majian0318
commited on
Commit
•
015a181
1
Parent(s):
4651ded
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,628 @@
|
|
1 |
-
---
|
2 |
-
license: apache-2.0
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- zh
|
5 |
+
- en
|
6 |
+
- fr
|
7 |
+
- de
|
8 |
+
- ja
|
9 |
+
- kg
|
10 |
+
base_model:
|
11 |
+
- stabilityai/stable-diffusion-xl-base-1.0
|
12 |
+
pipeline_tag: text-to-image
|
13 |
+
---
|
14 |
+
|
15 |
+
|
16 |
+
![FLUX.1 [schnell] Grid](./PEA-Diffusion.png)
|
17 |
+
|
18 |
+
|
19 |
+
Text-to-image diffusion models are well-known for their ability to generate realistic images based on textual prompts. However, the existing works have predominantly focused on English, lacking support for non-English text-to-image models. The most commonly used translation methods cannot solve the generation problem related to language culture, while training from scratch on a specific language dataset is prohibitively expensive. In this paper, we are inspired to propose a simple plug-and-play language transfer method based on knowledge distillation. All we need to do is train a lightweight MLP-like parameter-efficient adapter (PEA) with only 6M parameters under teacher knowledge distillation along with a small parallel data corpus. We are surprised to find that freezing the parameters of UNet can still achieve remarkable performance on the language-specific prompt evaluation set, demonstrating that PEA can stimulate the potential generation ability of the original UNet. Additionally, it closely approaches the performance of the English text-to-image model on a general prompt evaluation set. Furthermore, our adapter can be used as a plugin to achieve significant results in downstream tasks in cross-lingual text-to-image generation.
|
20 |
+
|
21 |
+
# Usage
|
22 |
+
We provide examples of adapters for models such as [SDXL](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0), [Playground v2.5](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic), and [stable-cascade](https://huggingface.co/stabilityai/stable-cascade). For SD3, please refer directly to https://huggingface.co/OPPOer/MultilingualSD3-adapter, and for FLUX. 1, please refer to https://huggingface.co/OPPOer/MultilingualFLUX.1-adapter
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
## `SDXL`
|
27 |
+
We used the multilingual encoder [Mul-OpenCLIP](https://huggingface.co/laion/CLIP-ViT-H-14-frozen-xlm-roberta-large-laion5B-s13B-b90k).
|
28 |
+
As mentioned in the article, you can replace the model here with any SDXL derived model, including sampling acceleration, which can also be directly adapted.
|
29 |
+
|
30 |
+
```python
|
31 |
+
import os
|
32 |
+
import torch
|
33 |
+
import torch.nn as nn
|
34 |
+
|
35 |
+
from PIL import Image
|
36 |
+
from diffusers import AutoencoderKL, StableDiffusionXLPipeline,DPMSolverMultistepScheduler
|
37 |
+
from diffusers.image_processor import VaeImageProcessor
|
38 |
+
from diffusers.models.attention_processor import (
|
39 |
+
AttnProcessor2_0,
|
40 |
+
LoRAAttnProcessor2_0,
|
41 |
+
LoRAXFormersAttnProcessor,
|
42 |
+
XFormersAttnProcessor,
|
43 |
+
)
|
44 |
+
|
45 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
46 |
+
import open_clip
|
47 |
+
|
48 |
+
|
49 |
+
def image_grid(imgs, rows, cols):
|
50 |
+
assert len(imgs) == rows*cols
|
51 |
+
|
52 |
+
w, h = imgs[0].size
|
53 |
+
grid = Image.new('RGB', size=(cols*w, rows*h))
|
54 |
+
grid_w, grid_h = grid.size
|
