|
140 |
+
|
141 |
+
The Corresponding Source for a work in source code form is that
|
142 |
+
same work.
|
143 |
+
|
144 |
+
2. Basic Permissions.
|
145 |
+
|
146 |
+
All rights granted under this License are granted for the term of
|
147 |
+
copyright on the Program, and are irrevocable provided the stated
|
148 |
+
conditions are met. This License explicitly affirms your unlimited
|
149 |
+
permission to run the unmodified Program. The output from running a
|
150 |
+
covered work is covered by this License only if the output, given its
|
151 |
+
content, constitutes a covered work. This License acknowledges your
|
152 |
+
rights of fair use or other equivalent, as provided by copyright law.
|
153 |
+
|
154 |
+
You may make, run and propagate covered works that you do not
|
155 |
+
convey, without conditions so long as your license otherwise remains
|
156 |
+
in force. You may convey covered works to others for the sole purpose
|
157 |
+
of having them make modifications exclusively for you, or provide you
|
158 |
+
with facilities for running those works, provided that you comply with
|
159 |
+
the terms of this License in conveying all material for which you do
|
160 |
+
not control copyright. Those thus making or running the covered works
|
161 |
+
for you must do so exclusively on your behalf, under your direction
|
162 |
+
and control, on terms that prohibit them from making any copies of
|
163 |
+
your copyrighted material outside their relationship with you.
|
164 |
+
|
165 |
+
Conveying under any other circumstances is permitted solely under
|
166 |
+
the conditions stated below. Sublicensing is not allowed; section 10
|
167 |
+
makes it unnecessary.
|
168 |
+
|
169 |
+
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
|
170 |
+
|
171 |
+
No covered work shall be deemed part of an effective technological
|
172 |
+
measure under any applicable law fulfilling obligations under article
|
173 |
+
11 of the WIPO copyright treaty adopted on 20 December 1996, or
|
174 |
+
similar laws prohibiting or restricting circumvention of such
|
175 |
+
measures.
|
176 |
+
|
177 |
+
When you convey a covered work, you waive any legal power to forbid
|
178 |
+
circumvention of technological measures to the extent such circumvention
|
179 |
+
is effected by exercising rights under this License with respect to
|
180 |
+
the covered work, and you disclaim any intention to limit operation or
|
181 |
+
modification of the work as a means of enforcing, against the work's
|
182 |
+
users, your or third parties' legal rights to forbid circumvention of
|
183 |
+
technological measures.
|
184 |
+
|
185 |
+
4. Conveying Verbatim Copies.
|
186 |
+
|
187 |
+
You may convey verbatim copies of the Program's source code as you
|
188 |
+
receive it, in any medium, provided that you conspicuously and
|
189 |
+
appropriately publish on each copy an appropriate copyright notice;
|
190 |
+
keep intact all notices stating that this License and any
|
191 |
+
non-permissive terms added in accord with section 7 apply to the code;
|
192 |
+
keep intact all notices of the absence of any warranty; and give all
|
193 |
+
recipients a copy of this License along with the Program.
|
194 |
+
|
195 |
+
You may charge any price or no price for each copy that you convey,
|
196 |
+
and you may offer support or warranty protection for a fee.
|
197 |
+
|
198 |
+
5. Conveying Modified Source Versions.
|
199 |
+
|
200 |
+
You may convey a work based on the Program, or the modifications to
|
201 |
+
produce it from the Program, in the form of source code under the
|
202 |
+
terms of section 4, provided that you also meet all of these conditions:
|
203 |
+
|
204 |
+
a) The work must carry prominent notices stating that you modified
|
205 |
+
it, and giving a relevant date.
|
206 |
+
|
207 |
+
b) The work must carry prominent notices stating that it is
|
208 |
+
released under this License and any conditions added under section
|
209 |
+
7. This requirement modifies the requirement in section 4 to
|
210 |
+
"keep intact all notices".
|
211 |
+
|
212 |
+
c) You must license the entire work, as a whole, under this
|
213 |
+
License to anyone who comes into possession of a copy. This
|
214 |
+
License will therefore apply, along with any applicable section 7
|
215 |
+
additional terms, to the whole of the work, and all its parts,
|
216 |
+
regardless of how they are packaged. This License gives no
|
217 |
+
permission to license the work in any other way, but it does not
|
218 |
+
invalidate such permission if you have separately received it.
|
219 |
+
|
220 |
+
d) If the work has interactive user interfaces, each must display
|
221 |
+
Appropriate Legal Notices; however, if the Program has interactive
|
222 |
+
interfaces that do not display Appropriate Legal Notices, your
|
223 |
+
work need not make them do so.
|
224 |
+
|
225 |
+
A compilation of a covered work with other separate and independent
|
226 |
+
works, which are not by their nature extensions of the covered work,
|
227 |
+
and which are not combined with it such as to form a larger program,
|
228 |
+
in or on a volume of a storage or distribution medium, is called an
|
229 |
+
"aggregate" if the compilation and its resulting copyright are not
|
230 |
+
used to limit the access or legal rights of the compilation's users
|
231 |
+
beyond what the individual works permit. Inclusion of a covered work
|
232 |
+
in an aggregate does not cause this License to apply to the other
|
233 |
+
parts of the aggregate.
|
234 |
+
|
235 |
+
6. Conveying Non-Source Forms.
|
236 |
+
|
237 |
+
You may convey a covered work in object code form under the terms
|
238 |
+
of sections 4 and 5, provided that you also convey the
|
239 |
+
machine-readable Corresponding Source under the terms of this License,
|
240 |
+
in one of these ways:
|
241 |
+
|
242 |
+
a) Convey the object code in, or embodied in, a physical product
|
243 |
+
(including a physical distribution medium), accompanied by the
|
244 |
+
Corresponding Source fixed on a durable physical medium
|
245 |
+
customarily used for software interchange.
|
246 |
+
|
247 |
+
b) Convey the object code in, or embodied in, a physical product
|
248 |
+
(including a physical distribution medium), accompanied by a
|
249 |
+
written offer, valid for at least three years and valid for as
|
250 |
+
long as you offer spare parts or customer support for that product
|
251 |
+
model, to give anyone who possesses the object code either (1) a
|
252 |
+
copy of the Corresponding Source for all the software in the
|
253 |
+
product that is covered by this License, on a durable physical
|
254 |
+
medium customarily used for software interchange, for a price no
|
255 |
+
more than your reasonable cost of physically performing this
|
256 |
+
conveying of source, or (2) access to copy the
|
257 |
+
Corresponding Source from a network server at no charge.
|
258 |
+
|
259 |
+
c) Convey individual copies of the object code with a copy of the
|
260 |
+
written offer to provide the Corresponding Source. This
|
261 |
+
alternative is allowed only occasionally and noncommercially, and
|
262 |
+
only if you received the object code with such an offer, in accord
|
263 |
+
with subsection 6b.
|
264 |
+
|
265 |
+
d) Convey the object code by offering access from a designated
|
266 |
+
place (gratis or for a charge), and offer equivalent access to the
|
267 |
+
Corresponding Source in the same way through the same place at no
|
268 |
+
further charge. You need not require recipients to copy the
|
269 |
+
Corresponding Source along with the object code. If the place to
|
270 |
+
copy the object code is a network server, the Corresponding Source
|
271 |
+
may be on a different server (operated by you or a third party)
|
272 |
+
that supports equivalent copying facilities, provided you maintain
|
273 |
+
clear directions next to the object code saying where to find the
|
274 |
+
Corresponding Source. Regardless of what server hosts the
|
275 |
+
Corresponding Source, you remain obligated to ensure that it is
|
276 |
+
available for as long as needed to satisfy these requirements.
|
277 |
+
|
278 |
+
e) Convey the object code using peer-to-peer transmission, provided
|
279 |
+
you inform other peers where the object code and Corresponding
|
280 |
+
Source of the work are being offered to the general public at no
|
281 |
+
charge under subsection 6d.
|
282 |
+
|
283 |
+
A separable portion of the object code, whose source code is excluded
|
284 |
+
from the Corresponding Source as a System Library, need not be
|
285 |
+
included in conveying the object code work.
|
286 |
+
|
287 |
+
A "User Product" is either (1) a "consumer product", which means any
|
288 |
+
tangible personal property which is normally used for personal, family,
|
289 |
+
or household purposes, or (2) anything designed or sold for incorporation
|
290 |
+
into a dwelling. In determining whether a product is a consumer product,
|
291 |
+
doubtful cases shall be resolved in favor of coverage. For a particular
|
292 |
+
product received by a particular user, "normally used" refers to a
|
293 |
+
typical or common use of that class of product, regardless of the status
|
294 |
+
of the particular user or of the way in which the particular user
|
295 |
+
actually uses, or expects or is expected to use, the product. A product
|
296 |
+
is a consumer product regardless of whether the product has substantial
|
297 |
+
commercial, industrial or non-consumer uses, unless such uses represent
|
298 |
+
the only significant mode of use of the product.
|
299 |
+
|
300 |
+
"Installation Information" for a User Product means any methods,
|
301 |
+
procedures, authorization keys, or other information required to install
|
302 |
+
and execute modified versions of a covered work in that User Product from
|
303 |
+
a modified version of its Corresponding Source. The information must
|
304 |
+
suffice to ensure that the continued functioning of the modified object
|
305 |
+
code is in no case prevented or interfered with solely because
|
306 |
+
modification has been made.
|
307 |
+
|
308 |
+
If you convey an object code work under this section in, or with, or
|
309 |
+
specifically for use in, a User Product, and the conveying occurs as
|
310 |
+
part of a transaction in which the right of possession and use of the
|
311 |
+
User Product is transferred to the recipient in perpetuity or for a
|
312 |
+
fixed term (regardless of how the transaction is characterized), the
|
313 |
+
Corresponding Source conveyed under this section must be accompanied
|
314 |
+
by the Installation Information. But this requirement does not apply
|
315 |
+
if neither you nor any third party retains the ability to install
|
316 |
+
modified object code on the User Product (for example, the work has
|
317 |
+
been installed in ROM).
|
318 |
+
|
319 |
+
The requirement to provide Installation Information does not include a
|
320 |
+
requirement to continue to provide support service, warranty, or updates
|
321 |
+
for a work that has been modified or installed by the recipient, or for
|
322 |
+
the User Product in which it has been modified or installed. Access to a
|
323 |
+
network may be denied when the modification itself materially and
|
324 |
+
adversely affects the operation of the network or violates the rules and
|
325 |
+
protocols for communication across the network.
|
326 |
+
|
327 |
+
Corresponding Source conveyed, and Installation Information provided,
|
328 |
+
in accord with this section must be in a format that is publicly
|
329 |
+
documented (and with an implementation available to the public in
|
330 |
+
source code form), and must require no special password or key for
|
331 |
+
unpacking, reading or copying.
|
332 |
+
|
333 |
+
7. Additional Terms.
|
334 |
+
|
335 |
+
"Additional permissions" are terms that supplement the terms of this
|
336 |
+
License by making exceptions from one or more of its conditions.
|
337 |
+
Additional permissions that are applicable to the entire Program shall
|
338 |
+
be treated as though they were included in this License, to the extent
|
339 |
+
that they are valid under applicable law. If additional permissions
|
340 |
+
apply only to part of the Program, that part may be used separately
|
341 |
+
under those permissions, but the entire Program remains governed by
|
342 |
+
this License without regard to the additional permissions.
|
343 |
+
|
344 |
+
When you convey a copy of a covered work, you may at your option
|
345 |
+
remove any additional permissions from that copy, or from any part of
|
346 |
+
it. (Additional permissions may be written to require their own
|
347 |
+
removal in certain cases when you modify the work.) You may place
|
348 |
+
additional permissions on material, added by you to a covered work,
|
349 |
+
for which you have or can give appropriate copyright permission.
|
350 |
+
|
351 |
+
Notwithstanding any other provision of this License, for material you
|
352 |
+
add to a covered work, you may (if authorized by the copyright holders of
|
353 |
+
that material) supplement the terms of this License with terms:
|
354 |
+
|
355 |
+
a) Disclaiming warranty or limiting liability differently from the
|
356 |
+
terms of sections 15 and 16 of this License; or
|
357 |
+
|
358 |
+
b) Requiring preservation of specified reasonable legal notices or
|
359 |
+
author attributions in that material or in the Appropriate Legal
|
360 |
+
Notices displayed by works containing it; or
|
361 |
+
|
362 |
+
c) Prohibiting misrepresentation of the origin of that material, or
|
363 |
+
requiring that modified versions of such material be marked in
|
364 |
+
reasonable ways as different from the original version; or
|
365 |
+
|
366 |
+
d) Limiting the use for publicity purposes of names of licensors or
|
367 |
+
authors of the material; or
|
368 |
+
|
369 |
+
e) Declining to grant rights under trademark law for use of some
|
370 |
+
trade names, trademarks, or service marks; or
|
371 |
+
|
372 |
+
f) Requiring indemnification of licensors and authors of that
|
373 |
+
material by anyone who conveys the material (or modified versions of
|
374 |
+
it) with contractual assumptions of liability to the recipient, for
|
375 |
+
any liability that these contractual assumptions directly impose on
|
376 |
+
those licensors and authors.
|
377 |
+
|
378 |
+
All other non-permissive additional terms are considered "further
|
379 |
+
restrictions" within the meaning of section 10. If the Program as you
|
380 |
+
received it, or any part of it, contains a notice stating that it is
|
381 |
+
governed by this License along with a term that is a further
|
382 |
+
restriction, you may remove that term. If a license document contains
|
383 |
+
a further restriction but permits relicensing or conveying under this
|
384 |
+
License, you may add to a covered work material governed by the terms
|
385 |
+
of that license document, provided that the further restriction does
|
386 |
+
not survive such relicensing or conveying.
|
387 |
+
|
388 |
+
If you add terms to a covered work in accord with this section, you
|
389 |
+
must place, in the relevant source files, a statement of the
|
390 |
+
additional terms that apply to those files, or a notice indicating
|
391 |
+
where to find the applicable terms.
|
392 |
+
|
393 |
+
Additional terms, permissive or non-permissive, may be stated in the
|
394 |
+
form of a separately written license, or stated as exceptions;
|
395 |
+
the above requirements apply either way.
|
396 |
+
|
397 |
+
8. Termination.
|
398 |
+
|
399 |
+
You may not propagate or modify a covered work except as expressly
|
400 |
+
provided under this License. Any attempt otherwise to propagate or
|
401 |
+
modify it is void, and will automatically terminate your rights under
|
402 |
+
this License (including any patent licenses granted under the third
|
403 |
+
paragraph of section 11).
|
404 |
+
|
405 |
+
However, if you cease all violation of this License, then your
|
406 |
+
license from a particular copyright holder is reinstated (a)
|
407 |
+
provisionally, unless and until the copyright holder explicitly and
|
408 |
+
finally terminates your license, and (b) permanently, if the copyright
|
409 |
+
holder fails to notify you of the violation by some reasonable means
|
410 |
+
prior to 60 days after the cessation.
|
411 |
+
|
412 |
+
Moreover, your license from a particular copyright holder is
|
413 |
+
reinstated permanently if the copyright holder notifies you of the
|
414 |
+
violation by some reasonable means, this is the first time you have
|
415 |
+
received notice of violation of this License (for any work) from that
|
416 |
+
copyright holder, and you cure the violation prior to 30 days after
|
417 |
+
your receipt of the notice.
|
418 |
+
|
419 |
+
Termination of your rights under this section does not terminate the
|
420 |
+
licenses of parties who have received copies or rights from you under
|
421 |
+
this License. If your rights have been terminated and not permanently
|
422 |
+
reinstated, you do not qualify to receive new licenses for the same
|
423 |
+
material under section 10.
|
424 |
+
|
425 |
+
9. Acceptance Not Required for Having Copies.
|
426 |
+
|
427 |
+
You are not required to accept this License in order to receive or
|
428 |
+
run a copy of the Program. Ancillary propagation of a covered work
|
429 |
+
occurring solely as a consequence of using peer-to-peer transmission
|
430 |
+
to receive a copy likewise does not require acceptance. However,
|
431 |
+
nothing other than this License grants you permission to propagate or
|
432 |
+
modify any covered work. These actions infringe copyright if you do
|
433 |
+
not accept this License. Therefore, by modifying or propagating a
|
434 |
+
covered work, you indicate your acceptance of this License to do so.
|
435 |
+
|
436 |
+
10. Automatic Licensing of Downstream Recipients.
|
437 |
+
|
438 |
+
Each time you convey a covered work, the recipient automatically
|
439 |
+
receives a license from the original licensors, to run, modify and
|
440 |
+
propagate that work, subject to this License. You are not responsible
|
441 |
+
for enforcing compliance by third parties with this License.
|
442 |
+
|
443 |
+
An "entity transaction" is a transaction transferring control of an
|
444 |
+
organization, or substantially all assets of one, or subdividing an
|
445 |
+
organization, or merging organizations. If propagation of a covered
|
446 |
+
work results from an entity transaction, each party to that
|
447 |
+
transaction who receives a copy of the work also receives whatever
|
448 |
+
licenses to the work the party's predecessor in interest had or could
|
449 |
+
give under the previous paragraph, plus a right to possession of the
|
450 |
+
Corresponding Source of the work from the predecessor in interest, if
|
451 |
+
the predecessor has it or can get it with reasonable efforts.
|
452 |
+
|
453 |
+
You may not impose any further restrictions on the exercise of the
|
454 |
+
rights granted or affirmed under this License. For example, you may
|
455 |
+
not impose a license fee, royalty, or other charge for exercise of
|
456 |
+
rights granted under this License, and you may not initiate litigation
|
457 |
+
(including a cross-claim or counterclaim in a lawsuit) alleging that
|
458 |
+
any patent claim is infringed by making, using, selling, offering for
|
459 |
+
sale, or importing the Program or any portion of it.
|
460 |
+
|
461 |
+
11. Patents.
|
462 |
+
|
463 |
+
A "contributor" is a copyright holder who authorizes use under this
|
464 |
+
License of the Program or a work on which the Program is based. The
|
465 |
+
work thus licensed is called the contributor's "contributor version".
|
466 |
+
|
467 |
+
A contributor's "essential patent claims" are all patent claims
|
468 |
+
owned or controlled by the contributor, whether already acquired or
|
469 |
+
hereafter acquired, that would be infringed by some manner, permitted
|
470 |
+
by this License, of making, using, or selling its contributor version,
|
471 |
+
but do not include claims that would be infringed only as a
|
472 |
+
consequence of further modification of the contributor version. For
|
473 |
+
purposes of this definition, "control" includes the right to grant
|
474 |
+
patent sublicenses in a manner consistent with the requirements of
|
475 |
+
this License.
|
476 |
+
|
477 |
+
Each contributor grants you a non-exclusive, worldwide, royalty-free
|
478 |
+
patent license under the contributor's essential patent claims, to
|
479 |
+
make, use, sell, offer for sale, import and otherwise run, modify and
|
480 |
+
propagate the contents of its contributor version.
|
481 |
+
|
482 |
+
In the following three paragraphs, a "patent license" is any express
|
483 |
+
agreement or commitment, however denominated, not to enforce a patent
|
484 |
+
(such as an express permission to practice a patent or covenant not to
|
485 |
+
sue for patent infringement). To "grant" such a patent license to a
|
486 |
+
party means to make such an agreement or commitment not to enforce a
|
487 |
+
patent against the party.
|
488 |
+
|
489 |
+
If you convey a covered work, knowingly relying on a patent license,
|
490 |
+
and the Corresponding Source of the work is not available for anyone
|
491 |
+
to copy, free of charge and under the terms of this License, through a
|
492 |
+
publicly available network server or other readily accessible means,
|
493 |
+
then you must either (1) cause the Corresponding Source to be so
|
494 |
+
available, or (2) arrange to deprive yourself of the benefit of the
|
495 |
+
patent license for this particular work, or (3) arrange, in a manner
|
496 |
+
consistent with the requirements of this License, to extend the patent
|
497 |
+
license to downstream recipients. "Knowingly relying" means you have
|
498 |
+
actual knowledge that, but for the patent license, your conveying the
|
499 |
+
covered work in a country, or your recipient's use of the covered work
|
500 |
+
in a country, would infringe one or more identifiable patents in that
|
501 |
+
country that you have reason to believe are valid.
|
502 |
+
|
503 |
+
If, pursuant to or in connection with a single transaction or
|
504 |
+
arrangement, you convey, or propagate by procuring conveyance of, a
|
505 |
+
covered work, and grant a patent license to some of the parties
|
506 |
+
receiving the covered work authorizing them to use, propagate, modify
|
507 |
+
or convey a specific copy of the covered work, then the patent license
|
508 |
+
you grant is automatically extended to all recipients of the covered
|
509 |
+
work and works based on it.
|
510 |
+
|
511 |
+
A patent license is "discriminatory" if it does not include within
|
512 |
+
the scope of its coverage, prohibits the exercise of, or is
|
513 |
+
conditioned on the non-exercise of one or more of the rights that are
|
514 |
+
specifically granted under this License. You may not convey a covered
|
515 |
+
work if you are a party to an arrangement with a third party that is
|
516 |
+
in the business of distributing software, under which you make payment
|
517 |
+
to the third party based on the extent of your activity of conveying
|
518 |
+
the work, and under which the third party grants, to any of the
|
519 |
+
parties who would receive the covered work from you, a discriminatory
|
520 |
+
patent license (a) in connection with copies of the covered work
|
521 |
+
conveyed by you (or copies made from those copies), or (b) primarily
|
522 |
+
for and in connection with specific products or compilations that
|
523 |
+
contain the covered work, unless you entered into that arrangement,
|
524 |
+
or that patent license was granted, prior to 28 March 2007.
|
525 |
+
|
526 |
+
Nothing in this License shall be construed as excluding or limiting
|
527 |
+
any implied license or other defenses to infringement that may
|
528 |
+
otherwise be available to you under applicable patent law.
|
529 |
+
|
530 |
+
12. No Surrender of Others' Freedom.
|
531 |
+
|
532 |
+
If conditions are imposed on you (whether by court order, agreement or
|
533 |
+
otherwise) that contradict the conditions of this License, they do not
|
534 |
+
excuse you from the conditions of this License. If you cannot convey a
|
535 |
+
covered work so as to satisfy simultaneously your obligations under this
|
536 |
+
License and any other pertinent obligations, then as a consequence you may
|
537 |
+
not convey it at all. For example, if you agree to terms that obligate you
|
538 |
+
to collect a royalty for further conveying from those to whom you convey
|
539 |
+
the Program, the only way you could satisfy both those terms and this
|
540 |
+
License would be to refrain entirely from conveying the Program.
|
541 |
+
|
542 |
+
13. Remote Network Interaction; Use with the GNU General Public License.
|
543 |
+
|
544 |
+
Notwithstanding any other provision of this License, if you modify the
|
545 |
+
Program, your modified version must prominently offer all users
|
546 |
+
interacting with it remotely through a computer network (if your version
|
547 |
+
supports such interaction) an opportunity to receive the Corresponding
|
548 |
+
Source of your version by providing access to the Corresponding Source
|
549 |
+
from a network server at no charge, through some standard or customary
|
550 |
+
means of facilitating copying of software. This Corresponding Source
|
551 |
+
shall include the Corresponding Source for any work covered by version 3
|
552 |
+
of the GNU General Public License that is incorporated pursuant to the
|
553 |
+
following paragraph.
