Disty0/sotediffusion-wuerstchen3

New verison is available: https://huggingface.co/Disty0/sotediffusion-v2

SoteDiffusion Wuerstchen3

Anime finetune of Würstchen V3.

Release Notes

• This release is sponsored by fal.ai/grants
• Trained on 6M images for 3 epochs using 8x A100 80G GPUs.

API Usage

This model can be used via API with Fal.AI
For more details: https://fal.ai/models/fal-ai/stable-cascade/sote-diffusion

UI Guide

SD.Next

URL: https://github.com/vladmandic/automatic/

Go to Models -> Huggingface and type Disty0/sotediffusion-wuerstchen3-decoder into the model name and press download.
Load Disty0/sotediffusion-wuerstchen3-decoder after the download process is complete.

Prompt:

newest, extremely aesthetic, best quality,

Negative Prompt:

very displeasing, worst quality, monochrome, realistic, oldest, loli,

Parameters:
Sampler: Default

Steps: 30 or 40
Refiner Steps: 10

CFG: 7
Secondary CFG: 2 or 1

Resolution: 1024x1536, 2048x1152
Anything works as long as it's a multiply of 128.

ComfyUI

Please refer to CivitAI: https://civitai.com/models/353284

Code Example

pip install diffusers

import torch
from diffusers import StableCascadeCombinedPipeline

device = "cuda"
dtype = torch.bfloat16 # or torch.float16
model = "Disty0/sotediffusion-wuerstchen3-decoder"

pipe = StableCascadeCombinedPipeline.from_pretrained(model, torch_dtype=dtype)

# send everything to the gpu:
pipe = pipe.to(device, dtype=dtype)
pipe.prior_pipe = pipe.prior_pipe.to(device, dtype=dtype)

# or enable model offload to save vram:
# pipe.enable_model_cpu_offload()



prompt = "newest, extremely aesthetic, best quality, 1girl, solo, cat ears, pink hair, orange eyes, long hair, bare shoulders, looking at viewer, smile, indoors, casual, living room, playing guitar,"
negative_prompt = "very displeasing, worst quality, monochrome, realistic, oldest, loli,"
output = pipe(
    width=1024,
    height=1536,
    prompt=prompt,
    negative_prompt=negative_prompt,
    decoder_guidance_scale=2.0,
    prior_guidance_scale=7.0,
    prior_num_inference_steps=30,
    output_type="pil",
    num_inference_steps=10
).images[0]

## do something with the output image

Training:

Software used: Kohya SD-Scripts with Stable Cascade branch.
https://github.com/kohya-ss/sd-scripts/tree/stable-cascade

GPU used: 8x Nvidia A100 80GB
GPU Hours: 220

Base

parameter	value
amp	bf16
weights	fp32
save weights	fp16
resolution	1024x1024
effective batch size	128
unet learning rate	1e-5
te learning rate	4e-6
optimizer	Adafactor
images	6M
epochs	3

Final

parameter	value
amp	bf16
weights	fp32
save weights	fp16
resolution	1024x1024
effective batch size	128
unet learning rate	4e-6
te learning rate	none
optimizer	Adafactor
images	120K
epochs	16

Dataset:

GPU used for captioning: 1x Intel ARC A770 16GB
GPU Hours: 350

Model used for captioning: SmilingWolf/wd-swinv2-tagger-v3
Model used for text: llava-hf/llava-1.5-7b-hf

Command:

python /mnt/DataSSD/AI/Apps/kohya_ss/sd-scripts/finetune/tag_images_by_wd14_tagger.py --model_dir "/mnt/DataSSD/AI/models/wd14_tagger_model" --repo_id "SmilingWolf/wd-swinv2-tagger-v3" --recursive --remove_underscore --use_rating_tags --character_tags_first --character_tag_expand --append_tags --onnx --caption_separator ", " --general_threshold 0.35 --character_threshold 0.50 --batch_size 4 --caption_extension ".txt" ./

dataset name	total images
newest	1.848.331
recent	1.380.630
mid	993.227
early	566.152
oldest	160.397
pixiv	343.614
visual novel cg	231.358
anime wallpaper	104.790
Total	5.628.499

Note:

• Smallest size is 1280x600 | 768.000 pixels
• Deduped based on image similarity using czkawka-cli
• Around 120K very high quality images got intentionally duplicated 5 times, making the total image count 6.2M

Tags:

Model is trained with random tag order but this is the order in the dataset if you are interested:

aesthetic tags, quality tags, date tags, custom tags, rating tags, character, series, rest of the tags

Date:

tag	date
newest	2022 to 2024
recent	2019 to 2021
mid	2015 to 2018
early	2011 to 2014
oldest	2005 to 2010

Aesthetic Tags:

Model used: shadowlilac/aesthetic-shadow-v2

score greater than	tag	count
0.90	extremely aesthetic	125.451
0.80	very aesthetic	887.382
0.70	aesthetic	1.049.857
0.50	slightly aesthetic	1.643.091
0.40	not displeasing	569.543
0.30	not aesthetic	445.188
0.20	slightly displeasing	341.424
0.10	displeasing	237.660
rest of them	very displeasing	328.712

Quality Tags:

Model used: https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/models/aes-B32-v0.pth

score greater than	tag	count
0.980	best quality	1.270.447
0.900	high quality	498.244
0.750	great quality	351.006
0.500	medium quality	366.448
0.250	normal quality	368.380
0.125	bad quality	279.050
0.025	low quality	538.958
rest of them	worst quality	1.955.966

Rating Tags:

tag	count
general	1.416.451
sensitive	3.447.664
nsfw	427.459
explicit nsfw	336.925

Custom Tags:

dataset name	custom tag
image boards	date,
text	The text says "text",
characters	character, series
pixiv	art by Display_Name,
visual novel cg	Full_VN_Name (short_3_letter_name), visual novel cg,
anime wallpaper	date, anime wallpaper,

Limitations and Bias

Bias

• This model is intended for anime illustrations.
  Realistic capabilites are not tested at all.

Limitations

• Can fall back to realistic.
  Add "realistic" tag to the negatives when this happens.
• Far shot eyes and hands can be bad.

License

SoteDiffusion models falls under Fair AI Public License 1.0-SD license, which is compatible with Stable Diffusion models’ license. Key points:

1. Modification Sharing: If you modify SoteDiffusion models, you must share both your changes and the original license.
2. Source Code Accessibility: If your modified version is network-accessible, provide a way (like a download link) for others to get the source code. This applies to derived models too.
3. Distribution Terms: Any distribution must be under this license or another with similar rules.
4. Compliance: Non-compliance must be fixed within 30 days to avoid license termination, emphasizing transparency and adherence to open-source values.

Notes: Anything not covered by Fair AI license is inherited from Stability AI Non-Commercial license which is named as LICENSE_INHERIT.