55 |
+
|
56 |
+
for i, img in enumerate(imgs):
|
57 |
+
grid.paste(img, box=(i%cols*w, i//cols*h))
|
58 |
+
return grid
|
59 |
+
|
60 |
+
class MLP(nn.Module):
|
61 |
+
def __init__(self, in_dim, out_dim, hidden_dim,out_dim1, use_residual=True):
|
62 |
+
super().__init__()
|
63 |
+
if use_residual:
|
64 |
+
assert in_dim == out_dim
|
65 |
+
self.layernorm = nn.LayerNorm(in_dim)
|
66 |
+
self.fc1 = nn.Linear(in_dim, hidden_dim)
|
67 |
+
self.fc2 = nn.Linear(hidden_dim, out_dim)
|
68 |
+
self.fc3 = nn.Linear(out_dim, out_dim1)
|
69 |
+
self.use_residual = use_residual
|
70 |
+
self.act_fn = nn.GELU()
|
71 |
+
|
72 |
+
def forward(self, x):
|
73 |
+
residual = x
|
74 |
+
x = self.layernorm(x)
|
75 |
+
x = self.fc1(x)
|
76 |
+
x = self.act_fn(x)
|
77 |
+
x = self.fc2(x)
|
78 |
+
x2 = self.act_fn(x)
|
79 |
+
x2 = self.fc3(x2)
|
80 |
+
if self.use_residual:
|
81 |
+
x = x + residual
|
82 |
+
x1 = torch.mean(x,1)
|
83 |
+
return x1,x2
|
84 |
+
|
85 |
+
|
86 |
+
class StableDiffusionTest():
|
87 |
+
|
88 |
+
def __init__(self, model_id,text_text_encoder_pathpath,proj_path):
|
89 |
+
super().__init__()
|
90 |
+
self.text_encoder, _, preprocess = open_clip.create_model_and_transforms('xlm-roberta-large-ViT-H-14', pretrained=text_encoder_path)
|
91 |
+
self.tokenizer = open_clip.get_tokenizer('xlm-roberta-large-ViT-H-14')
|
92 |
+
self.text_encoder.text.output_tokens = True
|
93 |
+
self.proj = MLP(1024, 1280, 1024,2048, use_residual=False).to(device,dtype=dtype)
|
94 |
+
self.text_encoder = self.text_encoder.to(device)
|
95 |
+
|
96 |
+
self.vae = AutoencoderKL.from_pretrained(model_id, subfolder="vae").to(device)
|
97 |
+
scheduler = DPMSolverMultistepScheduler.from_pretrained(model_id, subfolder="scheduler")
|
98 |
+
self.pipe = StableDiffusionXLPipeline.from_pretrained(model_id, scheduler=scheduler,torch_dtype=dtype).to(device)
|
99 |
+
self.image_processor = VaeImageProcessor(vae_scale_factor=self.pipe.vae_scale_factor)
|
100 |
+
self.proj.load_state_dict(torch.load(proj_path, map_location="cpu"))
|
101 |
+
|
102 |
+
|
103 |
+
def encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
|
104 |
+
batch_size = len(prompt) if isinstance(prompt, list) else 1
|
105 |
+
|
106 |
+
text_input_ids = self.tokenizer(prompt).to(device,dtype=dtype)
|
107 |
+
_,text_embeddings = self.text_encoder.encode_text(text_input_ids)
|
108 |
+
|
109 |
+
add_text_embeds,text_embeddings_2048 = self.proj(text_embeddings)
|
110 |
+
|
111 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
112 |
+
bs_embed, seq_len, _ = text_embeddings.shape
|
113 |
+
text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
|
114 |
+
text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
115 |
+
|
116 |
+
# get unconditional embeddings for classifier free guidance
|
117 |
+
if do_classifier_free_guidance:
|
118 |
+
uncond_tokens: List[str]
|
119 |
+
if negative_prompt is None:
|
120 |
+
uncond_tokens = [""] * batch_size
|
121 |
+
elif type(prompt) is not type(negative_prompt):
|
122 |
+
raise TypeError(
|
123 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
124 |
+
f" {type(prompt)}."
|
125 |
+
)
|
126 |
+
elif isinstance(negative_prompt, str):
|
127 |
+
uncond_tokens = [negative_prompt]
|
128 |
+
elif batch_size != len(negative_prompt):
|
129 |
+
raise ValueError(
|
130 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
131 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
132 |
+
" the batch size of `prompt`."