|
554 |
+
|
555 |
+
Notwithstanding any other provision of this License, you have
|
556 |
+
permission to link or combine any covered work with a work licensed
|
557 |
+
under version 3 of the GNU General Public License into a single
|
558 |
+
combined work, and to convey the resulting work. The terms of this
|
559 |
+
License will continue to apply to the part which is the covered work,
|
560 |
+
but the work with which it is combined will remain governed by version
|
561 |
+
3 of the GNU General Public License.
|
562 |
+
|
563 |
+
14. Revised Versions of this License.
|
564 |
+
|
565 |
+
The Free Software Foundation may publish revised and/or new versions of
|
566 |
+
the GNU Affero General Public License from time to time. Such new versions
|
567 |
+
will be similar in spirit to the present version, but may differ in detail to
|
568 |
+
address new problems or concerns.
|
569 |
+
|
570 |
+
Each version is given a distinguishing version number. If the
|
571 |
+
Program specifies that a certain numbered version of the GNU Affero General
|
572 |
+
Public License "or any later version" applies to it, you have the
|
573 |
+
option of following the terms and conditions either of that numbered
|
574 |
+
version or of any later version published by the Free Software
|
575 |
+
Foundation. If the Program does not specify a version number of the
|
576 |
+
GNU Affero General Public License, you may choose any version ever published
|
577 |
+
by the Free Software Foundation.
|
578 |
+
|
579 |
+
If the Program specifies that a proxy can decide which future
|
580 |
+
versions of the GNU Affero General Public License can be used, that proxy's
|
581 |
+
public statement of acceptance of a version permanently authorizes you
|
582 |
+
to choose that version for the Program.
|
583 |
+
|
584 |
+
Later license versions may give you additional or different
|
585 |
+
permissions. However, no additional obligations are imposed on any
|
586 |
+
author or copyright holder as a result of your choosing to follow a
|
587 |
+
later version.
|
588 |
+
|
589 |
+
15. Disclaimer of Warranty.
|
590 |
+
|
591 |
+
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
|
592 |
+
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
|
593 |
+
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
|
594 |
+
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
|
595 |
+
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
596 |
+
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
|
597 |
+
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
|
598 |
+
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
|
599 |
+
|
600 |
+
16. Limitation of Liability.
|
601 |
+
|
602 |
+
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
|
603 |
+
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
|
604 |
+
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
|
605 |
+
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
|
606 |
+
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
|
607 |
+
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
|
608 |
+
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
|
609 |
+
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
|
610 |
+
SUCH DAMAGES.
|
611 |
+
|
612 |
+
17. Interpretation of Sections 15 and 16.
|
613 |
+
|
614 |
+
If the disclaimer of warranty and limitation of liability provided
|
615 |
+
above cannot be given local legal effect according to their terms,
|
616 |
+
reviewing courts shall apply local law that most closely approximates
|
617 |
+
an absolute waiver of all civil liability in connection with the
|
618 |
+
Program, unless a warranty or assumption of liability accompanies a
|
619 |
+
copy of the Program in return for a fee.
|
620 |
+
|
621 |
+
END OF TERMS AND CONDITIONS
|
622 |
+
|
623 |
+
How to Apply These Terms to Your New Programs
|
624 |
+
|
625 |
+
If you develop a new program, and you want it to be of the greatest
|
626 |
+
possible use to the public, the best way to achieve this is to make it
|
627 |
+
free software which everyone can redistribute and change under these terms.
|
628 |
+
|
629 |
+
To do so, attach the following notices to the program. It is safest
|
630 |
+
to attach them to the start of each source file to most effectively
|
631 |
+
state the exclusion of warranty; and each file should have at least
|
632 |
+
the "copyright" line and a pointer to where the full notice is found.
|
633 |
+
|
634 |
+
<one line to give the program's name and a brief idea of what it does.>
|
635 |
+
Copyright (C) <year> <name of author>
|
636 |
+
|
637 |
+
This program is free software: you can redistribute it and/or modify
|
638 |
+
it under the terms of the GNU Affero General Public License as published by
|
639 |
+
the Free Software Foundation, either version 3 of the License, or
|
640 |
+
(at your option) any later version.
|
641 |
+
|
642 |
+
This program is distributed in the hope that it will be useful,
|
643 |
+
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
644 |
+
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
645 |
+
GNU Affero General Public License for more details.
|
646 |
+
|
647 |
+
You should have received a copy of the GNU Affero General Public License
|
648 |
+
along with this program. If not, see <https://www.gnu.org/licenses/>.
|
649 |
+
|
650 |
+
Also add information on how to contact you by electronic and paper mail.
|
651 |
+
|
652 |
+
If your software can interact with users remotely through a computer
|
653 |
+
network, you should also make sure that it provides a way for users to
|
654 |
+
get its source. For example, if your program is a web application, its
|
655 |
+
interface could display a "Source" link that leads users to an archive
|
656 |
+
of the code. There are many ways you could offer source, and different
|
657 |
+
solutions will be better for different programs; see section 13 for the
|
658 |
+
specific requirements.
|
659 |
+
|
660 |
+
You should also get your employer (if you work as a programmer) or school,
|
661 |
+
if any, to sign a "copyright disclaimer" for the program, if necessary.
|
662 |
+
For more information on this, and how to apply and follow the GNU AGPL, see
|
663 |
+
<https://www.gnu.org/licenses/>.
|
README.md
CHANGED
@@ -1,3 +1,177 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Stable Diffusion web UI
|
2 |
+
A browser interface based on Gradio library for Stable Diffusion.
|
3 |
+
|
4 |
+

|
5 |
+
|
6 |
+
## Features
|
7 |
+
[Detailed feature showcase with images](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features):
|
8 |
+
- Original txt2img and img2img modes
|
9 |
+
- One click install and run script (but you still must install python and git)
|
10 |
+
- Outpainting
|
11 |
+
- Inpainting
|
12 |
+
- Color Sketch
|
13 |
+
- Prompt Matrix
|
14 |
+
- Stable Diffusion Upscale
|
15 |
+
- Attention, specify parts of text that the model should pay more attention to
|
16 |
+
- a man in a `((tuxedo))` - will pay more attention to tuxedo
|
17 |
+
- a man in a `(tuxedo:1.21)` - alternative syntax
|
18 |
+
- select text and press `Ctrl+Up` or `Ctrl+Down` (or `Command+Up` or `Command+Down` if you're on a MacOS) to automatically adjust attention to selected text (code contributed by anonymous user)
|
19 |
+
- Loopback, run img2img processing multiple times
|
20 |
+
- X/Y/Z plot, a way to draw a 3 dimensional plot of images with different parameters
|
21 |
+
- Textual Inversion
|
22 |
+
- have as many embeddings as you want and use any names you like for them
|
23 |
+
- use multiple embeddings with different numbers of vectors per token
|
24 |
+
- works with half precision floating point numbers
|
25 |
+
- train embeddings on 8GB (also reports of 6GB working)
|
26 |
+
- Extras tab with:
|
27 |
+
- GFPGAN, neural network that fixes faces
|
28 |
+
- CodeFormer, face restoration tool as an alternative to GFPGAN
|
29 |
+
- RealESRGAN, neural network upscaler
|
30 |
+
- ESRGAN, neural network upscaler with a lot of third party models
|
31 |
+
- SwinIR and Swin2SR ([see here](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/2092)), neural network upscalers
|
32 |
+
- LDSR, Latent diffusion super resolution upscaling
|
33 |
+
- Resizing aspect ratio options
|
34 |
+
- Sampling method selection
|
35 |
+
- Adjust sampler eta values (noise multiplier)
|
36 |
+
- More advanced noise setting options
|
37 |
+
- Interrupt processing at any time
|
38 |
+
- 4GB video card support (also reports of 2GB working)
|
39 |
+
- Correct seeds for batches
|
40 |
+
- Live prompt token length validation
|
41 |
+
- Generation parameters
|
42 |
+
- parameters you used to generate images are saved with that image
|
43 |
+
- in PNG chunks for PNG, in EXIF for JPEG
|
44 |
+
- can drag the image to PNG info tab to restore generation parameters and automatically copy them into UI
|
45 |
+
- can be disabled in settings
|
46 |
+
- drag and drop an image/text-parameters to promptbox
|
47 |
+
- Read Generation Parameters Button, loads parameters in promptbox to UI
|
48 |
+
- Settings page
|
49 |
+
- Running arbitrary python code from UI (must run with `--allow-code` to enable)
|
50 |
+
- Mouseover hints for most UI elements
|
51 |
+
- Possible to change defaults/mix/max/step values for UI elements via text config
|
52 |
+
- Tiling support, a checkbox to create images that can be tiled like textures
|
53 |
+
- Progress bar and live image generation preview
|
54 |
+
- Can use a separate neural network to produce previews with almost none VRAM or compute requirement
|
55 |
+
- Negative prompt, an extra text field that allows you to list what you don't want to see in generated image
|
56 |
+
- Styles, a way to save part of prompt and easily apply them via dropdown later
|
57 |
+
- Variations, a way to generate same image but with tiny differences
|
58 |
+
- Seed resizing, a way to generate same image but at slightly different resolution
|
59 |
+
- CLIP interrogator, a button that tries to guess prompt from an image
|
60 |
+
- Prompt Editing, a way to change prompt mid-generation, say to start making a watermelon and switch to anime girl midway
|
61 |
+
- Batch Processing, process a group of files using img2img
|
62 |
+
- Img2img Alternative, reverse Euler method of cross attention control
|
63 |
+
- Highres Fix, a convenience option to produce high resolution pictures in one click without usual distortions
|
64 |
+
- Reloading checkpoints on the fly
|
65 |
+
- Checkpoint Merger, a tab that allows you to merge up to 3 checkpoints into one
|
66 |
+
- [Custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts) with many extensions from community
|
67 |
+
- [Composable-Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/), a way to use multiple prompts at once
|
68 |
+
- separate prompts using uppercase `AND`
|
69 |
+
- also supports weights for prompts: `a cat :1.2 AND a dog AND a penguin :2.2`
|
70 |
+
- No token limit for prompts (original stable diffusion lets you use up to 75 tokens)
|
71 |
+
- DeepDanbooru integration, creates danbooru style tags for anime prompts
|
72 |
+
- [xformers](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers), major speed increase for select cards: (add `--xformers` to commandline args)
|
73 |
+
- via extension: [History tab](https://github.com/yfszzx/stable-diffusion-webui-images-browser): view, direct and delete images conveniently within the UI
|
74 |
+
- Generate forever option
|
75 |
+
- Training tab
|
76 |
+
- hypernetworks and embeddings options
|
77 |
+
- Preprocessing images: cropping, mirroring, autotagging using BLIP or deepdanbooru (for anime)
|
78 |
+
- Clip skip
|
79 |
+
- Hypernetworks
|
80 |
+
- Loras (same as Hypernetworks but more pretty)
|
81 |
+
- A separate UI where you can choose, with preview, which embeddings, hypernetworks or Loras to add to your prompt
|
82 |
+
- Can select to load a different VAE from settings screen
|
83 |
+
- Estimated completion time in progress bar
|
84 |
+
- API
|
85 |
+
- Support for dedicated [inpainting model](https://github.com/runwayml/stable-diffusion#inpainting-with-stable-diffusion) by RunwayML
|
86 |
+
- via extension: [Aesthetic Gradients](https://github.com/AUTOMATIC1111/stable-diffusion-webui-aesthetic-gradients), a way to generate images with a specific aesthetic by using clip images embeds (implementation of [https://github.com/vicgalle/stable-diffusion-aesthetic-gradients](https://github.com/vicgalle/stable-diffusion-aesthetic-gradients))
|
87 |
+
- [Stable Diffusion 2.0](https://github.com/Stability-AI/stablediffusion) support - see [wiki](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#stable-diffusion-20) for instructions
|
88 |
+
- [Alt-Diffusion](https://arxiv.org/abs/2211.06679) support - see [wiki](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alt-diffusion) for instructions
|
89 |
+
- Now without any bad letters!
|
90 |
+
- Load checkpoints in safetensors format
|
91 |
+
- Eased resolution restriction: generated image's dimension must be a multiple of 8 rather than 64
|
92 |
+
- Now with a license!
|
93 |
+
- Reorder elements in the UI from settings screen
|
94 |
+
|
95 |
+
## Installation and Running
|
96 |
+
Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for:
|
97 |
+
- [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended)
|
98 |
+
- [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs.
|
99 |
+
- [Intel CPUs, Intel GPUs (both integrated and discrete)](https://github.com/openvinotoolkit/stable-diffusion-webui/wiki/Installation-on-Intel-Silicon) (external wiki page)
|
100 |
+
|
101 |
+
Alternatively, use online services (like Google Colab):
|
102 |
+
|
103 |
+
- [List of Online Services](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Online-Services)
|
104 |
+
|
105 |
+
### Installation on Windows 10/11 with NVidia-GPUs using release package
|
106 |
+
1. Download `sd.webui.zip` from [v1.0.0-pre](https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.0.0-pre) and extract it's contents.
|
107 |
+
2. Run `update.bat`.
|
108 |
+
3. Run `run.bat`.
|
109 |
+
> For more details see [Install-and-Run-on-NVidia-GPUs](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs)
|
110 |
+
|
111 |
+
### Automatic Installation on Windows
|
112 |
+
1. Install [Python 3.10.6](https://www.python.org/downloads/release/python-3106/) (Newer version of Python does not support torch), checking "Add Python to PATH".
|
113 |
+
2. Install [git](https://git-scm.com/download/win).
|
114 |
+
3. Download the stable-diffusion-webui repository, for example by running `git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git`.
|
115 |
+
4. Run `webui-user.bat` from Windows Explorer as normal, non-administrator, user.
|
116 |
+
|
117 |
+
### Automatic Installation on Linux
|
118 |
+
1. Install the dependencies:
|
119 |
+
```bash
|
120 |
+
# Debian-based:
|
121 |
+
sudo apt install wget git python3 python3-venv libgl1 libglib2.0-0
|
122 |
+
# Red Hat-based:
|
123 |
+
sudo dnf install wget git python3
|
124 |
+
# Arch-based:
|
125 |
+
sudo pacman -S wget git python3
|
126 |
+
```
|
127 |
+
2. Navigate to the directory you would like the webui to be installed and execute the following command:
|
128 |
+
```bash
|
129 |
+
wget -q https://raw.githubusercontent.com/AUTOMATIC1111/stable-diffusion-webui/master/webui.sh
|
130 |
+
```
|
131 |
+
3. Run `webui.sh`.
|
132 |
+
4. Check `webui-user.sh` for options.
|
133 |
+
### Installation on Apple Silicon
|
134 |
+
|
135 |
+
Find the instructions [here](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Installation-on-Apple-Silicon).
|
136 |
+
|
137 |
+
## Contributing
|
138 |
+
Here's how to add code to this repo: [Contributing](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Contributing)
|
139 |
+
|
140 |
+
## Documentation
|
141 |
+
|
142 |
+
The documentation was moved from this README over to the project's [wiki](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki).
|
143 |
+
|
144 |
+
For the purposes of getting Google and other search engines to crawl the wiki, here's a link to the (not for humans) [crawlable wiki](https://github-wiki-see.page/m/AUTOMATIC1111/stable-diffusion-webui/wiki).
|
145 |
+
|
146 |
+
## Credits
|
147 |
+
Licenses for borrowed code can be found in `Settings -> Licenses` screen, and also in `html/licenses.html` file.
|
148 |
+
|
149 |
+
- Stable Diffusion - https://github.com/CompVis/stable-diffusion, https://github.com/CompVis/taming-transformers
|
150 |
+
- k-diffusion - https://github.com/crowsonkb/k-diffusion.git
|
151 |
+
- GFPGAN - https://github.com/TencentARC/GFPGAN.git
|
152 |
+
- CodeFormer - https://github.com/sczhou/CodeFormer
|
153 |
+
- ESRGAN - https://github.com/xinntao/ESRGAN
|
154 |
+
- SwinIR - https://github.com/JingyunLiang/SwinIR
|
155 |
+
- Swin2SR - https://github.com/mv-lab/swin2sr
|
156 |
+
- LDSR - https://github.com/Hafiidz/latent-diffusion
|
157 |
+
- MiDaS - https://github.com/isl-org/MiDaS
|
158 |
+
- Ideas for optimizations - https://github.com/basujindal/stable-diffusion
|
159 |
+
- Cross Attention layer optimization - Doggettx - https://github.com/Doggettx/stable-diffusion, original idea for prompt editing.
|
160 |
+
- Cross Attention layer optimization - InvokeAI, lstein - https://github.com/invoke-ai/InvokeAI (originally http://github.com/lstein/stable-diffusion)
|
161 |
+
- Sub-quadratic Cross Attention layer optimization - Alex Birch (https://github.com/Birch-san/diffusers/pull/1), Amin Rezaei (https://github.com/AminRezaei0x443/memory-efficient-attention)
|
162 |
+
- Textual Inversion - Rinon Gal - https://github.com/rinongal/textual_inversion (we're not using his code, but we are using his ideas).
|
163 |
+
- Idea for SD upscale - https://github.com/jquesnelle/txt2imghd
|
164 |
+
- Noise generation for outpainting mk2 - https://github.com/parlance-zz/g-diffuser-bot
|
165 |
+
- CLIP interrogator idea and borrowing some code - https://github.com/pharmapsychotic/clip-interrogator
|
166 |
+
- Idea for Composable Diffusion - https://github.com/energy-based-model/Compositional-Visual-Generation-with-Composable-Diffusion-Models-PyTorch
|
167 |
+
- xformers - https://github.com/facebookresearch/xformers
|
168 |
+
- DeepDanbooru - interrogator for anime diffusers https://github.com/KichangKim/DeepDanbooru
|
169 |
+
- Sampling in float32 precision from a float16 UNet - marunine for the idea, Birch-san for the example Diffusers implementation (https://github.com/Birch-san/diffusers-play/tree/92feee6)
|
170 |
+
- Instruct pix2pix - Tim Brooks (star), Aleksander Holynski (star), Alexei A. Efros (no star) - https://github.com/timothybrooks/instruct-pix2pix
|
171 |
+
- Security advice - RyotaK
|
172 |
+
- UniPC sampler - Wenliang Zhao - https://github.com/wl-zhao/UniPC
|
173 |
+
- TAESD - Ollin Boer Bohan - https://github.com/madebyollin/taesd
|
174 |
+
- LyCORIS - KohakuBlueleaf
|
175 |
+
- Restart sampling - lambertae - https://github.com/Newbeeer/diffusion_restart_sampling
|
176 |
+
- Initial Gradio script - posted on 4chan by an Anonymous user. Thank you Anonymous user.
|
177 |
+
- (You)
|
__pycache__/launch.cpython-310.pyc
ADDED
Binary file (832 Bytes). View file
|
|
__pycache__/webui.cpython-310.pyc
ADDED
Binary file (15 kB). View file
|
|
cache.json
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"hashes": {
|
3 |
+
"checkpoint/absolutereality_v16.safetensors": {
|
4 |
+
"mtime": 1690338820.1232889,
|
5 |
+
"sha256": "be1d90c4abb7bb0183f267f899f38b44112ad6ef9a757a6723514ea4e9be15dc"
|
6 |
+
}
|
7 |
+
}
|
8 |
+
}
|
configs/alt-diffusion-inference.yaml
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 1.0e-04
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.00085
|
6 |
+
linear_end: 0.0120
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: "jpg"
|
11 |
+
cond_stage_key: "txt"
|
12 |
+
image_size: 64
|
13 |
+
channels: 4
|
14 |
+
cond_stage_trainable: false # Note: different from the one we trained before
|
15 |
+
conditioning_key: crossattn
|
16 |
+
monitor: val/loss_simple_ema
|
17 |
+
scale_factor: 0.18215
|
18 |
+
use_ema: False
|
19 |
+
|
20 |
+
scheduler_config: # 10000 warmup steps
|
21 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
22 |
+
params:
|
23 |
+
warm_up_steps: [ 10000 ]
|
24 |
+
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
|
25 |
+
f_start: [ 1.e-6 ]
|
26 |
+
f_max: [ 1. ]
|
27 |
+
f_min: [ 1. ]
|
28 |
+
|
29 |
+
unet_config:
|
30 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
31 |
+
params:
|
32 |
+
image_size: 32 # unused
|
33 |
+
in_channels: 4
|
34 |
+
out_channels: 4
|
35 |
+
model_channels: 320
|
36 |
+
attention_resolutions: [ 4, 2, 1 ]
|
37 |
+
num_res_blocks: 2
|
38 |
+
channel_mult: [ 1, 2, 4, 4 ]
|
39 |
+
num_heads: 8
|
40 |
+
use_spatial_transformer: True
|
41 |
+
transformer_depth: 1
|
42 |
+
context_dim: 768
|
43 |
+
use_checkpoint: True
|
44 |
+
legacy: False
|
45 |
+
|
46 |
+
first_stage_config:
|
47 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
48 |
+
params:
|
49 |
+
embed_dim: 4
|
50 |
+
monitor: val/rec_loss
|
51 |
+
ddconfig:
|
52 |
+
double_z: true
|
53 |
+
z_channels: 4
|
54 |
+
resolution: 256
|
55 |
+
in_channels: 3
|
56 |
+
out_ch: 3
|
57 |
+
ch: 128
|
58 |
+
ch_mult:
|
59 |
+
- 1
|
60 |
+
- 2
|
61 |
+
- 4
|
62 |
+
- 4
|
63 |
+
num_res_blocks: 2
|
64 |
+
attn_resolutions: []
|
65 |
+
dropout: 0.0
|
66 |
+
lossconfig:
|
67 |
+
target: torch.nn.Identity
|
68 |
+
|
69 |
+
cond_stage_config:
|
70 |
+
target: modules.xlmr.BertSeriesModelWithTransformation
|
71 |
+
params:
|
72 |
+
name: "XLMR-Large"