|
133 |
+
)
|
134 |
+
else:
|
135 |
+
uncond_tokens = negative_prompt
|
136 |
+
|
137 |
+
max_length = text_input_ids.shape[-1]
|
138 |
+
|
139 |
+
uncond_input_ids = self.tokenizer(uncond_tokens).to(device)
|
140 |
+
_,uncond_embeddings = self.text_encoder.encode_text(uncond_input_ids)
|
141 |
+
|
142 |
+
add_text_embeds_uncond,uncond_embeddings_2048 = self.proj(uncond_embeddings)
|
143 |
+
|
144 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
145 |
+
seq_len = uncond_embeddings_2048.shape[1]
|
146 |
+
uncond_embeddings_2048 = uncond_embeddings_2048.repeat(1, num_images_per_prompt, 1)
|
147 |
+
uncond_embeddings_2048 = uncond_embeddings_2048.view(batch_size * num_images_per_prompt, seq_len, -1)
|
148 |
+
|
149 |
+
text_embeddings_2048 = torch.cat([uncond_embeddings_2048, text_embeddings_2048])
|
150 |
+
add_text_embeds = torch.cat([add_text_embeds_uncond, add_text_embeds])
|
151 |
+
|
152 |
+
return text_embeddings_2048,add_text_embeds
|
153 |
+
|
154 |
+
def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
|
155 |
+
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
156 |
+
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
|
157 |
+
return add_time_ids
|
158 |
+
|
159 |
+
|
160 |
+
@torch.no_grad()
|
161 |
+
def __call__(
|
162 |
+
self,
|
163 |
+
prompt: Union[str, List[str]],
|
164 |
+
height: Optional[int] = 1024,
|
165 |
+
width: Optional[int] = 1024,
|
166 |
+
num_inference_steps: int = 30,
|
167 |
+
guidance_scale: float = 7.5,
|
168 |
+
original_size: Optional[Tuple[int, int]] = None,
|
169 |
+
crops_coords_top_left: Tuple[int, int] = (0, 0),
|
170 |
+
target_size: Optional[Tuple[int, int]] = None,
|
171 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
172 |
+
guidance_rescale: float = 0,
|
173 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
174 |
+
num_images_per_prompt: Optional[int] = 1,
|
175 |
+
eta: float = 0.0,
|
176 |
+
generator: Optional[torch.Generator] = None,
|
177 |
+
latents: Optional[torch.FloatTensor] = None,
|
178 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
179 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
180 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
181 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
182 |
+
output_type: Optional[str] = "pil",
|
183 |
+
return_dict: bool = True,
|
184 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
185 |
+
callback_steps: Optional[int] = 1,
|
186 |
+
**kwargs,
|
187 |
+
):
|
188 |
+
# 0. Default height and width to unet
|
189 |
+
height = height or self.pipe.unet.config.sample_size * self.pipe.vae_scale_factor
|
190 |
+
width = width or self.pipe.unet.config.sample_size * self.pipe.vae_scale_factor
|
191 |
+
original_size = original_size or (height, width)
|
192 |
+
target_size = target_size or (height, width)
|
193 |
+
|
194 |
+
# 1. Check inputs. Raise error if not correct
|
195 |
+
# self.pipe.check_inputs(prompt, height, width, callback_steps)
|
196 |
+
|
197 |
+
# 2. Define call parameters
|
198 |
+
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
199 |
+
device = self.pipe._execution_device
|
200 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
201 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
202 |
+
# corresponds to doing no classifier free guidance.
|
203 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
204 |
+
|
205 |
+
# 3. Encode input prompt
|
206 |
+
|
207 |
+
prompt_embeds,add_text_embeds = self.encode_prompt(prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt)
|
208 |
+
prompt_embeds = prompt_embeds
|
209 |
+
add_text_embeds = add_text_embeds
|
210 |
+
|
211 |
+
# 4. Prepare timesteps
|
212 |
+
self.pipe.scheduler.set_timesteps(num_inference_steps, device=device)
|
213 |
+
timesteps = self.pipe.scheduler.timesteps
|
214 |
+
|
215 |
+
# 5. Prepare latent variables
|
216 |
+
num_channels_latents = self.pipe.unet.in_channels
|
217 |
+
latents = self.pipe.prepare_latents(
|
218 |
+
batch_size * num_images_per_prompt,
|
219 |
+
num_channels_latents,
|
220 |
+
height,
|
221 |
+
width,
|
222 |
+
prompt_embeds.dtype,
|
223 |
+
device,
|
224 |
+
generator,
|
225 |
+
latents,
|
226 |
+
)
|
227 |
+
|
228 |
+
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
229 |
+
extra_step_kwargs = self.pipe.prepare_extra_step_kwargs(generator, eta)
|
230 |
+
|
231 |
+
add_time_ids = self._get_add_time_ids(original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype)
|
232 |
+
if do_classifier_free_guidance:
|
233 |
+
add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
|
234 |
+
|
235 |
+
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
|
236 |
+
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
237 |
+
|
238 |
+
# 7. Denoising loop
|
239 |
+
for i, t in enumerate(self.pipe.progress_bar(timesteps)):
|
240 |
+
# expand the latents if we are doing classifier free guidance
|
241 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
242 |
+
latent_model_input = self.pipe.scheduler.scale_model_input(latent_model_input, t)
|
243 |
+
|
244 |
+
# predict the noise residual
|
245 |
+
noise_pred = self.pipe.unet(
|
246 |
+
latent_model_input,
|
247 |
+
t,
|
248 |
+
encoder_hidden_states=prompt_embeds,
|
249 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
250 |
+
added_cond_kwargs=added_cond_kwargs,
|
251 |
+
return_dict=False,
|
252 |
+
)[0]
|
253 |
+
|
254 |
+
# noise_pred = self.pipe.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
|
255 |
+
|
256 |
+
# perform guidance
|
257 |
+
if do_classifier_free_guidance:
|
258 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
259 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
260 |
+
|
261 |
+
# compute the previous noisy sample x_t -> x_t-1
|
262 |
+
# latents = self.pipe.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
263 |
+
latents = self.pipe.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
264 |
+
|
265 |
+
# call the callback, if provided
|
266 |
+
if callback is not None and i % callback_steps == 0:
|
267 |
+
callback(i, t, latents)
|
268 |
+
|
269 |
+
self.vae.to(dtype=torch.float32)
|
270 |
+
|
271 |
+
use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
|
272 |
+
AttnProcessor2_0,
|
273 |
+
XFormersAttnProcessor,
|
274 |
+
LoRAXFormersAttnProcessor,
|
275 |
+
LoRAAttnProcessor2_0,
|
276 |
+
]
|
277 |
+
# if xformers or torch_2_0 is used attention block does not need
|
278 |
+
# to be in float32 which can save lots of memory
|
279 |
+
if not use_torch_2_0_or_xformers:
|
280 |
+
self.vae.post_quant_conv.to(latents.dtype)
|
281 |
+
self.vae.decoder.conv_in.to(latents.dtype)
|
282 |
+
self.vae.decoder.mid_block.to(latents.dtype)
|
283 |
+
else:
|
284 |
+
latents = latents.float()
|
285 |
+
|
286 |
+
# 8. Post-processing
|
287 |
+
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
288 |
+
image = self.image_processor.postprocess(image, output_type="np")
|
289 |
+
|
290 |
+
# 10. Convert to PIL
|
291 |
+
if output_type == "pil":
|
292 |
+
image = self.pipe.numpy_to_pil(image)
|
293 |
+
|
294 |
+
return image
|
295 |
+
|
296 |
+
|
297 |
+
if __name__ == '__main__':
|
298 |
+
device = "cuda"
|
299 |
+
dtype = torch.float16
|
300 |
+
|
301 |
+
text_encoder_path = 'laion/CLIP-ViT-H-14-frozen-xlm-roberta-large-laion5B-s13B-b90k/open_clip_pytorch_model.bin'
|
302 |
+
model_id = "stablediffusionapi/protovision-xl-v6.6"
|
303 |
+
proj_path = "OPPOer/PEA-Diffusion/pytorch_model.bin"
|
304 |
+
|
305 |
+
sdt = StableDiffusionTest(model_id,text_encoder_path,proj_path)
|
306 |
+
|
307 |
+
batch=2
|
308 |
+
height = 1024
|
309 |
+
width = 1024
|
310 |
+
while True:
|
311 |
+
raw_text = input("\nPlease Input Query (stop to exit) >>> ")
|
312 |
+
if not raw_text:
|
313 |
+
print('Query should not be empty!')