|
configs/instruct-pix2pix.yaml
ADDED
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
|
2 |
+
# See more details in LICENSE.
|
3 |
+
|
4 |
+
model:
|
5 |
+
base_learning_rate: 1.0e-04
|
6 |
+
target: modules.models.diffusion.ddpm_edit.LatentDiffusion
|
7 |
+
params:
|
8 |
+
linear_start: 0.00085
|
9 |
+
linear_end: 0.0120
|
10 |
+
num_timesteps_cond: 1
|
11 |
+
log_every_t: 200
|
12 |
+
timesteps: 1000
|
13 |
+
first_stage_key: edited
|
14 |
+
cond_stage_key: edit
|
15 |
+
# image_size: 64
|
16 |
+
# image_size: 32
|
17 |
+
image_size: 16
|
18 |
+
channels: 4
|
19 |
+
cond_stage_trainable: false # Note: different from the one we trained before
|
20 |
+
conditioning_key: hybrid
|
21 |
+
monitor: val/loss_simple_ema
|
22 |
+
scale_factor: 0.18215
|
23 |
+
use_ema: false
|
24 |
+
|
25 |
+
scheduler_config: # 10000 warmup steps
|
26 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
27 |
+
params:
|
28 |
+
warm_up_steps: [ 0 ]
|
29 |
+
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
|
30 |
+
f_start: [ 1.e-6 ]
|
31 |
+
f_max: [ 1. ]
|
32 |
+
f_min: [ 1. ]
|
33 |
+
|
34 |
+
unet_config:
|
35 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
36 |
+
params:
|
37 |
+
image_size: 32 # unused
|
38 |
+
in_channels: 8
|
39 |
+
out_channels: 4
|
40 |
+
model_channels: 320
|
41 |
+
attention_resolutions: [ 4, 2, 1 ]
|
42 |
+
num_res_blocks: 2
|
43 |
+
channel_mult: [ 1, 2, 4, 4 ]
|
44 |
+
num_heads: 8
|
45 |
+
use_spatial_transformer: True
|
46 |
+
transformer_depth: 1
|
47 |
+
context_dim: 768
|
48 |
+
use_checkpoint: True
|
49 |
+
legacy: False
|
50 |
+
|
51 |
+
first_stage_config:
|
52 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
53 |
+
params:
|
54 |
+
embed_dim: 4
|
55 |
+
monitor: val/rec_loss
|
56 |
+
ddconfig:
|
57 |
+
double_z: true
|
58 |
+
z_channels: 4
|
59 |
+
resolution: 256
|
60 |
+
in_channels: 3
|
61 |
+
out_ch: 3
|
62 |
+
ch: 128
|
63 |
+
ch_mult:
|
64 |
+
- 1
|
65 |
+
- 2
|
66 |
+
- 4
|
67 |
+
- 4
|
68 |
+
num_res_blocks: 2
|
69 |
+
attn_resolutions: []
|
70 |
+
dropout: 0.0
|
71 |
+
lossconfig:
|
72 |
+
target: torch.nn.Identity
|
73 |
+
|
74 |
+
cond_stage_config:
|
75 |
+
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
|
76 |
+
|
77 |
+
data:
|
78 |
+
target: main.DataModuleFromConfig
|
79 |
+
params:
|
80 |
+
batch_size: 128
|
81 |
+
num_workers: 1
|
82 |
+
wrap: false
|
83 |
+
validation:
|
84 |
+
target: edit_dataset.EditDataset
|
85 |
+
params:
|
86 |
+
path: data/clip-filtered-dataset
|
87 |
+
cache_dir: data/
|
88 |
+
cache_name: data_10k
|
89 |
+
split: val
|
90 |
+
min_text_sim: 0.2
|
91 |
+
min_image_sim: 0.75
|
92 |
+
min_direction_sim: 0.2
|
93 |
+
max_samples_per_prompt: 1
|
94 |
+
min_resize_res: 512
|
95 |
+
max_resize_res: 512
|
96 |
+
crop_res: 512
|
97 |
+
output_as_edit: False
|
98 |
+
real_input: True
|
configs/v1-inference.yaml
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 1.0e-04
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.00085
|
6 |
+
linear_end: 0.0120
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: "jpg"
|
11 |
+
cond_stage_key: "txt"
|
12 |
+
image_size: 64
|
13 |
+
channels: 4
|
14 |
+
cond_stage_trainable: false # Note: different from the one we trained before
|
15 |
+
conditioning_key: crossattn
|
16 |
+
monitor: val/loss_simple_ema
|
17 |
+
scale_factor: 0.18215
|
18 |
+
use_ema: False
|
19 |
+
|
20 |
+
scheduler_config: # 10000 warmup steps
|
21 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
22 |
+
params:
|
23 |
+
warm_up_steps: [ 10000 ]
|
24 |
+
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
|
25 |
+
f_start: [ 1.e-6 ]
|
26 |
+
f_max: [ 1. ]
|
27 |
+
f_min: [ 1. ]
|
28 |
+
|
29 |
+
unet_config:
|
30 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
31 |
+
params:
|
32 |
+
image_size: 32 # unused
|
33 |
+
in_channels: 4
|
34 |
+
out_channels: 4
|
35 |
+
model_channels: 320
|
36 |
+
attention_resolutions: [ 4, 2, 1 ]
|
37 |
+
num_res_blocks: 2
|
38 |
+
channel_mult: [ 1, 2, 4, 4 ]
|
39 |
+
num_heads: 8
|
40 |
+
use_spatial_transformer: True
|
41 |
+
transformer_depth: 1
|
42 |
+
context_dim: 768
|
43 |
+
use_checkpoint: True
|
44 |
+
legacy: False
|
45 |
+
|
46 |
+
first_stage_config:
|
47 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
48 |
+
params:
|
49 |
+
embed_dim: 4
|
50 |
+
monitor: val/rec_loss
|
51 |
+
ddconfig:
|
52 |
+
double_z: true
|
53 |
+
z_channels: 4
|
54 |
+
resolution: 256
|
55 |
+
in_channels: 3
|
56 |
+
out_ch: 3
|
57 |
+
ch: 128
|
58 |
+
ch_mult:
|
59 |
+
- 1
|
60 |
+
- 2
|
61 |
+
- 4
|
62 |
+
- 4
|
63 |
+
num_res_blocks: 2
|
64 |
+
attn_resolutions: []
|
65 |
+
dropout: 0.0
|
66 |
+
lossconfig:
|
67 |
+
target: torch.nn.Identity
|
68 |
+
|
69 |
+
cond_stage_config:
|
70 |
+
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
|
configs/v1-inpainting-inference.yaml
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 7.5e-05
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentInpaintDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.00085
|
6 |
+
linear_end: 0.0120
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: "jpg"
|
11 |
+
cond_stage_key: "txt"
|
12 |
+
image_size: 64
|
13 |
+
channels: 4
|
14 |
+
cond_stage_trainable: false # Note: different from the one we trained before
|
15 |
+
conditioning_key: hybrid # important
|
16 |
+
monitor: val/loss_simple_ema
|
17 |
+
scale_factor: 0.18215
|
18 |
+
finetune_keys: null
|
19 |
+
|
20 |
+
scheduler_config: # 10000 warmup steps
|
21 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
22 |
+
params:
|
23 |
+
warm_up_steps: [ 2500 ] # NOTE for resuming. use 10000 if starting from scratch
|
24 |
+
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
|
25 |
+
f_start: [ 1.e-6 ]
|
26 |
+
f_max: [ 1. ]
|
27 |
+
f_min: [ 1. ]
|
28 |
+
|
29 |
+
unet_config:
|
30 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
31 |
+
params:
|
32 |
+
image_size: 32 # unused
|
33 |
+
in_channels: 9 # 4 data + 4 downscaled image + 1 mask
|
34 |
+
out_channels: 4
|
35 |
+
model_channels: 320
|
36 |
+
attention_resolutions: [ 4, 2, 1 ]
|
37 |
+
num_res_blocks: 2
|
38 |
+
channel_mult: [ 1, 2, 4, 4 ]
|
39 |
+
num_heads: 8
|
40 |
+
use_spatial_transformer: True
|
41 |
+
transformer_depth: 1
|
42 |
+
context_dim: 768
|
43 |
+
use_checkpoint: True
|
44 |
+
legacy: False
|
45 |
+
|
46 |
+
first_stage_config:
|
47 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
48 |
+
params:
|
49 |
+
embed_dim: 4
|
50 |
+
monitor: val/rec_loss
|
51 |
+
ddconfig:
|
52 |
+
double_z: true
|
53 |
+
z_channels: 4
|
54 |
+
resolution: 256
|
55 |
+
in_channels: 3
|
56 |
+
out_ch: 3
|
57 |
+
ch: 128
|
58 |
+
ch_mult:
|
59 |
+
- 1
|
60 |
+
- 2
|
61 |
+
- 4
|
62 |
+
- 4
|
63 |
+
num_res_blocks: 2
|
64 |
+
attn_resolutions: []
|
65 |
+
dropout: 0.0
|
66 |
+
lossconfig:
|
67 |
+
target: torch.nn.Identity
|
68 |
+
|
69 |
+
cond_stage_config:
|
70 |
+
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
|
embeddings/Place Textual Inversion embeddings here.txt
ADDED
File without changes
|
environment-wsl2.yaml
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
name: automatic
|
2 |
+
channels:
|
3 |
+
- pytorch
|
4 |
+
- defaults
|
5 |
+
dependencies:
|
6 |
+
- python=3.10
|
7 |
+
- pip=23.0
|
8 |
+
- cudatoolkit=11.8
|
9 |
+
- pytorch=2.0
|
10 |
+
- torchvision=0.15
|
11 |
+
- numpy=1.23
|
extensions-builtin/.DS_Store
ADDED
Binary file (8.2 kB). View file
|
|
extensions-builtin/LDSR/__pycache__/ldsr_model_arch.cpython-310.pyc
ADDED
Binary file (6.71 kB). View file
|
|
extensions-builtin/LDSR/__pycache__/preload.cpython-310.pyc
ADDED
Binary file (515 Bytes). View file
|
|
extensions-builtin/LDSR/__pycache__/sd_hijack_autoencoder.cpython-310.pyc
ADDED
Binary file (8.95 kB). View file
|
|
extensions-builtin/LDSR/__pycache__/sd_hijack_ddpm_v1.cpython-310.pyc
ADDED
Binary file (42.4 kB). View file
|
|
extensions-builtin/LDSR/__pycache__/vqvae_quantize.cpython-310.pyc
ADDED
Binary file (3.67 kB). View file
|
|
extensions-builtin/LDSR/ldsr_model_arch.py
ADDED
@@ -0,0 +1,250 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gc
|
3 |
+
import time
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
import torchvision
|
8 |
+
from PIL import Image
|
9 |
+
from einops import rearrange, repeat
|
10 |
+
from omegaconf import OmegaConf
|
11 |
+
import safetensors.torch
|
12 |
+
|
13 |
+
from ldm.models.diffusion.ddim import DDIMSampler
|
14 |
+
from ldm.util import instantiate_from_config, ismap
|
15 |
+
from modules import shared, sd_hijack, devices
|
16 |
+
|
17 |
+
cached_ldsr_model: torch.nn.Module = None
|
18 |
+
|
19 |
+
|
20 |
+
# Create LDSR Class
|
21 |
+
class LDSR:
|
22 |
+
def load_model_from_config(self, half_attention):
|
23 |
+
global cached_ldsr_model
|
24 |
+
|
25 |
+
if shared.opts.ldsr_cached and cached_ldsr_model is not None:
|
26 |
+
print("Loading model from cache")
|
27 |
+
model: torch.nn.Module = cached_ldsr_model
|
28 |
+
else:
|
29 |
+
print(f"Loading model from {self.modelPath}")
|
30 |
+
_, extension = os.path.splitext(self.modelPath)
|
31 |
+
if extension.lower() == ".safetensors":
|
32 |
+
pl_sd = safetensors.torch.load_file(self.modelPath, device="cpu")
|
33 |
+
else:
|
34 |
+
pl_sd = torch.load(self.modelPath, map_location="cpu")
|
35 |
+
sd = pl_sd["state_dict"] if "state_dict" in pl_sd else pl_sd
|
36 |
+
config = OmegaConf.load(self.yamlPath)
|
37 |
+
config.model.target = "ldm.models.diffusion.ddpm.LatentDiffusionV1"
|
38 |
+
model: torch.nn.Module = instantiate_from_config(config.model)
|
39 |
+
model.load_state_dict(sd, strict=False)
|
40 |
+
model = model.to(shared.device)
|
41 |
+
if half_attention:
|
42 |
+
model = model.half()
|
43 |
+
if shared.cmd_opts.opt_channelslast:
|
44 |
+
model = model.to(memory_format=torch.channels_last)
|
45 |
+
|
46 |
+
sd_hijack.model_hijack.hijack(model) # apply optimization
|
47 |
+
model.eval()
|
48 |
+
|
49 |
+
if shared.opts.ldsr_cached:
|
50 |
+
cached_ldsr_model = model
|
51 |
+
|
52 |
+
return {"model": model}
|
53 |
+
|
54 |
+
def __init__(self, model_path, yaml_path):
|
55 |
+
self.modelPath = model_path
|
56 |
+
self.yamlPath = yaml_path
|
57 |
+
|
58 |
+
@staticmethod
|
59 |
+
def run(model, selected_path, custom_steps, eta):
|
60 |
+
example = get_cond(selected_path)
|
61 |
+
|
62 |
+
n_runs = 1
|
63 |
+
guider = None
|
64 |
+
ckwargs = None
|
65 |
+
ddim_use_x0_pred = False
|
66 |
+
temperature = 1.
|
67 |
+
eta = eta
|
68 |
+
custom_shape = None
|
69 |
+
|
70 |
+
height, width = example["image"].shape[1:3]
|
71 |
+
split_input = height >= 128 and width >= 128
|
72 |
+
|
73 |
+
if split_input:
|
74 |
+
ks = 128
|
75 |
+
stride = 64
|
76 |
+
vqf = 4 #
|
77 |
+
model.split_input_params = {"ks": (ks, ks), "stride": (stride, stride),
|
78 |
+
"vqf": vqf,
|
79 |
+
"patch_distributed_vq": True,
|
80 |
+
"tie_braker": False,
|
81 |
+
"clip_max_weight": 0.5,
|
82 |
+
"clip_min_weight": 0.01,
|
83 |
+
"clip_max_tie_weight": 0.5,
|
84 |
+
"clip_min_tie_weight": 0.01}
|
85 |
+
else:
|
86 |
+
if hasattr(model, "split_input_params"):
|
87 |
+
delattr(model, "split_input_params")
|
88 |
+
|
89 |
+
x_t = None
|
90 |
+
logs = None
|
91 |
+
for _ in range(n_runs):
|
92 |
+
if custom_shape is not None:
|
93 |
+
x_t = torch.randn(1, custom_shape[1], custom_shape[2], custom_shape[3]).to(model.device)
|
94 |
+
x_t = repeat(x_t, '1 c h w -> b c h w', b=custom_shape[0])
|
95 |
+
|
96 |
+
logs = make_convolutional_sample(example, model,
|
97 |
+
custom_steps=custom_steps,
|
98 |
+
eta=eta, quantize_x0=False,
|
99 |
+
custom_shape=custom_shape,
|
100 |
+
temperature=temperature, noise_dropout=0.,
|
101 |
+
corrector=guider, corrector_kwargs=ckwargs, x_T=x_t,
|
102 |
+
ddim_use_x0_pred=ddim_use_x0_pred
|
103 |
+
)
|
104 |
+
return logs
|
105 |
+
|
106 |
+
def super_resolution(self, image, steps=100, target_scale=2, half_attention=False):
|
107 |
+
model = self.load_model_from_config(half_attention)
|
108 |
+
|
109 |
+
# Run settings
|
110 |
+
diffusion_steps = int(steps)
|
111 |
+
eta = 1.0
|
112 |
+
|
113 |
+
|
114 |
+
gc.collect()
|
115 |
+
devices.torch_gc()
|
116 |
+
|
117 |
+
im_og = image
|
118 |
+
width_og, height_og = im_og.size
|
119 |
+
# If we can adjust the max upscale size, then the 4 below should be our variable
|
120 |
+
down_sample_rate = target_scale / 4
|
121 |
+
wd = width_og * down_sample_rate
|
122 |
+
hd = height_og * down_sample_rate
|
123 |
+
width_downsampled_pre = int(np.ceil(wd))
|
124 |
+
height_downsampled_pre = int(np.ceil(hd))
|
125 |
+
|
126 |
+
if down_sample_rate != 1:
|
127 |
+
print(
|
128 |
+
f'Downsampling from [{width_og}, {height_og}] to [{width_downsampled_pre}, {height_downsampled_pre}]')
|
129 |
+
im_og = im_og.resize((width_downsampled_pre, height_downsampled_pre), Image.LANCZOS)
|
130 |
+
else:
|
131 |
+
print(f"Down sample rate is 1 from {target_scale} / 4 (Not downsampling)")
|
132 |
+
|
133 |
+
# pad width and height to multiples of 64, pads with the edge values of image to avoid artifacts
|
134 |
+
pad_w, pad_h = np.max(((2, 2), np.ceil(np.array(im_og.size) / 64).astype(int)), axis=0) * 64 - im_og.size
|
135 |
+
im_padded = Image.fromarray(np.pad(np.array(im_og), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge'))
|
136 |
+
|
137 |
+
logs = self.run(model["model"], im_padded, diffusion_steps, eta)
|
138 |
+
|
139 |
+
sample = logs["sample"]
|
140 |
+
sample = sample.detach().cpu()
|
141 |
+
sample = torch.clamp(sample, -1., 1.)
|
142 |
+
sample = (sample + 1.) / 2. * 255
|
143 |
+
sample = sample.numpy().astype(np.uint8)
|
144 |
+
sample = np.transpose(sample, (0, 2, 3, 1))
|
145 |
+
a = Image.fromarray(sample[0])
|
146 |
+
|
147 |
+
# remove padding
|
148 |
+
a = a.crop((0, 0) + tuple(np.array(im_og.size) * 4))
|
149 |
+
|
150 |
+
del model
|
151 |
+
gc.collect()
|
152 |
+
devices.torch_gc()
|
153 |
+
|
154 |
+
return a
|
155 |
+
|
156 |
+
|
157 |
+
def get_cond(selected_path):
|
158 |
+
example = {}
|
159 |
+
up_f = 4
|
160 |
+
c = selected_path.convert('RGB')
|
161 |
+
c = torch.unsqueeze(torchvision.transforms.ToTensor()(c), 0)
|
162 |
+
c_up = torchvision.transforms.functional.resize(c, size=[up_f * c.shape[2], up_f * c.shape[3]],
|
163 |
+
antialias=True)
|
164 |
+
c_up = rearrange(c_up, '1 c h w -> 1 h w c')
|
165 |
+
c = rearrange(c, '1 c h w -> 1 h w c')
|
166 |
+
c = 2. * c - 1.
|
167 |
+
|
168 |
+
c = c.to(shared.device)
|
169 |
+
example["LR_image"] = c
|
170 |
+
example["image"] = c_up
|
171 |
+
|
172 |
+
return example
|
173 |
+
|
174 |
+
|
175 |
+
@torch.no_grad()
|
176 |
+
def convsample_ddim(model, cond, steps, shape, eta=1.0, callback=None, normals_sequence=None,
|
177 |
+
mask=None, x0=None, quantize_x0=False, temperature=1., score_corrector=None,
|
178 |
+
corrector_kwargs=None, x_t=None
|
179 |
+
):
|
180 |
+
ddim = DDIMSampler(model)
|
181 |
+
bs = shape[0]
|
182 |
+
shape = shape[1:]
|
183 |
+
print(f"Sampling with eta = {eta}; steps: {steps}")
|
184 |
+
samples, intermediates = ddim.sample(steps, batch_size=bs, shape=shape, conditioning=cond, callback=callback,
|
185 |
+
normals_sequence=normals_sequence, quantize_x0=quantize_x0, eta=eta,
|
186 |
+
mask=mask, x0=x0, temperature=temperature, verbose=False,
|
187 |
+
score_corrector=score_corrector,
|
188 |
+
corrector_kwargs=corrector_kwargs, x_t=x_t)
|
189 |
+
|
190 |
+
return samples, intermediates
|
191 |
+
|
192 |
+
|
193 |
+
@torch.no_grad()
|
194 |
+
def make_convolutional_sample(batch, model, custom_steps=None, eta=1.0, quantize_x0=False, custom_shape=None, temperature=1., noise_dropout=0., corrector=None,
|
195 |
+
corrector_kwargs=None, x_T=None, ddim_use_x0_pred=False):
|
196 |
+
log = {}
|
197 |
+
|
198 |
+
z, c, x, xrec, xc = model.get_input(batch, model.first_stage_key,
|
199 |
+
return_first_stage_outputs=True,
|
200 |
+
force_c_encode=not (hasattr(model, 'split_input_params')
|
201 |
+
and model.cond_stage_key == 'coordinates_bbox'),
|
202 |
+
return_original_cond=True)
|
203 |
+
|
204 |
+
if custom_shape is not None:
|
205 |
+
z = torch.randn(custom_shape)
|
206 |
+
print(f"Generating {custom_shape[0]} samples of shape {custom_shape[1:]}")
|
207 |
+
|
208 |
+
z0 = None
|
209 |
+
|
210 |
+
log["input"] = x
|
211 |
+
log["reconstruction"] = xrec
|
212 |
+
|
213 |
+
if ismap(xc):
|
214 |
+
log["original_conditioning"] = model.to_rgb(xc)
|
215 |
+
if hasattr(model, 'cond_stage_key'):
|
216 |
+
log[model.cond_stage_key] = model.to_rgb(xc)
|
217 |
+
|
218 |
+
else:
|
219 |
+
log["original_conditioning"] = xc if xc is not None else torch.zeros_like(x)
|
220 |
+
if model.cond_stage_model:
|
221 |
+
log[model.cond_stage_key] = xc if xc is not None else torch.zeros_like(x)
|
222 |
+
if model.cond_stage_key == 'class_label':
|
223 |
+
log[model.cond_stage_key] = xc[model.cond_stage_key]
|
224 |
+
|
225 |
+
with model.ema_scope("Plotting"):
|
226 |
+
t0 = time.time()
|
227 |
+
|
228 |
+
sample, intermediates = convsample_ddim(model, c, steps=custom_steps, shape=z.shape,
|
229 |
+
eta=eta,
|
230 |
+
quantize_x0=quantize_x0, mask=None, x0=z0,
|
231 |
+
temperature=temperature, score_corrector=corrector, corrector_kwargs=corrector_kwargs,
|
232 |
+
x_t=x_T)
|
233 |
+
t1 = time.time()
|
234 |
+
|
235 |
+
if ddim_use_x0_pred:
|
236 |
+
sample = intermediates['pred_x0'][-1]
|
237 |
+
|
238 |
+
x_sample = model.decode_first_stage(sample)
|
239 |
+
|
240 |
+
try:
|
241 |
+
x_sample_noquant = model.decode_first_stage(sample, force_not_quantize=True)
|
242 |
+
log["sample_noquant"] = x_sample_noquant
|
243 |
+
log["sample_diff"] = torch.abs(x_sample_noquant - x_sample)
|
244 |
+
except Exception:
|
245 |
+
pass
|
246 |
+
|
247 |
+
log["sample"] = x_sample
|
248 |
+
log["time"] = t1 - t0
|
249 |
+
|
250 |
+
return log
|
extensions-builtin/LDSR/preload.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from modules import paths
|
3 |
+
|
4 |
+
|
5 |
+
def preload(parser):
|
6 |
+
parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(paths.models_path, 'LDSR'))
|
extensions-builtin/LDSR/scripts/__pycache__/ldsr_model.cpython-310.pyc
ADDED
Binary file (3.22 kB). View file
|
|
extensions-builtin/LDSR/scripts/ldsr_model.py
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
from modules.modelloader import load_file_from_url
|
4 |
+
from modules.upscaler import Upscaler, UpscalerData
|
5 |
+
from ldsr_model_arch import LDSR
|
6 |
+
from modules import shared, script_callbacks, errors
|
7 |
+
import sd_hijack_autoencoder # noqa: F401
|
8 |
+
import sd_hijack_ddpm_v1 # noqa: F401
|
9 |
+
|
10 |
+
|
11 |
+
class UpscalerLDSR(Upscaler):
|
12 |
+
def __init__(self, user_path):
|
13 |
+
self.name = "LDSR"
|
14 |
+
self.user_path = user_path
|
15 |
+
self.model_url = "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1"
|
16 |
+
self.yaml_url = "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1"
|
17 |
+
super().__init__()
|
18 |
+
scaler_data = UpscalerData("LDSR", None, self)
|
19 |
+
self.scalers = [scaler_data]
|
20 |
+
|
21 |
+
def load_model(self, path: str):
|
22 |
+
# Remove incorrect project.yaml file if too big
|
23 |
+
yaml_path = os.path.join(self.model_path, "project.yaml")
|
24 |
+
old_model_path = os.path.join(self.model_path, "model.pth")
|
25 |
+
new_model_path = os.path.join(self.model_path, "model.ckpt")
|
26 |
+
|
27 |
+
local_model_paths = self.find_models(ext_filter=[".ckpt", ".safetensors"])
|
28 |
+
local_ckpt_path = next(iter([local_model for local_model in local_model_paths if local_model.endswith("model.ckpt")]), None)
|
29 |
+
local_safetensors_path = next(iter([local_model for local_model in local_model_paths if local_model.endswith("model.safetensors")]), None)
|
30 |
+
local_yaml_path = next(iter([local_model for local_model in local_model_paths if local_model.endswith("project.yaml")]), None)
|
31 |
+
|
32 |
+
if os.path.exists(yaml_path):
|
33 |
+
statinfo = os.stat(yaml_path)
|
34 |
+
if statinfo.st_size >= 10485760:
|
35 |
+
print("Removing invalid LDSR YAML file.")