|
314 |
+
continue
|
315 |
+
if raw_text == "stop":
|
316 |
+
break
|
317 |
+
images = sdt([raw_text]*batch,height=height,width=width)
|
318 |
+
grid = image_grid(images, rows=1, cols=batch)
|
319 |
+
grid.save("SDXL.png")
|
320 |
+
|
321 |
+
```
|
322 |
+
|
323 |
+
|
324 |
+
|
325 |
+
|
326 |
+
|
327 |
+
## `Playground v2.5`
|
328 |
+
We used the multilingual encoder [Mul-OpenCLIP](https://huggingface.co/laion/CLIP-ViT-H-14-frozen-xlm-roberta-large-laion5B-s13B-b90k)
|
329 |
+
|
330 |
+
```python
|
331 |
+
import os,sys
|
332 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
333 |
+
import sys
|
334 |
+
import random
|
335 |
+
from tqdm import tqdm
|
336 |
+
|
337 |
+
import torch
|
338 |
+
import torch.nn as nn
|
339 |
+
import numpy as np
|
340 |
+
|
341 |
+
import argparse
|
342 |
+
from PIL import Image
|
343 |
+
import json
|
344 |
+
from diffusers import AutoencoderKL, DiffusionPipeline
|
345 |
+
from diffusers.image_processor import VaeImageProcessor
|
346 |
+
from diffusers.models.attention_processor import (
|
347 |
+
AttnProcessor2_0,
|
348 |
+
LoRAAttnProcessor2_0,
|
349 |
+
LoRAXFormersAttnProcessor,
|
350 |
+
XFormersAttnProcessor,
|
351 |
+
)
|
352 |
+
import open_clip
|
353 |
+
|
354 |
+
|
355 |
+
def image_grid(imgs, rows, cols):
|
356 |
+
assert len(imgs) == rows*cols
|
357 |
+
|
358 |
+
w, h = imgs[0].size
|
359 |
+
grid = Image.new('RGB', size=(cols*w, rows*h))
|
360 |
+
grid_w, grid_h = grid.size
|
361 |
+
|
362 |
+
for i, img in enumerate(imgs):
|
363 |
+
grid.paste(img, box=(i%cols*w, i//cols*h))
|
364 |
+
return grid
|
365 |
+
|
366 |
+
|
367 |
+
class MLP(nn.Module):
|
368 |
+
def __init__(self, in_dim=1024, out_dim=1280, hidden_dim=2048, out_dim1=2048, use_residual=True):
|
369 |
+
super().__init__()
|
370 |
+
if use_residual:
|
371 |
+
assert in_dim == out_dim
|
372 |
+
self.layernorm = nn.LayerNorm(in_dim)
|
373 |
+
self.projector = nn.Sequential(
|
374 |
+
nn.Linear(in_dim, hidden_dim, bias=False),
|
375 |
+
nn.GELU(),
|
376 |
+
nn.Linear(hidden_dim, hidden_dim, bias=False),
|
377 |
+
nn.GELU(),
|
378 |
+
nn.Linear(hidden_dim, hidden_dim, bias=False),
|
379 |
+
nn.GELU(),
|
380 |
+
nn.Linear(hidden_dim, out_dim, bias=False),
|
381 |
+
)
|
382 |
+
self.fc = nn.Linear(out_dim, out_dim1)
|
383 |
+
self.use_residual = use_residual
|
384 |
+
def forward(self, x):
|
385 |
+
residual = x
|
386 |
+
x = self.layernorm(x)
|
387 |
+
x = self.projector(x)
|
388 |
+
x2 = nn.GELU()(x)
|
389 |
+
x2 = self.fc(x2)
|
390 |
+
if self.use_residual:
|
391 |
+
x = x + residual
|
392 |
+
x1 = torch.mean(x,1)
|
393 |
+
return x1,x2
|
394 |
+
|
395 |
+
|
396 |
+
class StableDiffusionTest():
|
397 |
+
def __init__(self, model_id,text_encoder_path,proj_path):
|
398 |
+
super().__init__()
|
399 |
+
self.text_encoder, _, preprocess = open_clip.create_model_and_transforms('xlm-roberta-large-ViT-H-14', pretrained=text_encoder_path)
|
400 |
+
self.tokenizer = open_clip.get_tokenizer('xlm-roberta-large-ViT-H-14')
|
401 |
+
self.text_encoder.text.output_tokens = True
|
402 |
+
self.text_encoder = self.text_encoder.to(device,dtype=dtype)
|
403 |
+
self.vae = AutoencoderKL.from_pretrained(model_id, subfolder="vae").to(device)
|
404 |
+
|
405 |
+
self.pipe = DiffusionPipeline.from_pretrained(model_id, subfolder="scheduler", torch_dtype=dtype, variant="fp16").to(device)
|
406 |
+
self.image_processor = VaeImageProcessor(vae_scale_factor=self.pipe.vae_scale_factor)
|
407 |
+
|
408 |
+
self.proj = MLP(1024, 1280, 2048, 2048, use_residual=False).to(device,dtype=dtype)
|
409 |
+
self.proj.load_state_dict(torch.load(proj_path, map_location="cpu"))
|
410 |
+
|
411 |
+
def encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
|
412 |
+
batch_size = len(prompt) if isinstance(prompt, list) else 1
|
413 |
+
text_input_ids = self.tokenizer(prompt).to(device)
|
414 |
+
_,text_embeddings = self.text_encoder.encode_text(text_input_ids)
|
415 |
+
add_text_embeds,text_embeddings_2048 = self.proj(text_embeddings)
|
416 |
+
|
417 |
+
bs_embed, seq_len, _ = text_embeddings.shape
|
418 |
+
text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
|
419 |
+
text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
420 |
+
|
421 |
+
if do_classifier_free_guidance:
|
422 |
+
uncond_tokens: List[str]
|
423 |
+
if negative_prompt is None:
|
424 |
+
uncond_tokens = [""] * batch_size
|
425 |
+
elif type(prompt) is not type(negative_prompt):
|
426 |
+
raise TypeError(
|
427 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
428 |
+
f" {type(prompt)}."
|
429 |
+
)
|
430 |
+
elif isinstance(negative_prompt, str):
|
431 |
+
uncond_tokens = [negative_prompt]
|
432 |
+
elif batch_size != len(negative_prompt):
|
433 |
+
raise ValueError(
|
434 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
435 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
436 |
+
" the batch size of `prompt`."