|
36 |
+
os.remove(yaml_path)
|
37 |
+
|
38 |
+
if os.path.exists(old_model_path):
|
39 |
+
print("Renaming model from model.pth to model.ckpt")
|
40 |
+
os.rename(old_model_path, new_model_path)
|
41 |
+
|
42 |
+
if local_safetensors_path is not None and os.path.exists(local_safetensors_path):
|
43 |
+
model = local_safetensors_path
|
44 |
+
else:
|
45 |
+
model = local_ckpt_path or load_file_from_url(self.model_url, model_dir=self.model_download_path, file_name="model.ckpt")
|
46 |
+
|
47 |
+
yaml = local_yaml_path or load_file_from_url(self.yaml_url, model_dir=self.model_download_path, file_name="project.yaml")
|
48 |
+
|
49 |
+
return LDSR(model, yaml)
|
50 |
+
|
51 |
+
def do_upscale(self, img, path):
|
52 |
+
try:
|
53 |
+
ldsr = self.load_model(path)
|
54 |
+
except Exception:
|
55 |
+
errors.report(f"Failed loading LDSR model {path}", exc_info=True)
|
56 |
+
return img
|
57 |
+
ddim_steps = shared.opts.ldsr_steps
|
58 |
+
return ldsr.super_resolution(img, ddim_steps, self.scale)
|
59 |
+
|
60 |
+
|
61 |
+
def on_ui_settings():
|
62 |
+
import gradio as gr
|
63 |
+
|
64 |
+
shared.opts.add_option("ldsr_steps", shared.OptionInfo(100, "LDSR processing steps. Lower = faster", gr.Slider, {"minimum": 1, "maximum": 200, "step": 1}, section=('upscaling', "Upscaling")))
|
65 |
+
shared.opts.add_option("ldsr_cached", shared.OptionInfo(False, "Cache LDSR model in memory", gr.Checkbox, {"interactive": True}, section=('upscaling', "Upscaling")))
|
66 |
+
|
67 |
+
|
68 |
+
script_callbacks.on_ui_settings(on_ui_settings)
|
extensions-builtin/LDSR/sd_hijack_autoencoder.py
ADDED
@@ -0,0 +1,293 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# The content of this file comes from the ldm/models/autoencoder.py file of the compvis/stable-diffusion repo
|
2 |
+
# The VQModel & VQModelInterface were subsequently removed from ldm/models/autoencoder.py when we moved to the stability-ai/stablediffusion repo
|
3 |
+
# As the LDSR upscaler relies on VQModel & VQModelInterface, the hijack aims to put them back into the ldm.models.autoencoder
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
import pytorch_lightning as pl
|
7 |
+
import torch.nn.functional as F
|
8 |
+
from contextlib import contextmanager
|
9 |
+
|
10 |
+
from torch.optim.lr_scheduler import LambdaLR
|
11 |
+
|
12 |
+
from ldm.modules.ema import LitEma
|
13 |
+
from vqvae_quantize import VectorQuantizer2 as VectorQuantizer
|
14 |
+
from ldm.modules.diffusionmodules.model import Encoder, Decoder
|
15 |
+
from ldm.util import instantiate_from_config
|
16 |
+
|
17 |
+
import ldm.models.autoencoder
|
18 |
+
from packaging import version
|
19 |
+
|
20 |
+
class VQModel(pl.LightningModule):
|
21 |
+
def __init__(self,
|
22 |
+
ddconfig,
|
23 |
+
lossconfig,
|
24 |
+
n_embed,
|
25 |
+
embed_dim,
|
26 |
+
ckpt_path=None,
|
27 |
+
ignore_keys=None,
|
28 |
+
image_key="image",
|
29 |
+
colorize_nlabels=None,
|
30 |
+
monitor=None,
|
31 |
+
batch_resize_range=None,
|
32 |
+
scheduler_config=None,
|
33 |
+
lr_g_factor=1.0,
|
34 |
+
remap=None,
|
35 |
+
sane_index_shape=False, # tell vector quantizer to return indices as bhw
|
36 |
+
use_ema=False
|
37 |
+
):
|
38 |
+
super().__init__()
|
39 |
+
self.embed_dim = embed_dim
|
40 |
+
self.n_embed = n_embed
|
41 |
+
self.image_key = image_key
|
42 |
+
self.encoder = Encoder(**ddconfig)
|
43 |
+
self.decoder = Decoder(**ddconfig)
|
44 |
+
self.loss = instantiate_from_config(lossconfig)
|
45 |
+
self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
|
46 |
+
remap=remap,
|
47 |
+
sane_index_shape=sane_index_shape)
|
48 |
+
self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
|
49 |
+
self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
|
50 |
+
if colorize_nlabels is not None:
|
51 |
+
assert type(colorize_nlabels)==int
|
52 |
+
self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
|
53 |
+
if monitor is not None:
|
54 |
+
self.monitor = monitor
|
55 |
+
self.batch_resize_range = batch_resize_range
|
56 |
+
if self.batch_resize_range is not None:
|
57 |
+
print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
|
58 |
+
|
59 |
+
self.use_ema = use_ema
|
60 |
+
if self.use_ema:
|
61 |
+
self.model_ema = LitEma(self)
|
62 |
+
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
63 |
+
|
64 |
+
if ckpt_path is not None:
|
65 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys or [])
|
66 |
+
self.scheduler_config = scheduler_config
|
67 |
+
self.lr_g_factor = lr_g_factor
|
68 |
+
|
69 |
+
@contextmanager
|
70 |
+
def ema_scope(self, context=None):
|
71 |
+
if self.use_ema:
|
72 |
+
self.model_ema.store(self.parameters())
|
73 |
+
self.model_ema.copy_to(self)
|
74 |
+
if context is not None:
|
75 |
+
print(f"{context}: Switched to EMA weights")
|
76 |
+
try:
|
77 |
+
yield None
|
78 |
+
finally:
|
79 |
+
if self.use_ema:
|
80 |
+
self.model_ema.restore(self.parameters())
|
81 |
+
if context is not None:
|
82 |
+
print(f"{context}: Restored training weights")
|
83 |
+
|
84 |
+
def init_from_ckpt(self, path, ignore_keys=None):
|
85 |
+
sd = torch.load(path, map_location="cpu")["state_dict"]
|
86 |
+
keys = list(sd.keys())
|
87 |
+
for k in keys:
|
88 |
+
for ik in ignore_keys or []:
|
89 |
+
if k.startswith(ik):
|
90 |
+
print("Deleting key {} from state_dict.".format(k))
|
91 |
+
del sd[k]
|
92 |
+
missing, unexpected = self.load_state_dict(sd, strict=False)
|
93 |
+
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
94 |
+
if missing:
|
95 |
+
print(f"Missing Keys: {missing}")
|
96 |
+
if unexpected:
|
97 |
+
print(f"Unexpected Keys: {unexpected}")
|
98 |
+
|
99 |
+
def on_train_batch_end(self, *args, **kwargs):
|
100 |
+
if self.use_ema:
|
101 |
+
self.model_ema(self)
|
102 |
+
|
103 |
+
def encode(self, x):
|
104 |
+
h = self.encoder(x)
|
105 |
+
h = self.quant_conv(h)
|
106 |
+
quant, emb_loss, info = self.quantize(h)
|
107 |
+
return quant, emb_loss, info
|
108 |
+
|
109 |
+
def encode_to_prequant(self, x):
|
110 |
+
h = self.encoder(x)
|
111 |
+
h = self.quant_conv(h)
|
112 |
+
return h
|
113 |
+
|
114 |
+
def decode(self, quant):
|
115 |
+
quant = self.post_quant_conv(quant)
|
116 |
+
dec = self.decoder(quant)
|
117 |
+
return dec
|
118 |
+
|
119 |
+
def decode_code(self, code_b):
|
120 |
+
quant_b = self.quantize.embed_code(code_b)
|
121 |
+
dec = self.decode(quant_b)
|
122 |
+
return dec
|
123 |
+
|
124 |
+
def forward(self, input, return_pred_indices=False):
|
125 |
+
quant, diff, (_,_,ind) = self.encode(input)
|
126 |
+
dec = self.decode(quant)
|
127 |
+
if return_pred_indices:
|
128 |
+
return dec, diff, ind
|
129 |
+
return dec, diff
|
130 |
+
|
131 |
+
def get_input(self, batch, k):
|
132 |
+
x = batch[k]
|
133 |
+
if len(x.shape) == 3:
|
134 |
+
x = x[..., None]
|
135 |
+
x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
|
136 |
+
if self.batch_resize_range is not None:
|
137 |
+
lower_size = self.batch_resize_range[0]
|
138 |
+
upper_size = self.batch_resize_range[1]
|
139 |
+
if self.global_step <= 4:
|
140 |
+
# do the first few batches with max size to avoid later oom
|
141 |
+
new_resize = upper_size
|
142 |
+
else:
|
143 |
+
new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
|
144 |
+
if new_resize != x.shape[2]:
|
145 |
+
x = F.interpolate(x, size=new_resize, mode="bicubic")
|
146 |
+
x = x.detach()
|
147 |
+
return x
|
148 |
+
|
149 |
+
def training_step(self, batch, batch_idx, optimizer_idx):
|
150 |
+
# https://github.com/pytorch/pytorch/issues/37142
|
151 |
+
# try not to fool the heuristics
|
152 |
+
x = self.get_input(batch, self.image_key)
|
153 |
+
xrec, qloss, ind = self(x, return_pred_indices=True)
|
154 |
+
|
155 |
+
if optimizer_idx == 0:
|
156 |
+
# autoencode
|
157 |
+
aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
|
158 |
+
last_layer=self.get_last_layer(), split="train",
|
159 |
+
predicted_indices=ind)
|
160 |
+
|
161 |
+
self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
|
162 |
+
return aeloss
|
163 |
+
|
164 |
+
if optimizer_idx == 1:
|
165 |
+
# discriminator
|
166 |
+
discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
|
167 |
+
last_layer=self.get_last_layer(), split="train")
|
168 |
+
self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
|
169 |
+
return discloss
|
170 |
+
|
171 |
+
def validation_step(self, batch, batch_idx):
|
172 |
+
log_dict = self._validation_step(batch, batch_idx)
|
173 |
+
with self.ema_scope():
|
174 |
+
self._validation_step(batch, batch_idx, suffix="_ema")
|
175 |
+
return log_dict
|
176 |
+
|
177 |
+
def _validation_step(self, batch, batch_idx, suffix=""):
|
178 |
+
x = self.get_input(batch, self.image_key)
|
179 |
+
xrec, qloss, ind = self(x, return_pred_indices=True)
|
180 |
+
aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
|
181 |
+
self.global_step,
|
182 |
+
last_layer=self.get_last_layer(),
|
183 |
+
split="val"+suffix,
|
184 |
+
predicted_indices=ind
|
185 |
+
)
|
186 |
+
|
187 |
+
discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
|
188 |
+
self.global_step,
|
189 |
+
last_layer=self.get_last_layer(),
|
190 |
+
split="val"+suffix,
|
191 |
+
predicted_indices=ind
|
192 |
+
)
|
193 |
+
rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
|
194 |
+
self.log(f"val{suffix}/rec_loss", rec_loss,
|
195 |
+
prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
|
196 |
+
self.log(f"val{suffix}/aeloss", aeloss,
|
197 |
+
prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
|
198 |
+
if version.parse(pl.__version__) >= version.parse('1.4.0'):
|
199 |
+
del log_dict_ae[f"val{suffix}/rec_loss"]
|
200 |
+
self.log_dict(log_dict_ae)
|
201 |
+
self.log_dict(log_dict_disc)
|
202 |
+
return self.log_dict
|
203 |
+
|
204 |
+
def configure_optimizers(self):
|
205 |
+
lr_d = self.learning_rate
|
206 |
+
lr_g = self.lr_g_factor*self.learning_rate
|
207 |
+
print("lr_d", lr_d)
|
208 |
+
print("lr_g", lr_g)
|
209 |
+
opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
|
210 |
+
list(self.decoder.parameters())+
|
211 |
+
list(self.quantize.parameters())+
|
212 |
+
list(self.quant_conv.parameters())+
|
213 |
+
list(self.post_quant_conv.parameters()),
|
214 |
+
lr=lr_g, betas=(0.5, 0.9))
|
215 |
+
opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
|
216 |
+
lr=lr_d, betas=(0.5, 0.9))
|
217 |
+
|
218 |
+
if self.scheduler_config is not None:
|
219 |
+
scheduler = instantiate_from_config(self.scheduler_config)
|
220 |
+
|
221 |
+
print("Setting up LambdaLR scheduler...")
|
222 |
+
scheduler = [
|
223 |
+
{
|
224 |
+
'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
|
225 |
+
'interval': 'step',
|
226 |
+
'frequency': 1
|
227 |
+
},
|
228 |
+
{
|
229 |
+
'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
|
230 |
+
'interval': 'step',
|
231 |
+
'frequency': 1
|
232 |
+
},
|
233 |
+
]
|
234 |
+
return [opt_ae, opt_disc], scheduler
|
235 |
+
return [opt_ae, opt_disc], []
|
236 |
+
|
237 |
+
def get_last_layer(self):
|
238 |
+
return self.decoder.conv_out.weight
|
239 |
+
|
240 |
+
def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
|
241 |
+
log = {}
|
242 |
+
x = self.get_input(batch, self.image_key)
|
243 |
+
x = x.to(self.device)
|
244 |
+
if only_inputs:
|
245 |
+
log["inputs"] = x
|
246 |
+
return log
|
247 |
+
xrec, _ = self(x)
|
248 |
+
if x.shape[1] > 3:
|
249 |
+
# colorize with random projection
|
250 |
+
assert xrec.shape[1] > 3
|
251 |
+
x = self.to_rgb(x)
|
252 |
+
xrec = self.to_rgb(xrec)
|
253 |
+
log["inputs"] = x
|
254 |
+
log["reconstructions"] = xrec
|
255 |
+
if plot_ema:
|
256 |
+
with self.ema_scope():
|
257 |
+
xrec_ema, _ = self(x)
|
258 |
+
if x.shape[1] > 3:
|
259 |
+
xrec_ema = self.to_rgb(xrec_ema)
|
260 |
+
log["reconstructions_ema"] = xrec_ema
|
261 |
+
return log
|
262 |
+
|
263 |
+
def to_rgb(self, x):
|
264 |
+
assert self.image_key == "segmentation"
|
265 |
+
if not hasattr(self, "colorize"):
|
266 |
+
self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
|
267 |
+
x = F.conv2d(x, weight=self.colorize)
|
268 |
+
x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
|
269 |
+
return x
|
270 |
+
|
271 |
+
|
272 |
+
class VQModelInterface(VQModel):
|
273 |
+
def __init__(self, embed_dim, *args, **kwargs):
|
274 |
+
super().__init__(*args, embed_dim=embed_dim, **kwargs)
|
275 |
+
self.embed_dim = embed_dim
|
276 |
+
|
277 |
+
def encode(self, x):
|
278 |
+
h = self.encoder(x)
|
279 |
+
h = self.quant_conv(h)
|
280 |
+
return h
|
281 |
+
|
282 |
+
def decode(self, h, force_not_quantize=False):
|
283 |
+
# also go through quantization layer
|
284 |
+
if not force_not_quantize:
|
285 |
+
quant, emb_loss, info = self.quantize(h)
|
286 |
+
else:
|
287 |
+
quant = h
|
288 |
+
quant = self.post_quant_conv(quant)
|
289 |
+
dec = self.decoder(quant)
|
290 |
+
return dec
|
291 |
+
|
292 |
+
ldm.models.autoencoder.VQModel = VQModel
|
293 |
+
ldm.models.autoencoder.VQModelInterface = VQModelInterface
|
extensions-builtin/LDSR/sd_hijack_ddpm_v1.py
ADDED
@@ -0,0 +1,1443 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# This script is copied from the compvis/stable-diffusion repo (aka the SD V1 repo)
|
2 |
+
# Original filename: ldm/models/diffusion/ddpm.py
|
3 |
+
# The purpose to reinstate the old DDPM logic which works with VQ, whereas the V2 one doesn't
|
4 |
+
# Some models such as LDSR require VQ to work correctly
|
5 |
+
# The classes are suffixed with "V1" and added back to the "ldm.models.diffusion.ddpm" module
|
6 |
+
|
7 |
+
import torch
|
8 |
+
import torch.nn as nn
|
9 |
+
import numpy as np
|
10 |
+
import pytorch_lightning as pl
|
11 |
+
from torch.optim.lr_scheduler import LambdaLR
|
12 |
+
from einops import rearrange, repeat
|
13 |
+
from contextlib import contextmanager
|
14 |
+
from functools import partial
|
15 |
+
from tqdm import tqdm
|
16 |
+
from torchvision.utils import make_grid
|
17 |
+
from pytorch_lightning.utilities.distributed import rank_zero_only
|
18 |
+
|
19 |
+
from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
|
20 |
+
from ldm.modules.ema import LitEma
|
21 |
+
from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
|
22 |
+
from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
|
23 |
+
from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
|
24 |
+
from ldm.models.diffusion.ddim import DDIMSampler
|
25 |
+
|
26 |
+
import ldm.models.diffusion.ddpm
|
27 |
+
|
28 |
+
__conditioning_keys__ = {'concat': 'c_concat',
|
29 |
+
'crossattn': 'c_crossattn',
|
30 |
+
'adm': 'y'}
|
31 |
+
|
32 |
+
|
33 |
+
def disabled_train(self, mode=True):
|
34 |
+
"""Overwrite model.train with this function to make sure train/eval mode
|
35 |
+
does not change anymore."""
|
36 |
+
return self
|
37 |
+
|
38 |
+
|
39 |
+
def uniform_on_device(r1, r2, shape, device):
|
40 |
+
return (r1 - r2) * torch.rand(*shape, device=device) + r2
|
41 |
+
|
42 |
+
|
43 |
+
class DDPMV1(pl.LightningModule):
|
44 |
+
# classic DDPM with Gaussian diffusion, in image space
|
45 |
+
def __init__(self,
|
46 |
+
unet_config,
|
47 |
+
timesteps=1000,
|
48 |
+
beta_schedule="linear",
|
49 |
+
loss_type="l2",
|
50 |
+
ckpt_path=None,
|
51 |
+
ignore_keys=None,
|
52 |
+
load_only_unet=False,
|
53 |
+
monitor="val/loss",
|
54 |
+
use_ema=True,
|
55 |
+
first_stage_key="image",
|
56 |
+
image_size=256,
|
57 |
+
channels=3,
|
58 |
+
log_every_t=100,
|
59 |
+
clip_denoised=True,
|
60 |
+
linear_start=1e-4,
|
61 |
+
linear_end=2e-2,
|
62 |
+
cosine_s=8e-3,
|
63 |
+
given_betas=None,
|
64 |
+
original_elbo_weight=0.,
|
65 |
+
v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
|
66 |
+
l_simple_weight=1.,
|
67 |
+
conditioning_key=None,
|
68 |
+
parameterization="eps", # all assuming fixed variance schedules
|
69 |
+
scheduler_config=None,
|
70 |
+
use_positional_encodings=False,
|
71 |
+
learn_logvar=False,
|
72 |
+
logvar_init=0.,
|
73 |
+
):
|
74 |
+
super().__init__()
|
75 |
+
assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
|
76 |
+
self.parameterization = parameterization
|
77 |
+
print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
|
78 |
+
self.cond_stage_model = None
|
79 |
+
self.clip_denoised = clip_denoised
|
80 |
+
self.log_every_t = log_every_t
|
81 |
+
self.first_stage_key = first_stage_key
|
82 |
+
self.image_size = image_size # try conv?
|
83 |
+
self.channels = channels
|
84 |
+
self.use_positional_encodings = use_positional_encodings
|
85 |
+
self.model = DiffusionWrapperV1(unet_config, conditioning_key)
|
86 |
+
count_params(self.model, verbose=True)
|
87 |
+
self.use_ema = use_ema
|
88 |
+
if self.use_ema:
|
89 |
+
self.model_ema = LitEma(self.model)
|
90 |
+
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
91 |
+
|
92 |
+
self.use_scheduler = scheduler_config is not None
|
93 |
+
if self.use_scheduler:
|
94 |
+
self.scheduler_config = scheduler_config
|
95 |
+
|
96 |
+
self.v_posterior = v_posterior
|
97 |
+
self.original_elbo_weight = original_elbo_weight
|
98 |
+
self.l_simple_weight = l_simple_weight
|
99 |
+
|
100 |
+
if monitor is not None:
|
101 |
+
self.monitor = monitor
|
102 |
+
if ckpt_path is not None:
|
103 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys or [], only_model=load_only_unet)
|
104 |
+
|
105 |
+
self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
|
106 |
+
linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
|
107 |
+
|
108 |
+
self.loss_type = loss_type
|
109 |
+
|
110 |
+
self.learn_logvar = learn_logvar
|
111 |
+
self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
|
112 |
+
if self.learn_logvar:
|
113 |
+
self.logvar = nn.Parameter(self.logvar, requires_grad=True)
|
114 |
+
|
115 |
+
|
116 |
+
def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
|
117 |
+
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
118 |
+
if exists(given_betas):
|
119 |
+
betas = given_betas
|
120 |
+
else:
|
121 |
+
betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
|
122 |
+
cosine_s=cosine_s)
|
123 |
+
alphas = 1. - betas
|
124 |
+
alphas_cumprod = np.cumprod(alphas, axis=0)
|
125 |
+
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
|
126 |
+
|
127 |
+
timesteps, = betas.shape
|
128 |
+
self.num_timesteps = int(timesteps)
|
129 |
+
self.linear_start = linear_start
|
130 |
+
self.linear_end = linear_end
|
131 |
+
assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
|
132 |
+
|
133 |
+
to_torch = partial(torch.tensor, dtype=torch.float32)
|
134 |
+
|
135 |
+
self.register_buffer('betas', to_torch(betas))
|
136 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
137 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
|
138 |
+
|
139 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
140 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
|
141 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
|
142 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
|
143 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
|
144 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
|
145 |
+
|
146 |
+
# calculations for posterior q(x_{t-1} | x_t, x_0)
|
147 |
+
posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
|
148 |
+
1. - alphas_cumprod) + self.v_posterior * betas
|
149 |
+
# above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
|
150 |
+
self.register_buffer('posterior_variance', to_torch(posterior_variance))
|
151 |
+
# below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
|
152 |
+
self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
|
153 |
+
self.register_buffer('posterior_mean_coef1', to_torch(
|
154 |
+
betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
|
155 |
+
self.register_buffer('posterior_mean_coef2', to_torch(
|
156 |
+
(1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
|
157 |
+
|
158 |
+
if self.parameterization == "eps":
|
159 |
+
lvlb_weights = self.betas ** 2 / (
|
160 |
+
2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
|
161 |
+
elif self.parameterization == "x0":
|
162 |
+
lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
|
163 |
+
else:
|
164 |
+
raise NotImplementedError("mu not supported")
|
165 |
+
# TODO how to choose this term
|
166 |
+
lvlb_weights[0] = lvlb_weights[1]
|
167 |
+
self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
|
168 |
+
assert not torch.isnan(self.lvlb_weights).all()
|
169 |
+
|
170 |
+
@contextmanager
|
171 |
+
def ema_scope(self, context=None):
|
172 |
+
if self.use_ema:
|
173 |
+
self.model_ema.store(self.model.parameters())
|
174 |
+
self.model_ema.copy_to(self.model)
|
175 |
+
if context is not None:
|
176 |
+
print(f"{context}: Switched to EMA weights")
|
177 |
+
try:
|
178 |
+
yield None
|
179 |
+
finally:
|
180 |
+
if self.use_ema:
|
181 |
+
self.model_ema.restore(self.model.parameters())
|
182 |
+
if context is not None:
|
183 |
+
print(f"{context}: Restored training weights")
|
184 |
+
|
185 |
+
def init_from_ckpt(self, path, ignore_keys=None, only_model=False):
|
186 |
+
sd = torch.load(path, map_location="cpu")
|
187 |
+
if "state_dict" in list(sd.keys()):
|
188 |
+
sd = sd["state_dict"]
|
189 |
+
keys = list(sd.keys())
|
190 |
+
for k in keys:
|
191 |
+
for ik in ignore_keys or []:
|
192 |
+
if k.startswith(ik):
|
193 |
+
print("Deleting key {} from state_dict.".format(k))
|
194 |
+
del sd[k]
|
195 |
+
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
|
196 |
+
sd, strict=False)
|
197 |
+
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
198 |
+
if missing:
|
199 |
+
print(f"Missing Keys: {missing}")
|
200 |
+
if unexpected:
|
201 |
+
print(f"Unexpected Keys: {unexpected}")
|
202 |
+
|
203 |
+
def q_mean_variance(self, x_start, t):
|
204 |
+
"""
|
205 |
+
Get the distribution q(x_t | x_0).
|
206 |
+
:param x_start: the [N x C x ...] tensor of noiseless inputs.
|
207 |
+
:param t: the number of diffusion steps (minus 1). Here, 0 means one step.
|
208 |
+
:return: A tuple (mean, variance, log_variance), all of x_start's shape.
|
209 |
+
"""
|
210 |
+
mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
|
211 |
+
variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
|
212 |
+
log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
|
213 |
+
return mean, variance, log_variance
|
214 |
+
|
215 |
+
def predict_start_from_noise(self, x_t, t, noise):
|
216 |
+
return (
|
217 |
+
extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
|
218 |
+
extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
|
219 |
+
)
|
220 |
+
|
221 |
+
def q_posterior(self, x_start, x_t, t):
|
222 |
+
posterior_mean = (
|
223 |
+
extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
|
224 |
+
extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
|
225 |
+
)
|
226 |
+
posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
|
227 |
+
posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
|
228 |
+
return posterior_mean, posterior_variance, posterior_log_variance_clipped
|
229 |
+
|
230 |
+
def p_mean_variance(self, x, t, clip_denoised: bool):
|
231 |
+
model_out = self.model(x, t)
|
232 |
+
if self.parameterization == "eps":
|
233 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
|
234 |
+
elif self.parameterization == "x0":
|
235 |
+
x_recon = model_out
|
236 |
+
if clip_denoised:
|
237 |
+
x_recon.clamp_(-1., 1.)
|
238 |
+
|
239 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
240 |
+
return model_mean, posterior_variance, posterior_log_variance
|
241 |
+
|
242 |
+
@torch.no_grad()
|
243 |
+
def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
|
244 |
+
b, *_, device = *x.shape, x.device
|
245 |
+
model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
|
246 |
+
noise = noise_like(x.shape, device, repeat_noise)
|
247 |
+
# no noise when t == 0
|
248 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
249 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
250 |
+
|
251 |
+
@torch.no_grad()
|
252 |
+
def p_sample_loop(self, shape, return_intermediates=False):
|
253 |
+
device = self.betas.device
|
254 |
+
b = shape[0]
|
255 |
+
img = torch.randn(shape, device=device)
|
256 |
+
intermediates = [img]
|
257 |
+
for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
|
258 |
+
img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
|
259 |
+
clip_denoised=self.clip_denoised)
|
260 |
+
if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
|
261 |
+
intermediates.append(img)
|
262 |
+
if return_intermediates:
|
263 |
+
return img, intermediates
|
264 |
+
return img
|
265 |
+
|
266 |
+
@torch.no_grad()
|
267 |
+
def sample(self, batch_size=16, return_intermediates=False):
|
268 |
+
image_size = self.image_size
|
269 |
+
channels = self.channels
|
270 |
+
return self.p_sample_loop((batch_size, channels, image_size, image_size),
|
271 |
+
return_intermediates=return_intermediates)
|
272 |
+
|
273 |
+
def q_sample(self, x_start, t, noise=None):
|
274 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
275 |
+
return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
|
276 |
+
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
|
277 |
+
|
278 |
+
def get_loss(self, pred, target, mean=True):
|
279 |
+
if self.loss_type == 'l1':
|
280 |
+
loss = (target - pred).abs()
|
281 |
+
if mean:
|
282 |
+
loss = loss.mean()
|
283 |
+
elif self.loss_type == 'l2':
|
284 |
+
if mean:
|
285 |
+
loss = torch.nn.functional.mse_loss(target, pred)
|
286 |
+
else:
|
287 |
+
loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
|
288 |
+
else:
|
289 |
+
raise NotImplementedError("unknown loss type '{loss_type}'")
|
290 |
+
|
291 |
+
return loss
|
292 |
+
|
293 |
+
def p_losses(self, x_start, t, noise=None):
|
294 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
295 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
296 |
+
model_out = self.model(x_noisy, t)
|
297 |
+
|
298 |
+
loss_dict = {}
|
299 |
+
if self.parameterization == "eps":
|
300 |
+
target = noise
|
301 |
+
elif self.parameterization == "x0":
|
302 |
+
target = x_start
|
303 |
+
else:
|
304 |
+
raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
|
305 |
+
|
306 |
+
loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
|
307 |
+
|
308 |
+
log_prefix = 'train' if self.training else 'val'
|
309 |
+
|
310 |
+
loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
|
311 |
+
loss_simple = loss.mean() * self.l_simple_weight
|
312 |
+
|
313 |
+
loss_vlb = (self.lvlb_weights[t] * loss).mean()
|
314 |
+
loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
|
315 |
+
|
316 |
+
loss = loss_simple + self.original_elbo_weight * loss_vlb
|
317 |
+
|
318 |
+
loss_dict.update({f'{log_prefix}/loss': loss})
|
319 |
+
|
320 |
+
return loss, loss_dict
|
321 |
+
|
322 |
+
def forward(self, x, *args, **kwargs):
|
323 |
+
# b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
|
324 |
+
# assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
|
325 |
+
t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
|
326 |
+
return self.p_losses(x, t, *args, **kwargs)
|
327 |
+
|
328 |
+
def get_input(self, batch, k):
|
329 |
+
x = batch[k]
|
330 |
+
if len(x.shape) == 3:
|
331 |
+
x = x[..., None]
|
332 |
+
x = rearrange(x, 'b h w c -> b c h w')
|
333 |
+
x = x.to(memory_format=torch.contiguous_format).float()
|
334 |
+
return x
|
335 |
+
|
336 |
+
def shared_step(self, batch):
|
337 |
+
x = self.get_input(batch, self.first_stage_key)
|
338 |
+
loss, loss_dict = self(x)
|
339 |
+
return loss, loss_dict
|
340 |
+
|
341 |
+
def training_step(self, batch, batch_idx):
|
342 |
+
loss, loss_dict = self.shared_step(batch)
|
343 |
+
|
344 |
+
self.log_dict(loss_dict, prog_bar=True,
|
345 |
+
logger=True, on_step=True, on_epoch=True)
|
346 |
+
|
347 |
+
self.log("global_step", self.global_step,
|
348 |
+
prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
349 |
+
|
350 |
+
if self.use_scheduler:
|
351 |
+
lr = self.optimizers().param_groups[0]['lr']
|
352 |
+
self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
353 |
+
|
354 |
+
return loss
|
355 |
+
|
356 |
+
@torch.no_grad()
|
357 |
+
def validation_step(self, batch, batch_idx):
|
358 |
+
_, loss_dict_no_ema = self.shared_step(batch)
|
359 |
+
with self.ema_scope():
|
360 |
+
_, loss_dict_ema = self.shared_step(batch)
|
361 |
+
loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
|
362 |
+
self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
|
363 |
+
self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
|
364 |
+
|
365 |
+
def on_train_batch_end(self, *args, **kwargs):
|
366 |
+
if self.use_ema:
|
367 |
+
self.model_ema(self.model)
|
368 |
+
|
369 |
+
def _get_rows_from_list(self, samples):
|
370 |
+
n_imgs_per_row = len(samples)
|
371 |
+
denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
|
372 |
+
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
|
373 |
+
denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
|
374 |
+
return denoise_grid
|
375 |
+
|
376 |
+
@torch.no_grad()
|
377 |
+
def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
|
378 |
+
log = {}
|
379 |
+
x = self.get_input(batch, self.first_stage_key)
|
380 |
+
N = min(x.shape[0], N)
|
381 |
+
n_row = min(x.shape[0], n_row)
|
382 |
+
x = x.to(self.device)[:N]
|
383 |
+
log["inputs"] = x
|
384 |
+
|
385 |
+
# get diffusion row
|
386 |
+
diffusion_row = []
|
387 |
+
x_start = x[:n_row]
|
388 |
+
|
389 |
+
for t in range(self.num_timesteps):
|
390 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
391 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
392 |
+
t = t.to(self.device).long()
|
393 |
+
noise = torch.randn_like(x_start)
|
394 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
395 |
+
diffusion_row.append(x_noisy)
|
396 |
+
|
397 |
+
log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
|
398 |
+
|
399 |
+
if sample:
|
400 |
+
# get denoise row
|
401 |
+
with self.ema_scope("Plotting"):
|
402 |
+
samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
|
403 |
+
|
404 |
+
log["samples"] = samples
|
405 |
+
log["denoise_row"] = self._get_rows_from_list(denoise_row)
|
406 |
+
|
407 |
+
if return_keys:
|
408 |
+
if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
|
409 |
+
return log
|
410 |
+
else:
|
411 |
+
return {key: log[key] for key in return_keys}
|
412 |
+
return log
|
413 |
+
|
414 |
+
def configure_optimizers(self):
|
415 |
+
lr = self.learning_rate
|
416 |
+
params = list(self.model.parameters())
|
417 |
+
if self.learn_logvar:
|
418 |
+
params = params + [self.logvar]
|
419 |
+
opt = torch.optim.AdamW(params, lr=lr)
|
420 |
+
return opt
|
421 |
+
|
422 |
+
|
423 |
+
class LatentDiffusionV1(DDPMV1):
|
424 |
+
"""main class"""
|
425 |
+
def __init__(self,
|
426 |
+
first_stage_config,
|
427 |
+
cond_stage_config,
|
428 |
+
num_timesteps_cond=None,
|
429 |
+
cond_stage_key="image",
|
430 |
+
cond_stage_trainable=False,
|
431 |
+
concat_mode=True,
|
432 |
+
cond_stage_forward=None,
|
433 |
+
conditioning_key=None,
|
434 |
+
scale_factor=1.0,
|
435 |
+
scale_by_std=False,
|
436 |
+
*args, **kwargs):
|
437 |
+
self.num_timesteps_cond = default(num_timesteps_cond, 1)
|
438 |
+
self.scale_by_std = scale_by_std
|
439 |
+
assert self.num_timesteps_cond <= kwargs['timesteps']
|
440 |
+
# for backwards compatibility after implementation of DiffusionWrapper
|
441 |
+
if conditioning_key is None:
|
442 |
+
conditioning_key = 'concat' if concat_mode else 'crossattn'
|
443 |
+
if cond_stage_config == '__is_unconditional__':
|
444 |
+
conditioning_key = None
|
445 |
+
ckpt_path = kwargs.pop("ckpt_path", None)
|
446 |
+
ignore_keys = kwargs.pop("ignore_keys", [])
|
447 |
+
super().__init__(*args, conditioning_key=conditioning_key, **kwargs)
|
448 |
+
self.concat_mode = concat_mode
|
449 |
+
self.cond_stage_trainable = cond_stage_trainable
|
450 |
+
self.cond_stage_key = cond_stage_key
|
451 |
+
try:
|
452 |
+
self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
|
453 |
+
except Exception:
|
454 |
+
self.num_downs = 0
|
455 |
+
if not scale_by_std:
|
456 |
+
self.scale_factor = scale_factor
|
457 |
+
else:
|
458 |
+
self.register_buffer('scale_factor', torch.tensor(scale_factor))
|
459 |
+
self.instantiate_first_stage(first_stage_config)
|
460 |
+
self.instantiate_cond_stage(cond_stage_config)
|
461 |
+
self.cond_stage_forward = cond_stage_forward
|
462 |
+
self.clip_denoised = False
|
463 |
+
self.bbox_tokenizer = None
|
464 |
+
|
465 |
+
self.restarted_from_ckpt = False
|
466 |
+
if ckpt_path is not None:
|
467 |
+
self.init_from_ckpt(ckpt_path, ignore_keys)
|
468 |
+
self.restarted_from_ckpt = True
|
469 |
+
|
470 |
+
def make_cond_schedule(self, ):
|
471 |
+
self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
|
472 |
+
ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
|
473 |
+
self.cond_ids[:self.num_timesteps_cond] = ids
|
474 |
+
|
475 |
+
@rank_zero_only
|
476 |
+
@torch.no_grad()
|
477 |
+
def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
|
478 |
+
# only for very first batch
|
479 |
+
if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
|
480 |
+
assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
|
481 |
+
# set rescale weight to 1./std of encodings
|
482 |
+
print("### USING STD-RESCALING ###")
|
483 |
+
x = super().get_input(batch, self.first_stage_key)
|
484 |
+
x = x.to(self.device)
|
485 |
+
encoder_posterior = self.encode_first_stage(x)
|
486 |
+
z = self.get_first_stage_encoding(encoder_posterior).detach()
|
487 |
+
del self.scale_factor
|
488 |
+
self.register_buffer('scale_factor', 1. / z.flatten().std())
|
489 |
+
print(f"setting self.scale_factor to {self.scale_factor}")
|
490 |
+
print("### USING STD-RESCALING ###")
|
491 |
+
|
492 |
+
def register_schedule(self,
|
493 |
+
given_betas=None, beta_schedule="linear", timesteps=1000,
|
494 |
+
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
495 |
+
super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
|
496 |
+
|
497 |
+
self.shorten_cond_schedule = self.num_timesteps_cond > 1
|
498 |
+
if self.shorten_cond_schedule:
|
499 |
+
self.make_cond_schedule()
|
500 |
+
|
501 |
+
def instantiate_first_stage(self, config):
|
502 |
+
model = instantiate_from_config(config)
|
503 |
+
self.first_stage_model = model.eval()
|
504 |
+
self.first_stage_model.train = disabled_train
|
505 |
+
for param in self.first_stage_model.parameters():
|
506 |
+
param.requires_grad = False
|
507 |
+
|
508 |
+
def instantiate_cond_stage(self, config):
|
509 |
+
if not self.cond_stage_trainable:
|
510 |
+
if config == "__is_first_stage__":
|
511 |
+
print("Using first stage also as cond stage.")
|
512 |
+
self.cond_stage_model = self.first_stage_model
|
513 |
+
elif config == "__is_unconditional__":
|
514 |
+
print(f"Training {self.__class__.__name__} as an unconditional model.")