|
437 |
+
)
|
438 |
+
else:
|
439 |
+
uncond_tokens = negative_prompt
|
440 |
+
|
441 |
+
max_length = text_input_ids.shape[-1]
|
442 |
+
uncond_input_ids = self.tokenizer(uncond_tokens).to(device)
|
443 |
+
_,uncond_embeddings = self.text_encoder.encode_text(uncond_input_ids)
|
444 |
+
add_text_embeds_uncond,uncond_embeddings_2048 = self.proj(uncond_embeddings)
|
445 |
+
|
446 |
+
seq_len = uncond_embeddings_2048.shape[1]
|
447 |
+
uncond_embeddings_2048 = uncond_embeddings_2048.repeat(1, num_images_per_prompt, 1)
|
448 |
+
uncond_embeddings_2048 = uncond_embeddings_2048.view(batch_size * num_images_per_prompt, seq_len, -1)
|
449 |
+
|
450 |
+
text_embeddings_2048 = torch.cat([uncond_embeddings_2048, text_embeddings_2048])
|
451 |
+
add_text_embeds = torch.cat([add_text_embeds_uncond, add_text_embeds])
|
452 |
+
|
453 |
+
return text_embeddings_2048,add_text_embeds
|
454 |
+
|
455 |
+
def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
|
456 |
+
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
457 |
+
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
|
458 |
+
return add_time_ids
|
459 |
+
|
460 |
+
|
461 |
+
@torch.no_grad()
|
462 |
+
def __call__(
|
463 |
+
self,
|
464 |
+
prompt: Union[str, List[str]],
|
465 |
+
height: Optional[int] = 1024,
|
466 |
+
width: Optional[int] = 1024,
|
467 |
+
num_inference_steps: int = 50,
|
468 |
+
guidance_scale: float = 3,
|
469 |
+
original_size: Optional[Tuple[int, int]] = None,
|
470 |
+
crops_coords_top_left: Tuple[int, int] = (0, 0),
|
471 |
+
target_size: Optional[Tuple[int, int]] = None,
|
472 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
473 |
+
guidance_rescale: float = 0,
|
474 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
475 |
+
num_images_per_prompt: Optional[int] = 1,
|
476 |
+
eta: float = 0.0,
|
477 |
+
generator: Optional[torch.Generator] = None,
|
478 |
+
latents: Optional[torch.FloatTensor] = None,
|
479 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
480 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
481 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
482 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
483 |
+
output_type: Optional[str] = "pil",
|
484 |
+
return_dict: bool = True,
|
485 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
486 |
+
callback_steps: Optional[int] = 1,
|
487 |
+
**kwargs,
|
488 |
+
):
|
489 |
+
height = height or self.pipe.unet.config.sample_size * self.pipe.vae_scale_factor
|
490 |
+
width = width or self.pipe.unet.config.sample_size * self.pipe.vae_scale_factor
|
491 |
+
original_size = original_size or (height, width)
|
492 |
+
target_size = target_size or (height, width)
|
493 |
+
|
494 |
+
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
495 |
+
device = self.pipe._execution_device
|
496 |
+
|
497 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
498 |
+
|
499 |
+
prompt_embeds,add_text_embeds = self.encode_prompt(prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt)
|
500 |
+
|
501 |
+
self.pipe.scheduler.set_timesteps(num_inference_steps, device=device)
|
502 |
+
timesteps = self.pipe.scheduler.timesteps
|
503 |
+
num_channels_latents = self.pipe.unet.in_channels
|
504 |
+
latents = self.pipe.prepare_latents(
|
505 |
+
batch_size * num_images_per_prompt,
|
506 |
+
num_channels_latents,
|
507 |
+
height,
|
508 |
+
width,
|
509 |
+
prompt_embeds.dtype,
|
510 |
+
device,
|
511 |
+
generator,
|
512 |
+
latents,
|
513 |
+
)
|
514 |
+
|
515 |
+
extra_step_kwargs = self.pipe.prepare_extra_step_kwargs(generator, eta)
|
516 |
+
|
517 |
+
add_time_ids = self._get_add_time_ids(original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype)
|
518 |
+
if do_classifier_free_guidance:
|
519 |
+