|
515 |
+
self.cond_stage_model = None
|
516 |
+
# self.be_unconditional = True
|
517 |
+
else:
|
518 |
+
model = instantiate_from_config(config)
|
519 |
+
self.cond_stage_model = model.eval()
|
520 |
+
self.cond_stage_model.train = disabled_train
|
521 |
+
for param in self.cond_stage_model.parameters():
|
522 |
+
param.requires_grad = False
|
523 |
+
else:
|
524 |
+
assert config != '__is_first_stage__'
|
525 |
+
assert config != '__is_unconditional__'
|
526 |
+
model = instantiate_from_config(config)
|
527 |
+
self.cond_stage_model = model
|
528 |
+
|
529 |
+
def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
|
530 |
+
denoise_row = []
|
531 |
+
for zd in tqdm(samples, desc=desc):
|
532 |
+
denoise_row.append(self.decode_first_stage(zd.to(self.device),
|
533 |
+
force_not_quantize=force_no_decoder_quantization))
|
534 |
+
n_imgs_per_row = len(denoise_row)
|
535 |
+
denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W
|
536 |
+
denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
|
537 |
+
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
|
538 |
+
denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
|
539 |
+
return denoise_grid
|
540 |
+
|
541 |
+
def get_first_stage_encoding(self, encoder_posterior):
|
542 |
+
if isinstance(encoder_posterior, DiagonalGaussianDistribution):
|
543 |
+
z = encoder_posterior.sample()
|
544 |
+
elif isinstance(encoder_posterior, torch.Tensor):
|
545 |
+
z = encoder_posterior
|
546 |
+
else:
|
547 |
+
raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
|
548 |
+
return self.scale_factor * z
|
549 |
+
|
550 |
+
def get_learned_conditioning(self, c):
|
551 |
+
if self.cond_stage_forward is None:
|
552 |
+
if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
|
553 |
+
c = self.cond_stage_model.encode(c)
|
554 |
+
if isinstance(c, DiagonalGaussianDistribution):
|
555 |
+
c = c.mode()
|
556 |
+
else:
|
557 |
+
c = self.cond_stage_model(c)
|
558 |
+
else:
|
559 |
+
assert hasattr(self.cond_stage_model, self.cond_stage_forward)
|
560 |
+
c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
|
561 |
+
return c
|
562 |
+
|
563 |
+
def meshgrid(self, h, w):
|
564 |
+
y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
|
565 |
+
x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
|
566 |
+
|
567 |
+
arr = torch.cat([y, x], dim=-1)
|
568 |
+
return arr
|
569 |
+
|
570 |
+
def delta_border(self, h, w):
|
571 |
+
"""
|
572 |
+
:param h: height
|
573 |
+
:param w: width
|
574 |
+
:return: normalized distance to image border,
|
575 |
+
wtith min distance = 0 at border and max dist = 0.5 at image center
|
576 |
+
"""
|
577 |
+
lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
|
578 |
+
arr = self.meshgrid(h, w) / lower_right_corner
|
579 |
+
dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
|
580 |
+
dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
|
581 |
+
edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
|
582 |
+
return edge_dist
|
583 |
+
|
584 |
+
def get_weighting(self, h, w, Ly, Lx, device):
|
585 |
+
weighting = self.delta_border(h, w)
|
586 |
+
weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
|
587 |
+
self.split_input_params["clip_max_weight"], )
|
588 |
+
weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
|
589 |
+
|
590 |
+
if self.split_input_params["tie_braker"]:
|
591 |
+
L_weighting = self.delta_border(Ly, Lx)
|
592 |
+
L_weighting = torch.clip(L_weighting,
|
593 |
+
self.split_input_params["clip_min_tie_weight"],
|
594 |
+
self.split_input_params["clip_max_tie_weight"])
|
595 |
+
|
596 |
+
L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
|
597 |
+
weighting = weighting * L_weighting
|
598 |
+
return weighting
|
599 |
+
|
600 |
+
def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code
|
601 |
+
"""
|
602 |
+
:param x: img of size (bs, c, h, w)
|
603 |
+
:return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
|
604 |
+
"""
|
605 |
+
bs, nc, h, w = x.shape
|
606 |
+
|
607 |
+
# number of crops in image
|
608 |
+
Ly = (h - kernel_size[0]) // stride[0] + 1
|
609 |
+
Lx = (w - kernel_size[1]) // stride[1] + 1
|
610 |
+
|
611 |
+
if uf == 1 and df == 1:
|
612 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
613 |
+
unfold = torch.nn.Unfold(**fold_params)
|
614 |
+
|
615 |
+
fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
|
616 |
+
|
617 |
+
weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
|
618 |
+
normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap
|
619 |
+
weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
|
620 |
+
|
621 |
+
elif uf > 1 and df == 1:
|
622 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
623 |
+
unfold = torch.nn.Unfold(**fold_params)
|
624 |
+
|
625 |
+
fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
|
626 |
+
dilation=1, padding=0,
|
627 |
+
stride=(stride[0] * uf, stride[1] * uf))
|
628 |
+
fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
|
629 |
+
|
630 |
+
weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
|
631 |
+
normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap
|
632 |
+
weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
|
633 |
+
|
634 |
+
elif df > 1 and uf == 1:
|
635 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
636 |
+
unfold = torch.nn.Unfold(**fold_params)
|
637 |
+
|
638 |
+
fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
|
639 |
+
dilation=1, padding=0,
|
640 |
+
stride=(stride[0] // df, stride[1] // df))
|
641 |
+
fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
|
642 |
+
|
643 |
+
weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
|
644 |
+
normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap
|
645 |
+
weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
|
646 |
+
|
647 |
+
else:
|
648 |
+
raise NotImplementedError
|
649 |
+
|
650 |
+
return fold, unfold, normalization, weighting
|
651 |
+
|
652 |
+
@torch.no_grad()
|
653 |
+
def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
|
654 |
+
cond_key=None, return_original_cond=False, bs=None):
|
655 |
+
x = super().get_input(batch, k)
|
656 |
+
if bs is not None:
|
657 |
+
x = x[:bs]
|
658 |
+
x = x.to(self.device)
|
659 |
+
encoder_posterior = self.encode_first_stage(x)
|
660 |
+
z = self.get_first_stage_encoding(encoder_posterior).detach()
|
661 |
+
|
662 |
+
if self.model.conditioning_key is not None:
|
663 |
+
if cond_key is None:
|
664 |
+
cond_key = self.cond_stage_key
|
665 |
+
if cond_key != self.first_stage_key:
|
666 |
+
if cond_key in ['caption', 'coordinates_bbox']:
|
667 |
+
xc = batch[cond_key]
|
668 |
+
elif cond_key == 'class_label':
|
669 |
+
xc = batch
|
670 |
+
else:
|
671 |
+
xc = super().get_input(batch, cond_key).to(self.device)
|
672 |
+
else:
|
673 |
+
xc = x
|
674 |
+
if not self.cond_stage_trainable or force_c_encode:
|
675 |
+
if isinstance(xc, dict) or isinstance(xc, list):
|
676 |
+
# import pudb; pudb.set_trace()
|
677 |
+
c = self.get_learned_conditioning(xc)
|
678 |
+
else:
|
679 |
+
c = self.get_learned_conditioning(xc.to(self.device))
|
680 |
+
else:
|
681 |
+
c = xc
|
682 |
+
if bs is not None:
|
683 |
+
c = c[:bs]
|
684 |
+
|
685 |
+
if self.use_positional_encodings:
|
686 |
+
pos_x, pos_y = self.compute_latent_shifts(batch)
|
687 |
+
ckey = __conditioning_keys__[self.model.conditioning_key]
|
688 |
+
c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
|
689 |
+
|
690 |
+
else:
|
691 |
+
c = None
|
692 |
+
xc = None
|
693 |
+
if self.use_positional_encodings:
|
694 |
+
pos_x, pos_y = self.compute_latent_shifts(batch)
|
695 |
+
c = {'pos_x': pos_x, 'pos_y': pos_y}
|
696 |
+
out = [z, c]
|
697 |
+
if return_first_stage_outputs:
|
698 |
+
xrec = self.decode_first_stage(z)
|
699 |
+
out.extend([x, xrec])
|
700 |
+
if return_original_cond:
|
701 |
+
out.append(xc)
|
702 |
+
return out
|
703 |
+
|
704 |
+
@torch.no_grad()
|
705 |
+
def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
|
706 |
+
if predict_cids:
|
707 |
+
if z.dim() == 4:
|
708 |
+
z = torch.argmax(z.exp(), dim=1).long()
|
709 |
+
z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
|
710 |
+
z = rearrange(z, 'b h w c -> b c h w').contiguous()
|
711 |
+
|
712 |
+
z = 1. / self.scale_factor * z
|
713 |
+
|
714 |
+
if hasattr(self, "split_input_params"):
|
715 |
+
if self.split_input_params["patch_distributed_vq"]:
|
716 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
717 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
718 |
+
uf = self.split_input_params["vqf"]
|
719 |
+
bs, nc, h, w = z.shape
|
720 |
+
if ks[0] > h or ks[1] > w:
|
721 |
+
ks = (min(ks[0], h), min(ks[1], w))
|
722 |
+
print("reducing Kernel")
|
723 |
+
|
724 |
+
if stride[0] > h or stride[1] > w:
|
725 |
+
stride = (min(stride[0], h), min(stride[1], w))
|
726 |
+
print("reducing stride")
|
727 |
+
|
728 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
|
729 |
+
|
730 |
+
z = unfold(z) # (bn, nc * prod(**ks), L)
|
731 |
+
# 1. Reshape to img shape
|
732 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
733 |
+
|
734 |
+
# 2. apply model loop over last dim
|
735 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
736 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
|
737 |
+
force_not_quantize=predict_cids or force_not_quantize)
|
738 |
+
for i in range(z.shape[-1])]
|
739 |
+
else:
|
740 |
+
|
741 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
|
742 |
+
for i in range(z.shape[-1])]
|
743 |
+
|
744 |
+
o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
|
745 |
+
o = o * weighting
|
746 |
+
# Reverse 1. reshape to img shape
|
747 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
748 |
+
# stitch crops together
|
749 |
+
decoded = fold(o)
|
750 |
+
decoded = decoded / normalization # norm is shape (1, 1, h, w)
|
751 |
+
return decoded
|
752 |
+
else:
|
753 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
754 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
755 |
+
else:
|
756 |
+
return self.first_stage_model.decode(z)
|
757 |
+
|
758 |
+
else:
|
759 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
760 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
761 |
+
else:
|
762 |
+
return self.first_stage_model.decode(z)
|
763 |
+
|
764 |
+
# same as above but without decorator
|
765 |
+
def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
|
766 |
+
if predict_cids:
|
767 |
+
if z.dim() == 4:
|
768 |
+
z = torch.argmax(z.exp(), dim=1).long()
|
769 |
+
z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
|
770 |
+
z = rearrange(z, 'b h w c -> b c h w').contiguous()
|
771 |
+
|
772 |
+
z = 1. / self.scale_factor * z
|
773 |
+
|
774 |
+
if hasattr(self, "split_input_params"):
|
775 |
+
if self.split_input_params["patch_distributed_vq"]:
|
776 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
777 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
778 |
+
uf = self.split_input_params["vqf"]
|
779 |
+
bs, nc, h, w = z.shape
|
780 |
+
if ks[0] > h or ks[1] > w:
|
781 |
+
ks = (min(ks[0], h), min(ks[1], w))
|
782 |
+
print("reducing Kernel")
|
783 |
+
|
784 |
+
if stride[0] > h or stride[1] > w:
|
785 |
+
stride = (min(stride[0], h), min(stride[1], w))
|
786 |
+
print("reducing stride")
|
787 |
+
|
788 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
|
789 |
+
|
790 |
+
z = unfold(z) # (bn, nc * prod(**ks), L)
|
791 |
+
# 1. Reshape to img shape
|
792 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
793 |
+
|
794 |
+
# 2. apply model loop over last dim
|
795 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
796 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
|
797 |
+
force_not_quantize=predict_cids or force_not_quantize)
|
798 |
+
for i in range(z.shape[-1])]
|
799 |
+
else:
|
800 |
+
|
801 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
|
802 |
+
for i in range(z.shape[-1])]
|
803 |
+
|
804 |
+
o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
|
805 |
+
o = o * weighting
|
806 |
+
# Reverse 1. reshape to img shape
|
807 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
808 |
+
# stitch crops together
|
809 |
+
decoded = fold(o)
|
810 |
+
decoded = decoded / normalization # norm is shape (1, 1, h, w)
|
811 |
+
return decoded
|
812 |
+
else:
|
813 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
814 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
815 |
+
else:
|
816 |
+
return self.first_stage_model.decode(z)
|
817 |
+
|
818 |
+
else:
|
819 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
820 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
821 |
+
else:
|
822 |
+
return self.first_stage_model.decode(z)
|
823 |
+
|
824 |
+
@torch.no_grad()
|
825 |
+
def encode_first_stage(self, x):
|
826 |
+
if hasattr(self, "split_input_params"):
|
827 |
+
if self.split_input_params["patch_distributed_vq"]:
|
828 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
829 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
830 |
+
df = self.split_input_params["vqf"]
|
831 |
+
self.split_input_params['original_image_size'] = x.shape[-2:]
|
832 |
+
bs, nc, h, w = x.shape
|
833 |
+
if ks[0] > h or ks[1] > w:
|
834 |
+
ks = (min(ks[0], h), min(ks[1], w))
|
835 |
+
print("reducing Kernel")
|
836 |
+
|
837 |
+
if stride[0] > h or stride[1] > w:
|
838 |
+
stride = (min(stride[0], h), min(stride[1], w))
|
839 |
+
print("reducing stride")
|
840 |
+
|
841 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
|
842 |
+
z = unfold(x) # (bn, nc * prod(**ks), L)
|
843 |
+
# Reshape to img shape
|
844 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
845 |
+
|
846 |
+
output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
|
847 |
+
for i in range(z.shape[-1])]
|
848 |
+
|
849 |
+
o = torch.stack(output_list, axis=-1)
|
850 |
+
o = o * weighting
|
851 |
+
|
852 |
+
# Reverse reshape to img shape
|
853 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
854 |
+
# stitch crops together
|
855 |
+
decoded = fold(o)
|
856 |
+
decoded = decoded / normalization
|
857 |
+
return decoded
|
858 |
+
|
859 |
+
else:
|
860 |
+
return self.first_stage_model.encode(x)
|
861 |
+
else:
|
862 |
+
return self.first_stage_model.encode(x)
|
863 |
+
|
864 |
+
def shared_step(self, batch, **kwargs):
|
865 |
+
x, c = self.get_input(batch, self.first_stage_key)
|
866 |
+
loss = self(x, c)
|
867 |
+
return loss
|
868 |
+
|
869 |
+
def forward(self, x, c, *args, **kwargs):
|
870 |
+
t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
|
871 |
+
if self.model.conditioning_key is not None:
|
872 |
+
assert c is not None
|
873 |
+
if self.cond_stage_trainable:
|
874 |
+
c = self.get_learned_conditioning(c)
|
875 |
+
if self.shorten_cond_schedule: # TODO: drop this option
|
876 |
+
tc = self.cond_ids[t].to(self.device)
|
877 |
+
c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
|
878 |
+
return self.p_losses(x, c, t, *args, **kwargs)
|
879 |
+
|
880 |
+
def apply_model(self, x_noisy, t, cond, return_ids=False):
|
881 |
+
|
882 |
+
if isinstance(cond, dict):
|
883 |
+
# hybrid case, cond is exptected to be a dict
|
884 |
+
pass
|
885 |
+
else:
|
886 |
+
if not isinstance(cond, list):
|
887 |
+
cond = [cond]
|
888 |
+
key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
|
889 |
+
cond = {key: cond}
|
890 |
+
|
891 |
+
if hasattr(self, "split_input_params"):
|
892 |
+
assert len(cond) == 1 # todo can only deal with one conditioning atm
|
893 |
+
assert not return_ids
|
894 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
895 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
896 |
+
|
897 |
+
h, w = x_noisy.shape[-2:]
|
898 |
+
|
899 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
|
900 |
+
|
901 |
+
z = unfold(x_noisy) # (bn, nc * prod(**ks), L)
|
902 |
+
# Reshape to img shape
|
903 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
904 |
+
z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
|
905 |
+
|
906 |
+
if self.cond_stage_key in ["image", "LR_image", "segmentation",
|
907 |
+
'bbox_img'] and self.model.conditioning_key: # todo check for completeness
|
908 |
+
c_key = next(iter(cond.keys())) # get key
|
909 |
+
c = next(iter(cond.values())) # get value
|
910 |
+
assert (len(c) == 1) # todo extend to list with more than one elem
|
911 |
+
c = c[0] # get element
|
912 |
+
|
913 |
+
c = unfold(c)
|
914 |
+
c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
915 |
+
|
916 |
+
cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
|
917 |
+
|
918 |
+
elif self.cond_stage_key == 'coordinates_bbox':
|
919 |
+
assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
|
920 |
+
|
921 |
+
# assuming padding of unfold is always 0 and its dilation is always 1
|
922 |
+
n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
|
923 |
+
full_img_h, full_img_w = self.split_input_params['original_image_size']
|
924 |
+
# as we are operating on latents, we need the factor from the original image size to the
|
925 |
+
# spatial latent size to properly rescale the crops for regenerating the bbox annotations
|
926 |
+
num_downs = self.first_stage_model.encoder.num_resolutions - 1
|
927 |
+
rescale_latent = 2 ** (num_downs)
|
928 |
+
|
929 |
+
# get top left postions of patches as conforming for the bbbox tokenizer, therefore we
|
930 |
+
# need to rescale the tl patch coordinates to be in between (0,1)
|
931 |
+
tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
|
932 |
+
rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
|
933 |
+
for patch_nr in range(z.shape[-1])]
|
934 |
+
|
935 |
+
# patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
|
936 |
+
patch_limits = [(x_tl, y_tl,
|
937 |
+
rescale_latent * ks[0] / full_img_w,
|
938 |
+
rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
|
939 |
+
# patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
|
940 |
+
|
941 |
+
# tokenize crop coordinates for the bounding boxes of the respective patches
|
942 |
+
patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
|
943 |
+
for bbox in patch_limits] # list of length l with tensors of shape (1, 2)
|
944 |
+
print(patch_limits_tknzd[0].shape)
|
945 |
+
# cut tknzd crop position from conditioning
|
946 |
+
assert isinstance(cond, dict), 'cond must be dict to be fed into model'
|
947 |
+
cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
|
948 |
+
print(cut_cond.shape)
|
949 |
+
|
950 |
+
adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
|
951 |
+
adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
|
952 |
+
print(adapted_cond.shape)
|
953 |
+
adapted_cond = self.get_learned_conditioning(adapted_cond)
|
954 |
+
print(adapted_cond.shape)
|
955 |
+
adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
|
956 |
+
print(adapted_cond.shape)
|
957 |
+
|
958 |
+
cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
|
959 |
+
|
960 |
+
else:
|
961 |
+
cond_list = [cond for i in range(z.shape[-1])] # Todo make this more efficient
|
962 |
+
|
963 |
+
# apply model by loop over crops
|
964 |
+
output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
|
965 |
+
assert not isinstance(output_list[0],
|
966 |
+
tuple) # todo cant deal with multiple model outputs check this never happens
|
967 |
+
|
968 |
+
o = torch.stack(output_list, axis=-1)
|
969 |
+
o = o * weighting
|
970 |
+
# Reverse reshape to img shape
|
971 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
972 |
+
# stitch crops together
|
973 |
+
x_recon = fold(o) / normalization
|
974 |
+
|
975 |
+
else:
|
976 |
+
x_recon = self.model(x_noisy, t, **cond)
|
977 |
+
|
978 |
+
if isinstance(x_recon, tuple) and not return_ids:
|
979 |
+
return x_recon[0]
|
980 |
+
else:
|
981 |
+
return x_recon
|
982 |
+
|
983 |
+
def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
|
984 |
+
return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
|
985 |
+
extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
|
986 |
+
|
987 |
+
def _prior_bpd(self, x_start):
|
988 |
+
"""
|
989 |
+
Get the prior KL term for the variational lower-bound, measured in
|
990 |
+
bits-per-dim.
|
991 |
+
This term can't be optimized, as it only depends on the encoder.
|
992 |
+
:param x_start: the [N x C x ...] tensor of inputs.
|
993 |
+
:return: a batch of [N] KL values (in bits), one per batch element.
|
994 |
+
"""
|
995 |
+
batch_size = x_start.shape[0]
|
996 |
+
t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
|
997 |
+
qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
|
998 |
+
kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
|
999 |
+
return mean_flat(kl_prior) / np.log(2.0)
|
1000 |
+
|
1001 |
+
def p_losses(self, x_start, cond, t, noise=None):
|
1002 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
1003 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
1004 |
+
model_output = self.apply_model(x_noisy, t, cond)
|
1005 |
+
|
1006 |
+
loss_dict = {}
|
1007 |
+
prefix = 'train' if self.training else 'val'
|
1008 |
+
|
1009 |
+
if self.parameterization == "x0":
|
1010 |
+
target = x_start
|
1011 |
+
elif self.parameterization == "eps":
|
1012 |
+
target = noise
|
1013 |
+
else:
|
1014 |
+
raise NotImplementedError()
|
1015 |
+
|
1016 |
+
loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
|
1017 |
+
loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
|
1018 |
+
|
1019 |
+
logvar_t = self.logvar[t].to(self.device)
|
1020 |
+
loss = loss_simple / torch.exp(logvar_t) + logvar_t
|
1021 |
+
# loss = loss_simple / torch.exp(self.logvar) + self.logvar
|
1022 |
+
if self.learn_logvar:
|
1023 |
+
loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
|
1024 |
+
loss_dict.update({'logvar': self.logvar.data.mean()})
|
1025 |
+
|
1026 |
+
loss = self.l_simple_weight * loss.mean()
|
1027 |
+
|
1028 |
+
loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
|
1029 |
+
loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
|
1030 |
+
loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
|
1031 |
+
loss += (self.original_elbo_weight * loss_vlb)
|
1032 |
+
loss_dict.update({f'{prefix}/loss': loss})
|
1033 |
+
|
1034 |
+
return loss, loss_dict
|
1035 |
+
|
1036 |
+
def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
|
1037 |
+
return_x0=False, score_corrector=None, corrector_kwargs=None):
|
1038 |
+
t_in = t
|
1039 |
+
model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
|
1040 |
+
|
1041 |
+
if score_corrector is not None:
|
1042 |
+
assert self.parameterization == "eps"
|
1043 |
+
model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
|
1044 |
+
|
1045 |
+
if return_codebook_ids:
|
1046 |
+
model_out, logits = model_out
|
1047 |
+
|
1048 |
+
if self.parameterization == "eps":
|
1049 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
|
1050 |
+
elif self.parameterization == "x0":
|
1051 |
+
x_recon = model_out
|
1052 |
+
else:
|
1053 |
+
raise NotImplementedError()
|
1054 |
+
|
1055 |
+
if clip_denoised:
|
1056 |
+
x_recon.clamp_(-1., 1.)
|
1057 |
+
if quantize_denoised:
|
1058 |
+
x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
|
1059 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
1060 |
+
if return_codebook_ids:
|
1061 |
+
return model_mean, posterior_variance, posterior_log_variance, logits
|
1062 |
+
elif return_x0:
|
1063 |
+
return model_mean, posterior_variance, posterior_log_variance, x_recon
|
1064 |
+
else:
|
1065 |
+
return model_mean, posterior_variance, posterior_log_variance
|
1066 |
+
|
1067 |
+
@torch.no_grad()
|
1068 |
+
def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
|
1069 |
+
return_codebook_ids=False, quantize_denoised=False, return_x0=False,
|
1070 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
|
1071 |
+
b, *_, device = *x.shape, x.device
|
1072 |
+
outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
|
1073 |
+
return_codebook_ids=return_codebook_ids,
|
1074 |
+
quantize_denoised=quantize_denoised,
|
1075 |
+
return_x0=return_x0,
|
1076 |
+
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
|
1077 |
+
if return_codebook_ids:
|
1078 |
+
raise DeprecationWarning("Support dropped.")
|
1079 |
+
model_mean, _, model_log_variance, logits = outputs
|
1080 |
+
elif return_x0:
|
1081 |
+
model_mean, _, model_log_variance, x0 = outputs
|
1082 |
+
else:
|
1083 |
+
model_mean, _, model_log_variance = outputs
|
1084 |
+
|
1085 |
+
noise = noise_like(x.shape, device, repeat_noise) * temperature
|
1086 |
+
if noise_dropout > 0.:
|
1087 |
+
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
1088 |
+
# no noise when t == 0
|
1089 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
1090 |
+
|
1091 |
+
if return_codebook_ids:
|
1092 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
|
1093 |
+
if return_x0:
|
1094 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
|
1095 |
+
else:
|
1096 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
1097 |
+
|
1098 |
+
@torch.no_grad()
|
1099 |
+
def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
|
1100 |
+
img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
|
1101 |
+
score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
|
1102 |
+
log_every_t=None):
|
1103 |
+
if not log_every_t:
|
1104 |
+
log_every_t = self.log_every_t
|
1105 |
+
timesteps = self.num_timesteps
|
1106 |
+
if batch_size is not None:
|
1107 |
+
b = batch_size if batch_size is not None else shape[0]
|
1108 |
+
shape = [batch_size] + list(shape)
|
1109 |
+
else:
|
1110 |
+
b = batch_size = shape[0]
|
1111 |
+
if x_T is None:
|
1112 |
+
img = torch.randn(shape, device=self.device)
|
1113 |
+
else:
|
1114 |
+
img = x_T
|
1115 |
+
intermediates = []
|
1116 |
+
if cond is not None:
|
1117 |
+
if isinstance(cond, dict):
|
1118 |
+
cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
|
1119 |
+
[x[:batch_size] for x in cond[key]] for key in cond}
|
1120 |
+
else:
|
1121 |
+
cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
|
1122 |
+
|
1123 |
+
if start_T is not None:
|
1124 |
+
timesteps = min(timesteps, start_T)
|
1125 |
+
iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
|
1126 |
+
total=timesteps) if verbose else reversed(
|
1127 |
+
range(0, timesteps))
|
1128 |
+
if type(temperature) == float:
|
1129 |
+
temperature = [temperature] * timesteps
|
1130 |
+
|
1131 |
+
for i in iterator:
|
1132 |
+
ts = torch.full((b,), i, device=self.device, dtype=torch.long)
|
1133 |
+
if self.shorten_cond_schedule:
|
1134 |
+
assert self.model.conditioning_key != 'hybrid'
|
1135 |
+
tc = self.cond_ids[ts].to(cond.device)
|
1136 |
+
cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
|
1137 |
+
|
1138 |
+
img, x0_partial = self.p_sample(img, cond, ts,
|
1139 |
+
clip_denoised=self.clip_denoised,
|
1140 |
+
quantize_denoised=quantize_denoised, return_x0=True,
|
1141 |
+
temperature=temperature[i], noise_dropout=noise_dropout,
|
1142 |
+
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
|
1143 |
+
if mask is not None:
|
1144 |
+
assert x0 is not None
|
1145 |
+
img_orig = self.q_sample(x0, ts)
|
1146 |
+
img = img_orig * mask + (1. - mask) * img
|
1147 |
+
|
1148 |
+
if i % log_every_t == 0 or i == timesteps - 1:
|
1149 |
+
intermediates.append(x0_partial)
|
1150 |
+
if callback:
|
1151 |
+
callback(i)
|
1152 |
+
if img_callback:
|
1153 |
+
img_callback(img, i)
|
1154 |
+
return img, intermediates
|
1155 |
+
|
1156 |
+
@torch.no_grad()
|
1157 |
+
def p_sample_loop(self, cond, shape, return_intermediates=False,
|
1158 |
+
x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
|
1159 |
+
mask=None, x0=None, img_callback=None, start_T=None,
|
1160 |
+
log_every_t=None):
|
1161 |
+
|
1162 |
+
if not log_every_t:
|
1163 |
+
log_every_t = self.log_every_t
|
1164 |
+
device = self.betas.device
|
1165 |
+
b = shape[0]
|
1166 |
+
if x_T is None:
|
1167 |
+
img = torch.randn(shape, device=device)
|
1168 |
+
else:
|
1169 |
+
img = x_T
|
1170 |
+
|
1171 |
+
intermediates = [img]
|
1172 |
+
if timesteps is None:
|
1173 |
+
timesteps = self.num_timesteps
|
1174 |
+
|
1175 |
+
if start_T is not None:
|
1176 |
+
timesteps = min(timesteps, start_T)
|
1177 |
+
iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
|
1178 |
+
range(0, timesteps))
|
1179 |
+
|
1180 |
+
if mask is not None:
|
1181 |
+
assert x0 is not None
|
1182 |
+
assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match
|
1183 |
+
|
1184 |
+
for i in iterator:
|
1185 |
+
ts = torch.full((b,), i, device=device, dtype=torch.long)
|
1186 |
+
if self.shorten_cond_schedule:
|
1187 |
+
assert self.model.conditioning_key != 'hybrid'
|
1188 |
+
tc = self.cond_ids[ts].to(cond.device)
|
1189 |
+
cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
|
1190 |
+
|
1191 |
+
img = self.p_sample(img, cond, ts,
|
1192 |
+
clip_denoised=self.clip_denoised,
|
1193 |
+
quantize_denoised=quantize_denoised)
|
1194 |
+
if mask is not None:
|
1195 |
+
img_orig = self.q_sample(x0, ts)
|
1196 |
+
img = img_orig * mask + (1. - mask) * img
|
1197 |
+
|
1198 |
+
if i % log_every_t == 0 or i == timesteps - 1:
|
1199 |
+
intermediates.append(img)
|
1200 |
+
if callback:
|
1201 |
+
callback(i)
|
1202 |
+
if img_callback:
|
1203 |
+
img_callback(img, i)
|
1204 |
+
|
1205 |
+
if return_intermediates:
|
1206 |
+
return img, intermediates
|
1207 |
+
return img
|
1208 |
+
|
1209 |
+
@torch.no_grad()
|
1210 |
+
def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
|
1211 |
+
verbose=True, timesteps=None, quantize_denoised=False,
|
1212 |
+
mask=None, x0=None, shape=None,**kwargs):
|
1213 |
+
if shape is None:
|
1214 |
+
shape = (batch_size, self.channels, self.image_size, self.image_size)
|
1215 |
+
if cond is not None:
|
1216 |
+
if isinstance(cond, dict):
|
1217 |
+
cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
|
1218 |
+
[x[:batch_size] for x in cond[key]] for key in cond}
|
1219 |
+
else:
|
1220 |
+
cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
|
1221 |
+
return self.p_sample_loop(cond,
|
1222 |
+
shape,
|
1223 |
+
return_intermediates=return_intermediates, x_T=x_T,
|
1224 |
+
verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
|
1225 |
+
mask=mask, x0=x0)
|
1226 |
+
|
1227 |
+
@torch.no_grad()
|
1228 |
+
def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
|
1229 |
+
|
1230 |
+
if ddim:
|
1231 |
+
ddim_sampler = DDIMSampler(self)
|
1232 |
+
shape = (self.channels, self.image_size, self.image_size)
|
1233 |
+
samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
|
1234 |
+
shape,cond,verbose=False,**kwargs)
|
1235 |
+
|
1236 |
+
else:
|
1237 |
+
samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
|
1238 |
+
return_intermediates=True,**kwargs)
|
1239 |
+
|
1240 |
+
return samples, intermediates
|
1241 |
+
|
1242 |
+
|
1243 |
+
@torch.no_grad()
|
1244 |
+
def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
|
1245 |
+
quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
|
1246 |
+
plot_diffusion_rows=True, **kwargs):
|
1247 |
+
|
1248 |
+
use_ddim = ddim_steps is not None
|
1249 |
+
|
1250 |
+
log = {}
|
1251 |
+
z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
|
1252 |
+
return_first_stage_outputs=True,
|
1253 |
+
force_c_encode=True,
|
1254 |
+
return_original_cond=True,
|
1255 |
+
bs=N)
|
1256 |
+
N = min(x.shape[0], N)
|
1257 |
+
n_row = min(x.shape[0], n_row)
|
1258 |
+
log["inputs"] = x
|
1259 |
+
log["reconstruction"] = xrec
|
1260 |
+
if self.model.conditioning_key is not None:
|
1261 |
+
if hasattr(self.cond_stage_model, "decode"):
|
1262 |
+
xc = self.cond_stage_model.decode(c)
|
1263 |
+
log["conditioning"] = xc
|
1264 |
+
elif self.cond_stage_key in ["caption"]:
|
1265 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
|
1266 |
+
log["conditioning"] = xc
|
1267 |
+
elif self.cond_stage_key == 'class_label':
|
1268 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
|
1269 |
+
log['conditioning'] = xc
|
1270 |
+
elif isimage(xc):
|
1271 |
+
log["conditioning"] = xc
|
1272 |
+
if ismap(xc):
|
1273 |
+
log["original_conditioning"] = self.to_rgb(xc)
|
1274 |
+
|
1275 |
+
if plot_diffusion_rows:
|
1276 |
+
# get diffusion row
|
1277 |
+
diffusion_row = []
|
1278 |
+
z_start = z[:n_row]
|
1279 |
+
for t in range(self.num_timesteps):
|
1280 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
1281 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
1282 |
+
t = t.to(self.device).long()
|
1283 |
+
noise = torch.randn_like(z_start)
|
1284 |
+
z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
|
1285 |
+
diffusion_row.append(self.decode_first_stage(z_noisy))
|
1286 |
+
|
1287 |
+
diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
|
1288 |
+
diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
|
1289 |
+
diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
|
1290 |
+
diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
|
1291 |
+
log["diffusion_row"] = diffusion_grid
|
1292 |
+
|
1293 |
+
if sample:
|
1294 |
+
# get denoise row
|
1295 |
+
with self.ema_scope("Plotting"):
|
1296 |
+
samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
|
1297 |
+
ddim_steps=ddim_steps,eta=ddim_eta)
|
1298 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
|
1299 |
+
x_samples = self.decode_first_stage(samples)
|
1300 |
+
log["samples"] = x_samples
|
1301 |
+
if plot_denoise_rows:
|
1302 |
+
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
|
1303 |
+
log["denoise_row"] = denoise_grid
|
1304 |
+
|
1305 |
+
if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
|
1306 |
+
self.first_stage_model, IdentityFirstStage):
|
1307 |
+
# also display when quantizing x0 while sampling
|
1308 |
+
with self.ema_scope("Plotting Quantized Denoised"):
|
1309 |
+
samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
|
1310 |
+
ddim_steps=ddim_steps,eta=ddim_eta,
|
1311 |
+
quantize_denoised=True)
|
1312 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
|
1313 |
+
# quantize_denoised=True)
|
1314 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1315 |
+
log["samples_x0_quantized"] = x_samples
|
1316 |
+
|
1317 |
+
if inpaint:
|
1318 |
+
# make a simple center square
|
1319 |
+
h, w = z.shape[2], z.shape[3]
|
1320 |
+
mask = torch.ones(N, h, w).to(self.device)
|
1321 |
+
# zeros will be filled in
|
1322 |
+
mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
|
1323 |
+
mask = mask[:, None, ...]
|
1324 |
+
with self.ema_scope("Plotting Inpaint"):
|
1325 |
+
|
1326 |
+
samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
|
1327 |
+
ddim_steps=ddim_steps, x0=z[:N], mask=mask)
|
1328 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1329 |
+
log["samples_inpainting"] = x_samples
|
1330 |
+
log["mask"] = mask
|
1331 |
+
|
1332 |
+
# outpaint
|
1333 |
+
with self.ema_scope("Plotting Outpaint"):
|
1334 |
+
samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
|
1335 |
+
ddim_steps=ddim_steps, x0=z[:N], mask=mask)
|
1336 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1337 |
+
log["samples_outpainting"] = x_samples
|
1338 |
+
|
1339 |
+
if plot_progressive_rows:
|
1340 |
+
with self.ema_scope("Plotting Progressives"):
|
1341 |
+
img, progressives = self.progressive_denoising(c,
|
1342 |
+
shape=(self.channels, self.image_size, self.image_size),
|
1343 |
+
batch_size=N)
|
1344 |
+
prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
|
1345 |
+
log["progressive_row"] = prog_row
|
1346 |
+
|
1347 |
+
if return_keys:
|
1348 |
+
if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
|
1349 |
+
return log
|
1350 |
+
else:
|
1351 |
+
return {key: log[key] for key in return_keys}
|
1352 |
+
return log
|
1353 |
+
|
1354 |
+
def configure_optimizers(self):
|
1355 |
+
lr = self.learning_rate
|
1356 |
+
params = list(self.model.parameters())
|
1357 |
+
if self.cond_stage_trainable:
|
1358 |
+
print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
|
1359 |
+
params = params + list(self.cond_stage_model.parameters())
|
1360 |
+
if self.learn_logvar:
|
1361 |
+
print('Diffusion model optimizing logvar')
|
1362 |
+
params.append(self.logvar)
|
1363 |
+
opt = torch.optim.AdamW(params, lr=lr)
|
1364 |
+
if self.use_scheduler:
|
1365 |
+
assert 'target' in self.scheduler_config
|
1366 |
+
scheduler = instantiate_from_config(self.scheduler_config)
|
1367 |
+
|
1368 |
+
print("Setting up LambdaLR scheduler...")
|
1369 |
+
scheduler = [
|
1370 |
+
{
|
1371 |
+
'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
|
1372 |
+
'interval': 'step',
|
1373 |
+
'frequency': 1
|
1374 |
+
}]
|
1375 |
+
return [opt], scheduler
|
1376 |
+
return opt
|
1377 |
+
|
1378 |
+
@torch.no_grad()
|
1379 |
+
def to_rgb(self, x):
|
1380 |
+
x = x.float()
|
1381 |
+
if not hasattr(self, "colorize"):
|
1382 |
+
self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
|
1383 |
+
x = nn.functional.conv2d(x, weight=self.colorize)
|
1384 |
+
x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
|
1385 |
+
return x
|
1386 |
+
|
1387 |
+
|
1388 |
+
class DiffusionWrapperV1(pl.LightningModule):
|
1389 |
+
def __init__(self, diff_model_config, conditioning_key):
|
1390 |
+
super().__init__()
|
1391 |
+
self.diffusion_model = instantiate_from_config(diff_model_config)
|
1392 |
+
self.conditioning_key = conditioning_key
|
1393 |
+
assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm']
|
1394 |
+
|
1395 |
+
def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
|
1396 |
+
if self.conditioning_key is None:
|
1397 |
+
out = self.diffusion_model(x, t)
|
1398 |
+
elif self.conditioning_key == 'concat':
|
1399 |
+
xc = torch.cat([x] + c_concat, dim=1)
|
1400 |
+
out = self.diffusion_model(xc, t)
|
1401 |
+
elif self.conditioning_key == 'crossattn':
|
1402 |
+
cc = torch.cat(c_crossattn, 1)
|
1403 |
+
out = self.diffusion_model(x, t, context=cc)
|
1404 |
+
elif self.conditioning_key == 'hybrid':
|
1405 |
+
xc = torch.cat([x] + c_concat, dim=1)
|
1406 |
+
cc = torch.cat(c_crossattn, 1)
|
1407 |
+
out = self.diffusion_model(xc, t, context=cc)
|
1408 |
+
elif self.conditioning_key == 'adm':
|
1409 |
+
cc = c_crossattn[0]
|
1410 |
+
out = self.diffusion_model(x, t, y=cc)
|
1411 |
+
else:
|
1412 |
+
raise NotImplementedError()
|
1413 |
+
|
1414 |
+
return out
|
1415 |
+
|
1416 |
+
|
1417 |
+
class Layout2ImgDiffusionV1(LatentDiffusionV1):
|
1418 |
+
# TODO: move all layout-specific hacks to this class
|
1419 |
+
def __init__(self, cond_stage_key, *args, **kwargs):
|
1420 |
+
assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
|
1421 |
+
super().__init__(*args, cond_stage_key=cond_stage_key, **kwargs)
|
1422 |
+
|
1423 |
+
def log_images(self, batch, N=8, *args, **kwargs):
|
1424 |
+
logs = super().log_images(*args, batch=batch, N=N, **kwargs)
|
1425 |
+
|
1426 |
+
key = 'train' if self.training else 'validation'
|
1427 |
+
dset = self.trainer.datamodule.datasets[key]
|
1428 |
+
mapper = dset.conditional_builders[self.cond_stage_key]
|
1429 |
+
|
1430 |
+
bbox_imgs = []
|
1431 |
+
map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
|
1432 |
+
for tknzd_bbox in batch[self.cond_stage_key][:N]:
|
1433 |
+
bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
|
1434 |
+
bbox_imgs.append(bboximg)
|
1435 |
+
|
1436 |
+
cond_img = torch.stack(bbox_imgs, dim=0)
|
1437 |
+
logs['bbox_image'] = cond_img
|
1438 |
+
return logs
|
1439 |
+
|
1440 |
+
ldm.models.diffusion.ddpm.DDPMV1 = DDPMV1
|
1441 |
+
ldm.models.diffusion.ddpm.LatentDiffusionV1 = LatentDiffusionV1
|
1442 |
+
ldm.models.diffusion.ddpm.DiffusionWrapperV1 = DiffusionWrapperV1
|
1443 |
+
ldm.models.diffusion.ddpm.Layout2ImgDiffusionV1 = Layout2ImgDiffusionV1
|
extensions-builtin/LDSR/vqvae_quantize.py
ADDED
@@ -0,0 +1,147 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Vendored from https://raw.githubusercontent.com/CompVis/taming-transformers/24268930bf1dce879235a7fddd0b2355b84d7ea6/taming/modules/vqvae/quantize.py,
|
2 |
+
# where the license is as follows:
|
3 |
+
#
|
4 |
+
# Copyright (c) 2020 Patrick Esser and Robin Rombach and Björn Ommer
|
5 |
+
#
|
6 |
+
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
7 |
+
# of this software and associated documentation files (the "Software"), to deal
|
8 |
+
# in the Software without restriction, including without limitation the rights
|
9 |
+
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
10 |
+
# copies of the Software, and to permit persons to whom the Software is
|
11 |
+
# furnished to do so, subject to the following conditions:
|
12 |
+
#
|
13 |
+
# The above copyright notice and this permission notice shall be included in all
|
14 |
+
# copies or substantial portions of the Software.
|
15 |
+
#
|
16 |
+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
|
17 |
+
# EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
|
18 |
+
# MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
|
19 |
+
# IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
|
20 |
+
# DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
|
21 |
+
# OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE
|
22 |
+
# OR OTHER DEALINGS IN THE SOFTWARE./
|
23 |
+
|
24 |
+
import torch
|
25 |
+
import torch.nn as nn
|
26 |
+
import numpy as np
|
27 |
+
from einops import rearrange
|
28 |
+
|
29 |
+
|
30 |
+
class VectorQuantizer2(nn.Module):
|
31 |
+
"""
|
32 |
+
Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly
|
33 |
+
avoids costly matrix multiplications and allows for post-hoc remapping of indices.
|
34 |
+
"""
|
35 |
+
|
36 |
+
# NOTE: due to a bug the beta term was applied to the wrong term. for
|
37 |
+
# backwards compatibility we use the buggy version by default, but you can
|
38 |
+
# specify legacy=False to fix it.
|
39 |
+
def __init__(self, n_e, e_dim, beta, remap=None, unknown_index="random",
|
40 |
+
sane_index_shape=False, legacy=True):
|
41 |
+
super().__init__()
|
42 |
+
self.n_e = n_e
|
43 |
+
self.e_dim = e_dim
|
44 |
+
self.beta = beta
|
45 |
+
self.legacy = legacy
|
46 |
+
|
47 |
+
self.embedding = nn.Embedding(self.n_e, self.e_dim)
|
48 |
+
self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
|
49 |
+
|
50 |
+
self.remap = remap
|
51 |
+
if self.remap is not None:
|
52 |
+
self.register_buffer("used", torch.tensor(np.load(self.remap)))
|
53 |
+
self.re_embed = self.used.shape[0]
|
54 |
+
self.unknown_index = unknown_index # "random" or "extra" or integer
|
55 |
+
if self.unknown_index == "extra":
|
56 |
+
self.unknown_index = self.re_embed
|
57 |
+
self.re_embed = self.re_embed + 1
|
58 |
+
print(f"Remapping {self.n_e} indices to {self.re_embed} indices. "
|
59 |
+
f"Using {self.unknown_index} for unknown indices.")
|
60 |
+
else:
|
61 |
+
self.re_embed = n_e
|
62 |
+
|
63 |
+
self.sane_index_shape = sane_index_shape
|
64 |
+
|
65 |
+
def remap_to_used(self, inds):
|
66 |
+
ishape = inds.shape
|
67 |
+
assert len(ishape) > 1
|
68 |
+
inds = inds.reshape(ishape[0], -1)
|
69 |
+
used = self.used.to(inds)
|
70 |
+
match = (inds[:, :, None] == used[None, None, ...]).long()
|
71 |
+
new = match.argmax(-1)
|
72 |
+
unknown = match.sum(2) < 1
|
73 |
+
if self.unknown_index == "random":
|
74 |
+
new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
|
75 |
+
else:
|
76 |
+
new[unknown] = self.unknown_index
|
77 |
+
return new.reshape(ishape)
|
78 |
+
|
79 |
+
def unmap_to_all(self, inds):
|
80 |
+
ishape = inds.shape
|
81 |
+
assert len(ishape) > 1
|
82 |
+
inds = inds.reshape(ishape[0], -1)
|
83 |
+
used = self.used.to(inds)
|
84 |
+
if self.re_embed > self.used.shape[0]: # extra token
|
85 |
+
inds[inds >= self.used.shape[0]] = 0 # simply set to zero
|
86 |
+
back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
|
87 |
+
return back.reshape(ishape)
|
88 |
+
|
89 |
+
def forward(self, z, temp=None, rescale_logits=False, return_logits=False):
|
90 |
+
assert temp is None or temp == 1.0, "Only for interface compatible with Gumbel"
|
91 |
+
assert rescale_logits is False, "Only for interface compatible with Gumbel"
|
92 |
+
assert return_logits is False, "Only for interface compatible with Gumbel"
|
93 |
+
# reshape z -> (batch, height, width, channel) and flatten
|
94 |
+
z = rearrange(z, 'b c h w -> b h w c').contiguous()
|
95 |
+
z_flattened = z.view(-1, self.e_dim)
|
96 |
+
# distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
|
97 |
+
|
98 |
+
d = torch.sum(z_flattened ** 2, dim=1, keepdim=True) + \
|
99 |
+
torch.sum(self.embedding.weight ** 2, dim=1) - 2 * \
|
100 |
+
torch.einsum('bd,dn->bn', z_flattened, rearrange(self.embedding.weight, 'n d -> d n'))
|
101 |
+
|
102 |
+
min_encoding_indices = torch.argmin(d, dim=1)
|
103 |
+
z_q = self.embedding(min_encoding_indices).view(z.shape)
|
104 |
+
perplexity = None
|
105 |
+
min_encodings = None
|
106 |
+
|
107 |
+
# compute loss for embedding
|
108 |
+
if not self.legacy:
|
109 |
+
loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + \
|
110 |
+
torch.mean((z_q - z.detach()) ** 2)
|
111 |
+
else:
|
112 |
+
loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * \
|
113 |
+
torch.mean((z_q - z.detach()) ** 2)
|
114 |
+
|
115 |
+
# preserve gradients
|
116 |
+
z_q = z + (z_q - z).detach()
|
117 |
+
|
118 |
+
# reshape back to match original input shape
|
119 |
+
z_q = rearrange(z_q, 'b h w c -> b c h w').contiguous()
|
120 |
+
|
121 |
+
if self.remap is not None:
|
122 |
+
min_encoding_indices = min_encoding_indices.reshape(z.shape[0], -1) # add batch axis
|
123 |
+
min_encoding_indices = self.remap_to_used(min_encoding_indices)
|
124 |
+
min_encoding_indices = min_encoding_indices.reshape(-1, 1) # flatten
|
125 |
+
|
126 |
+
if self.sane_index_shape:
|
127 |
+
min_encoding_indices = min_encoding_indices.reshape(
|
128 |
+
z_q.shape[0], z_q.shape[2], z_q.shape[3])
|
129 |
+
|
130 |
+
return z_q, loss, (perplexity, min_encodings, min_encoding_indices)
|
131 |
+
|
132 |
+
def get_codebook_entry(self, indices, shape):
|
133 |
+
# shape specifying (batch, height, width, channel)
|
134 |
+
if self.remap is not None:
|
135 |
+
indices = indices.reshape(shape[0], -1) # add batch axis
|
136 |
+
indices = self.unmap_to_all(indices)
|
137 |
+
indices = indices.reshape(-1) # flatten again
|
138 |
+
|
139 |
+
# get quantized latent vectors
|
140 |
+
z_q = self.embedding(indices)
|
141 |
+
|
142 |
+
if shape is not None:
|
143 |
+
z_q = z_q.view(shape)
|
144 |
+
# reshape back to match original input shape
|
145 |
+
z_q = z_q.permute(0, 3, 1, 2).contiguous()
|
146 |
+
|
147 |
+
return z_q
|
extensions-builtin/Lora/__pycache__/extra_networks_lora.cpython-310.pyc
ADDED
Binary file (2.38 kB). View file
|
|
extensions-builtin/Lora/__pycache__/lora.cpython-310.pyc
ADDED
Binary file (524 Bytes). View file
|
|
extensions-builtin/Lora/__pycache__/lyco_helpers.cpython-310.pyc
ADDED
Binary file (923 Bytes). View file
|
|
extensions-builtin/Lora/__pycache__/network.cpython-310.pyc
ADDED
Binary file (5.63 kB). View file
|
|
extensions-builtin/Lora/__pycache__/network_full.cpython-310.pyc
ADDED
Binary file (1.48 kB). View file
|
|
extensions-builtin/Lora/__pycache__/network_hada.cpython-310.pyc
ADDED
Binary file (2.21 kB). View file
|
|
extensions-builtin/Lora/__pycache__/network_ia3.cpython-310.pyc
ADDED
Binary file (1.6 kB). View file
|
|
extensions-builtin/Lora/__pycache__/network_lokr.cpython-310.pyc
ADDED
Binary file (2.41 kB). View file
|
|
extensions-builtin/Lora/__pycache__/network_lora.cpython-310.pyc
ADDED
Binary file (3.48 kB). View file
|
|
extensions-builtin/Lora/__pycache__/networks.cpython-310.pyc
ADDED
Binary file (12.7 kB). View file
|
|
extensions-builtin/Lora/__pycache__/preload.cpython-310.pyc
ADDED
Binary file (664 Bytes). View file
|
|
extensions-builtin/Lora/__pycache__/ui_edit_user_metadata.cpython-310.pyc
ADDED
Binary file (7.51 kB). View file
|
|
extensions-builtin/Lora/__pycache__/ui_extra_networks_lora.cpython-310.pyc
ADDED
Binary file (3.05 kB). View file
|
|
extensions-builtin/Lora/extra_networks_lora.py
ADDED
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from modules import extra_networks, shared
|
2 |
+
import networks
|
3 |
+
|
4 |
+
|
5 |
+
class ExtraNetworkLora(extra_networks.ExtraNetwork):
|
6 |
+
def __init__(self):
|
7 |
+
super().__init__('lora')
|
8 |
+
|
9 |
+
self.errors = {}
|
10 |
+
"""mapping of network names to the number of errors the network had during operation"""
|
11 |
+
|
12 |
+
def activate(self, p, params_list):
|
13 |
+
additional = shared.opts.sd_lora
|
14 |
+
|
15 |
+
self.errors.clear()
|
16 |
+
|
17 |
+
if additional != "None" and additional in networks.available_networks and not any(x for x in params_list if x.items[0] == additional):
|
18 |
+
p.all_prompts = [x + f"<lora:{additional}:{shared.opts.extra_networks_default_multiplier}>" for x in p.all_prompts]
|
19 |
+
params_list.append(extra_networks.ExtraNetworkParams(items=[additional, shared.opts.extra_networks_default_multiplier]))
|
20 |
+
|
21 |
+
names = []
|
22 |
+
te_multipliers = []
|
23 |
+
unet_multipliers = []
|
24 |
+
dyn_dims = []
|
25 |
+
for params in params_list:
|
26 |
+
assert params.items
|
27 |
+
|
28 |
+
names.append(params.positional[0])
|
29 |
+
|
30 |
+
te_multiplier = float(params.positional[1]) if len(params.positional) > 1 else 1.0
|
31 |
+
te_multiplier = float(params.named.get("te", te_multiplier))
|
32 |
+
|
33 |
+
unet_multiplier = float(params.positional[2]) if len(params.positional) > 2 else te_multiplier
|
34 |
+
unet_multiplier = float(params.named.get("unet", unet_multiplier))
|
35 |
+
|
36 |
+
dyn_dim = int(params.positional[3]) if len(params.positional) > 3 else None
|
37 |
+
dyn_dim = int(params.named["dyn"]) if "dyn" in params.named else dyn_dim
|
38 |
+
|
39 |
+
te_multipliers.append(te_multiplier)
|
40 |
+
unet_multipliers.append(unet_multiplier)
|
41 |
+
dyn_dims.append(dyn_dim)
|
42 |
+
|
43 |
+
networks.load_networks(names, te_multipliers, unet_multipliers, dyn_dims)
|
44 |
+
|
45 |
+
if shared.opts.lora_add_hashes_to_infotext:
|
46 |
+
network_hashes = []
|
47 |
+
for item in networks.loaded_networks:
|
48 |
+
shorthash = item.network_on_disk.shorthash
|
49 |
+
if not shorthash:
|
50 |
+
continue
|
51 |
+
|
52 |
+
alias = item.mentioned_name
|
53 |
+
if not alias:
|
54 |
+
continue
|
55 |
+
|
56 |
+
alias = alias.replace(":", "").replace(",", "")
|
57 |
+
|
58 |
+
network_hashes.append(f"{alias}: {shorthash}")
|
59 |
+
|
60 |
+
if network_hashes:
|
61 |
+
p.extra_generation_params["Lora hashes"] = ", ".join(network_hashes)
|
62 |
+
|
63 |
+
def deactivate(self, p):
|
64 |
+
if self.errors:
|
65 |
+
p.comment("Networks with errors: " + ", ".join(f"{k} ({v})" for k, v in self.errors.items()))
|
66 |
+
|
67 |
+
self.errors.clear()
|
extensions-builtin/Lora/lora.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import networks
|
2 |
+
|
3 |
+
list_available_loras = networks.list_available_networks
|
4 |
+
|
5 |
+
available_loras = networks.available_networks
|
6 |
+
available_lora_aliases = networks.available_network_aliases
|
7 |
+
available_lora_hash_lookup = networks.available_network_hash_lookup
|
8 |
+
forbidden_lora_aliases = networks.forbidden_network_aliases
|
9 |
+
loaded_loras = networks.loaded_networks
|
extensions-builtin/Lora/lora_patches.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
import networks
|
4 |
+
from modules import patches
|
5 |
+
|
6 |
+
|
7 |
+
class LoraPatches:
|
8 |
+
def __init__(self):
|
9 |
+
self.Linear_forward = patches.patch(__name__, torch.nn.Linear, 'forward', networks.network_Linear_forward)
|
10 |
+
self.Linear_load_state_dict = patches.patch(__name__, torch.nn.Linear, '_load_from_state_dict', networks.network_Linear_load_state_dict)
|
11 |
+
self.Conv2d_forward = patches.patch(__name__, torch.nn.Conv2d, 'forward', networks.network_Conv2d_forward)
|
12 |
+
self.Conv2d_load_state_dict = patches.patch(__name__, torch.nn.Conv2d, '_load_from_state_dict', networks.network_Conv2d_load_state_dict)
|
13 |
+
self.GroupNorm_forward = patches.patch(__name__, torch.nn.GroupNorm, 'forward', networks.network_GroupNorm_forward)
|
14 |
+
self.GroupNorm_load_state_dict = patches.patch(__name__, torch.nn.GroupNorm, '_load_from_state_dict', networks.network_GroupNorm_load_state_dict)
|
15 |
+
self.LayerNorm_forward = patches.patch(__name__, torch.nn.LayerNorm, 'forward', networks.network_LayerNorm_forward)
|
16 |
+
self.LayerNorm_load_state_dict = patches.patch(__name__, torch.nn.LayerNorm, '_load_from_state_dict', networks.network_LayerNorm_load_state_dict)
|
17 |
+
self.MultiheadAttention_forward = patches.patch(__name__, torch.nn.MultiheadAttention, 'forward', networks.network_MultiheadAttention_forward)
|
18 |
+
self.MultiheadAttention_load_state_dict = patches.patch(__name__, torch.nn.MultiheadAttention, '_load_from_state_dict', networks.network_MultiheadAttention_load_state_dict)
|
19 |
+
|
20 |
+
def undo(self):
|
21 |
+
self.Linear_forward = patches.undo(__name__, torch.nn.Linear, 'forward')
|
22 |
+
self.Linear_load_state_dict = patches.undo(__name__, torch.nn.Linear, '_load_from_state_dict')
|
23 |
+
self.Conv2d_forward = patches.undo(__name__, torch.nn.Conv2d, 'forward')
|
24 |
+
self.Conv2d_load_state_dict = patches.undo(__name__, torch.nn.Conv2d, '_load_from_state_dict')
|
25 |
+
self.GroupNorm_forward = patches.undo(__name__, torch.nn.GroupNorm, 'forward')
|
26 |
+
self.GroupNorm_load_state_dict = patches.undo(__name__, torch.nn.GroupNorm, '_load_from_state_dict')
|
27 |
+
self.LayerNorm_forward = patches.undo(__name__, torch.nn.LayerNorm, 'forward')
|
28 |
+
self.LayerNorm_load_state_dict = patches.undo(__name__, torch.nn.LayerNorm, '_load_from_state_dict')
|
29 |
+
self.MultiheadAttention_forward = patches.undo(__name__, torch.nn.MultiheadAttention, 'forward')
|
30 |
+
self.MultiheadAttention_load_state_dict = patches.undo(__name__, torch.nn.MultiheadAttention, '_load_from_state_dict')
|
31 |
+
|
extensions-builtin/Lora/lyco_helpers.py
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
|
4 |
+
def make_weight_cp(t, wa, wb):
|
5 |
+
temp = torch.einsum('i j k l, j r -> i r k l', t, wb)
|
6 |
+
return torch.einsum('i j k l, i r -> r j k l', temp, wa)
|
7 |
+
|
8 |
+
|
9 |
+
def rebuild_conventional(up, down, shape, dyn_dim=None):
|
10 |
+
up = up.reshape(up.size(0), -1)
|
11 |
+
down = down.reshape(down.size(0), -1)
|
12 |
+
if dyn_dim is not None:
|
13 |
+
up = up[:, :dyn_dim]
|
14 |
+
down = down[:dyn_dim, :]
|
15 |
+
return (up @ down).reshape(shape)
|
16 |
+
|
17 |
+
|
18 |
+
def rebuild_cp_decomposition(up, down, mid):
|
19 |
+
up = up.reshape(up.size(0), -1)
|
20 |
+
down = down.reshape(down.size(0), -1)
|
21 |
+
return torch.einsum('n m k l, i n, m j -> i j k l', mid, up, down)
|
extensions-builtin/Lora/network.py
ADDED
@@ -0,0 +1,158 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
import os
|
3 |
+
from collections import namedtuple
|
4 |
+
import enum
|
5 |
+
|
6 |
+
from modules import sd_models, cache, errors, hashes, shared
|
7 |
+
|
8 |
+
NetworkWeights = namedtuple('NetworkWeights', ['network_key', 'sd_key', 'w', 'sd_module'])
|
9 |
+
|
10 |
+
metadata_tags_order = {"ss_sd_model_name": 1, "ss_resolution": 2, "ss_clip_skip": 3, "ss_num_train_images": 10, "ss_tag_frequency": 20}
|
11 |
+
|
12 |
+
|
13 |
+
class SdVersion(enum.Enum):
|
14 |
+
Unknown = 1
|
15 |
+
SD1 = 2
|
16 |
+
SD2 = 3
|
17 |
+
SDXL = 4
|
18 |
+
|
19 |
+
|
20 |
+
class NetworkOnDisk:
|
21 |
+
def __init__(self, name, filename):
|
22 |
+
self.name = name
|
23 |
+
self.filename = filename
|
24 |
+
self.metadata = {}
|
25 |
+
self.is_safetensors = os.path.splitext(filename)[1].lower() == ".safetensors"
|
26 |
+
|
27 |
+
def read_metadata():
|
28 |
+
metadata = sd_models.read_metadata_from_safetensors(filename)
|
29 |
+
metadata.pop('ssmd_cover_images', None) # those are cover images, and they are too big to display in UI as text
|
30 |
+
|
31 |
+
return metadata
|
32 |
+
|
33 |
+
if self.is_safetensors:
|
34 |
+
try:
|
35 |
+
self.metadata = cache.cached_data_for_file('safetensors-metadata', "lora/" + self.name, filename, read_metadata)
|
36 |
+
except Exception as e:
|
37 |
+
errors.display(e, f"reading lora {filename}")
|
38 |
+
|
39 |
+
if self.metadata:
|
40 |
+
m = {}
|
41 |
+
for k, v in sorted(self.metadata.items(), key=lambda x: metadata_tags_order.get(x[0], 999)):
|
42 |
+
m[k] = v
|
43 |
+
|
44 |
+
self.metadata = m
|
45 |
+
|
46 |
+
self.alias = self.metadata.get('ss_output_name', self.name)
|
47 |
+
|
48 |
+
self.hash = None
|
49 |
+
self.shorthash = None
|
50 |
+
self.set_hash(
|
51 |
+
self.metadata.get('sshs_model_hash') or
|
52 |
+
hashes.sha256_from_cache(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or
|
53 |
+
''
|
54 |
+
)
|
55 |
+
|
56 |
+
self.sd_version = self.detect_version()
|
57 |
+
|
58 |
+
def detect_version(self):
|
59 |
+
if str(self.metadata.get('ss_base_model_version', "")).startswith("sdxl_"):
|
60 |
+
return SdVersion.SDXL
|
61 |
+
elif str(self.metadata.get('ss_v2', "")) == "True":
|
62 |
+
return SdVersion.SD2
|
63 |
+
elif len(self.metadata):
|
64 |
+
return SdVersion.SD1
|
65 |
+
|
66 |
+
return SdVersion.Unknown
|
67 |
+
|
68 |
+
def set_hash(self, v):
|
69 |
+
self.hash = v
|
70 |
+
self.shorthash = self.hash[0:12]
|
71 |
+
|
72 |
+
if self.shorthash:
|
73 |
+
import networks
|
74 |
+
networks.available_network_hash_lookup[self.shorthash] = self
|
75 |
+
|
76 |
+
def read_hash(self):
|
77 |
+
if not self.hash:
|
78 |
+
self.set_hash(hashes.sha256(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or '')
|
79 |
+
|
80 |
+
def get_alias(self):
|
81 |
+
import networks
|
82 |
+
if shared.opts.lora_preferred_name == "Filename" or self.alias.lower() in networks.forbidden_network_aliases:
|
83 |
+
return self.name
|
84 |
+
else:
|
85 |
+
return self.alias
|
86 |
+
|
87 |
+
|
88 |
+
class Network: # LoraModule
|
89 |
+
def __init__(self, name, network_on_disk: NetworkOnDisk):
|
90 |
+
self.name = name
|
91 |
+
self.network_on_disk = network_on_disk
|
92 |
+
self.te_multiplier = 1.0
|
93 |
+
self.unet_multiplier = 1.0
|
94 |
+
self.dyn_dim = None
|
95 |
+
self.modules = {}
|
96 |
+
self.mtime = None
|
97 |
+
|
98 |
+
self.mentioned_name = None
|
99 |
+
"""the text that was used to add the network to prompt - can be either name or an alias"""
|
100 |
+
|
101 |
+
|
102 |
+
class ModuleType:
|
103 |
+
def create_module(self, net: Network, weights: NetworkWeights) -> Network | None:
|
104 |
+
return None
|
105 |
+
|
106 |
+
|
107 |
+
class NetworkModule:
|
108 |
+
def __init__(self, net: Network, weights: NetworkWeights):
|
109 |
+
self.network = net
|
110 |
+
self.network_key = weights.network_key
|
111 |
+
self.sd_key = weights.sd_key
|
112 |
+
self.sd_module = weights.sd_module
|
113 |
+
|
114 |
+
if hasattr(self.sd_module, 'weight'):
|
115 |
+
self.shape = self.sd_module.weight.shape
|
116 |
+
|
117 |
+
self.dim = None
|
118 |
+
self.bias = weights.w.get("bias")
|
119 |
+
self.alpha = weights.w["alpha"].item() if "alpha" in weights.w else None
|
120 |
+
self.scale = weights.w["scale"].item() if "scale" in weights.w else None
|
121 |
+
|
122 |
+
def multiplier(self):
|
123 |
+
if 'transformer' in self.sd_key[:20]:
|
124 |
+
return self.network.te_multiplier
|
125 |
+
else:
|
126 |
+
return self.network.unet_multiplier
|
127 |
+
|
128 |
+
def calc_scale(self):
|
129 |
+
if self.scale is not None:
|
130 |
+
return self.scale
|
131 |
+
if self.dim is not None and self.alpha is not None:
|
132 |
+
return self.alpha / self.dim
|
133 |
+
|
134 |
+
return 1.0
|
135 |
+
|
136 |
+
def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None):
|
137 |
+
if self.bias is not None:
|
138 |
+
updown = updown.reshape(self.bias.shape)
|
139 |
+
updown += self.bias.to(orig_weight.device, dtype=orig_weight.dtype)
|
140 |
+
updown = updown.reshape(output_shape)
|
141 |
+
|
142 |
+
if len(output_shape) == 4:
|
143 |
+
updown = updown.reshape(output_shape)
|
144 |
+
|
145 |
+
if orig_weight.size().numel() == updown.size().numel():
|
146 |
+
updown = updown.reshape(orig_weight.shape)
|
147 |
+
|
148 |
+
if ex_bias is not None:
|
149 |
+
ex_bias = ex_bias * self.multiplier()
|
150 |
+
|
151 |
+
return updown * self.calc_scale() * self.multiplier(), ex_bias
|
152 |
+
|
153 |
+
def calc_updown(self, target):
|
154 |
+
raise NotImplementedError()
|
155 |
+
|
156 |
+
def forward(self, x, y):
|
157 |
+
raise NotImplementedError()
|
158 |
+
|
extensions-builtin/Lora/network_full.py
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import network
|
2 |
+
|
3 |
+
|
4 |
+
class ModuleTypeFull(network.ModuleType):
|
5 |
+
def create_module(self, net: network.Network, weights: network.NetworkWeights):
|
6 |
+
if all(x in weights.w for x in ["diff"]):
|
7 |
+
return NetworkModuleFull(net, weights)
|
8 |
+
|
9 |
+
return None
|
10 |
+
|
11 |
+
|
12 |
+
class NetworkModuleFull(network.NetworkModule):
|
13 |
+
def __init__(self, net: network.Network, weights: network.NetworkWeights):
|
14 |
+
super().__init__(net, weights)
|
15 |
+
|
16 |
+
self.weight = weights.w.get("diff")
|
17 |
+
self.ex_bias = weights.w.get("diff_b")
|
18 |
+
|
19 |
+
def calc_updown(self, orig_weight):
|
20 |
+
output_shape = self.weight.shape
|
21 |
+
updown = self.weight.to(orig_weight.device, dtype=orig_weight.dtype)
|
22 |
+
if self.ex_bias is not None:
|
23 |
+
ex_bias = self.ex_bias.to(orig_weight.device, dtype=orig_weight.dtype)
|
24 |
+
else:
|
25 |
+
ex_bias = None
|
26 |
+
|
27 |
+
return self.finalize_updown(updown, orig_weight, output_shape, ex_bias)
|
extensions-builtin/Lora/network_hada.py
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import lyco_helpers
|
2 |
+
import network
|
3 |
+
|
4 |
+
|
5 |
+
class ModuleTypeHada(network.ModuleType):
|
6 |
+
def create_module(self, net: network.Network, weights: network.NetworkWeights):
|
7 |
+
if all(x in weights.w for x in ["hada_w1_a", "hada_w1_b", "hada_w2_a", "hada_w2_b"]):
|
8 |
+
return NetworkModuleHada(net, weights)
|
9 |
+
|
10 |
+
return None
|
11 |
+
|
12 |
+
|
13 |
+
class NetworkModuleHada(network.NetworkModule):
|
14 |
+
def __init__(self, net: network.Network, weights: network.NetworkWeights):
|
15 |
+
super().__init__(net, weights)
|
16 |
+
|
17 |
+
if hasattr(self.sd_module, 'weight'):
|
18 |
+
self.shape = self.sd_module.weight.shape
|
19 |
+
|
20 |
+
self.w1a = weights.w["hada_w1_a"]
|
21 |
+
self.w1b = weights.w["hada_w1_b"]
|
22 |
+
self.dim = self.w1b.shape[0]
|
23 |
+
self.w2a = weights.w["hada_w2_a"]
|
24 |
+
self.w2b = weights.w["hada_w2_b"]
|
25 |
+
|
26 |
+
self.t1 = weights.w.get("hada_t1")
|
27 |
+
self.t2 = weights.w.get("hada_t2")
|
28 |
+
|
29 |
+
def calc_updown(self, orig_weight):
|
30 |
+
w1a = self.w1a.to(orig_weight.device, dtype=orig_weight.dtype)
|
31 |
+
w1b = self.w1b.to(orig_weight.device, dtype=orig_weight.dtype)
|
32 |
+
w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype)
|
33 |
+
w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype)
|
34 |
+
|
35 |
+
output_shape = [w1a.size(0), w1b.size(1)]
|
36 |
+
|
37 |
+
if self.t1 is not None:
|
38 |
+
output_shape = [w1a.size(1), w1b.size(1)]
|
39 |
+
t1 = self.t1.to(orig_weight.device, dtype=orig_weight.dtype)
|
40 |
+
updown1 = lyco_helpers.make_weight_cp(t1, w1a, w1b)
|
41 |
+
output_shape += t1.shape[2:]
|
42 |
+
else:
|
43 |
+
if len(w1b.shape) == 4:
|
44 |
+
output_shape += w1b.shape[2:]
|
45 |
+
updown1 = lyco_helpers.rebuild_conventional(w1a, w1b, output_shape)
|
46 |
+
|
47 |
+
if self.t2 is not None:
|
48 |
+
t2 = self.t2.to(orig_weight.device, dtype=orig_weight.dtype)
|
49 |
+
updown2 = lyco_helpers.make_weight_cp(t2, w2a, w2b)
|
50 |
+
else:
|
51 |
+
updown2 = lyco_helpers.rebuild_conventional(w2a, w2b, output_shape)
|
52 |
+
|
53 |
+
updown = updown1 * updown2
|
54 |
+
|
55 |
+
return self.finalize_updown(updown, orig_weight, output_shape)
|
extensions-builtin/Lora/network_ia3.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import network
|
2 |
+
|
3 |
+
|
4 |
+
class ModuleTypeIa3(network.ModuleType):
|
5 |
+
def create_module(self, net: network.Network, weights: network.NetworkWeights):
|
6 |
+
if all(x in weights.w for x in ["weight"]):
|
7 |
+
return NetworkModuleIa3(net, weights)
|
8 |
+
|
9 |
+
return None
|
10 |
+
|
11 |
+
|
12 |
+
class NetworkModuleIa3(network.NetworkModule):
|
13 |
+
def __init__(self, net: network.Network, weights: network.NetworkWeights):
|
14 |
+
super().__init__(net, weights)
|
15 |
+
|
16 |
+
self.w = weights.w["weight"]
|
17 |
+
self.on_input = weights.w["on_input"].item()
|
18 |
+
|
19 |
+
def calc_updown(self, orig_weight):
|
20 |
+
w = self.w.to(orig_weight.device, dtype=orig_weight.dtype)
|
21 |
+
|
22 |
+
output_shape = [w.size(0), orig_weight.size(1)]
|
23 |
+
if self.on_input:
|
24 |
+
output_shape.reverse()
|
25 |
+
else:
|
26 |
+
w = w.reshape(-1, 1)
|
27 |
+
|
28 |
+
updown = orig_weight * w
|
29 |
+
|
30 |
+
return self.finalize_updown(updown, orig_weight, output_shape)
|
extensions-builtin/Lora/network_lokr.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
import lyco_helpers
|
4 |
+
import network
|
5 |
+
|
6 |
+
|
7 |
+
class ModuleTypeLokr(network.ModuleType):
|
8 |
+
def create_module(self, net: network.Network, weights: network.NetworkWeights):
|
9 |
+
has_1 = "lokr_w1" in weights.w or ("lokr_w1_a" in weights.w and "lokr_w1_b" in weights.w)
|
10 |
+
has_2 = "lokr_w2" in weights.w or ("lokr_w2_a" in weights.w and "lokr_w2_b" in weights.w)
|
11 |
+
if has_1 and has_2:
|
12 |
+
return NetworkModuleLokr(net, weights)
|
13 |
+
|
14 |
+
return None
|
15 |
+
|
16 |
+
|
17 |
+
def make_kron(orig_shape, w1, w2):
|
18 |
+
if len(w2.shape) == 4:
|
19 |
+
w1 = w1.unsqueeze(2).unsqueeze(2)
|
20 |
+
w2 = w2.contiguous()
|
21 |
+
return torch.kron(w1, w2).reshape(orig_shape)
|
22 |
+
|
23 |
+
|
24 |
+
class NetworkModuleLokr(network.NetworkModule):
|
25 |
+
def __init__(self, net: network.Network, weights: network.NetworkWeights):
|
26 |
+
super().__init__(net, weights)
|
27 |
+
|
28 |
+
self.w1 = weights.w.get("lokr_w1")
|
29 |
+
self.w1a = weights.w.get("lokr_w1_a")
|
30 |
+
self.w1b = weights.w.get("lokr_w1_b")
|
31 |
+
self.dim = self.w1b.shape[0] if self.w1b is not None else self.dim
|
32 |
+
self.w2 = weights.w.get("lokr_w2")
|
33 |
+
self.w2a = weights.w.get("lokr_w2_a")
|
34 |
+
self.w2b = weights.w.get("lokr_w2_b")
|
35 |
+
self.dim = self.w2b.shape[0] if self.w2b is not None else self.dim
|
36 |
+
self.t2 = weights.w.get("lokr_t2")
|
37 |
+
|
38 |
+
def calc_updown(self, orig_weight):
|
39 |
+
if self.w1 is not None:
|
40 |
+
w1 = self.w1.to(orig_weight.device, dtype=orig_weight.dtype)
|
41 |
+
else:
|
42 |
+
w1a = self.w1a.to(orig_weight.device, dtype=orig_weight.dtype)
|
43 |
+
w1b = self.w1b.to(orig_weight.device, dtype=orig_weight.dtype)
|
44 |
+
w1 = w1a @ w1b
|
45 |
+
|
46 |
+
if self.w2 is not None:
|
47 |
+
w2 = self.w2.to(orig_weight.device, dtype=orig_weight.dtype)
|
48 |
+
elif self.t2 is None:
|
49 |
+
w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype)
|
50 |
+
w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype)
|
51 |
+
w2 = w2a @ w2b
|
52 |
+
else:
|
53 |
+
t2 = self.t2.to(orig_weight.device, dtype=orig_weight.dtype)
|
54 |
+
w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype)
|
55 |
+
w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype)
|
56 |
+
w2 = lyco_helpers.make_weight_cp(t2, w2a, w2b)
|
57 |
+
|
58 |
+
output_shape = [w1.size(0) * w2.size(0), w1.size(1) * w2.size(1)]
|
59 |
+
if len(orig_weight.shape) == 4:
|
60 |
+
output_shape = orig_weight.shape
|
61 |
+
|
62 |
+
updown = make_kron(output_shape, w1, w2)
|
63 |
+
|
64 |
+
return self.finalize_updown(updown, orig_weight, output_shape)
|