add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
|
520 |
+
|
521 |
+
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
|
522 |
+
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
523 |
+
|
524 |
+
for i, t in enumerate(self.pipe.progress_bar(timesteps)):
|
525 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
526 |
+
latent_model_input = self.pipe.scheduler.scale_model_input(latent_model_input, t)
|
527 |
+
|
528 |
+
noise_pred = self.pipe.unet(
|
529 |
+
latent_model_input,
|
530 |
+
t,
|
531 |
+
encoder_hidden_states=prompt_embeds,
|
532 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
533 |
+
added_cond_kwargs=added_cond_kwargs,
|
534 |
+
return_dict=False,
|
535 |
+
)[0]
|
536 |
+
|
537 |
+
if do_classifier_free_guidance:
|
538 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
539 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
540 |
+
|
541 |
+
latents = self.pipe.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
542 |
+
|
543 |
+
if callback is not None and i % callback_steps == 0:
|
544 |
+
callback(i, t, latents)
|
545 |
+
|
546 |
+
self.vae.to(dtype=torch.float32)
|
547 |
+
|
548 |
+
use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
|
549 |
+
AttnProcessor2_0,
|
550 |
+
XFormersAttnProcessor,
|
551 |
+
LoRAXFormersAttnProcessor,
|
552 |
+
LoRAAttnProcessor2_0,
|
553 |
+
]
|
554 |
+
|
555 |
+
if not use_torch_2_0_or_xformers:
|
556 |
+
self.vae.post_quant_conv.to(latents.dtype)
|
557 |
+
self.vae.decoder.conv_in.to(latents.dtype)
|
558 |
+
self.vae.decoder.mid_block.to(latents.dtype)
|
559 |
+
else:
|
560 |
+
latents = latents.float()
|
561 |
+
|
562 |
+
has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
|
563 |
+
has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
|
564 |
+
if has_latents_mean and has_latents_std:
|
565 |
+
latents_mean = (
|
566 |
+
torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
|
567 |
+
)
|
568 |
+
latents_std = (
|
569 |
+
torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
|
570 |
+
)
|
571 |
+
latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
|
572 |
+
else:
|
573 |
+
latents = latents / self.vae.config.scaling_factor
|
574 |
+
|
575 |
+
image = self.vae.decode(latents, return_dict=False)[0]
|
576 |
+
image = self.image_processor.postprocess(image, output_type="np")
|
577 |
+
|
578 |
+
if output_type == "pil":
|
579 |
+
image = self.pipe.numpy_to_pil(image)
|
580 |
+
|
581 |
+
return image
|
582 |
+
|
583 |
+
|
584 |
+
if __name__ == '__main__':
|
585 |
+
device = "cuda"
|
586 |
+
dtype = torch.float16
|
587 |
+
|
588 |
+
model_id = "playgroundai/playground-v2.5-1024px-aesthetic"
|
589 |
+
text_encoder_path = 'laion/CLIP-ViT-H-14-frozen-xlm-roberta-large-laion5B-s13B-b90k/open_clip_pytorch_model.bin'
|
590 |
+
proj_path = "OPPOer/PEA-Diffusion/pytorch_model_pg.bin"
|
591 |
+
|
592 |
+
sdt = StableDiffusionTest(model_id,text_encoder_path,proj_path)
|
593 |
+
|
594 |
+
batch=2
|
595 |
+
height = 1024
|
596 |
+
width = 1024
|
597 |
+
|
598 |
+
while True:
|
599 |
+
raw_text = input("\nPlease Input Query (stop to exit) >>> ")
|
600 |
+
if not raw_text:
|
601 |
+
print('Query should not be empty!')
|
602 |
+
continue
|
603 |
+
if raw_text == "stop":
|
604 |
+
break
|
605 |
+
images = sdt([raw_text]*batch,height=height,width=width)
|
606 |
+
grid = image_grid(images, rows=1, cols=batch)
|
607 |
+
grid.save("PG.png")
|
608 |
+
|
609 |
+
|
610 |
+
```
|
611 |
+
To learn more check out the [diffusers](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux) documentation
|
612 |
+
|
613 |
+
|
614 |
+
# License
|
615 |
+
The adapter itself is Apache License 2.0, but it must follow the license of the main model.
|
616 |
+
|
617 |
+
|
618 |
+
# Citation
|
619 |
+
```
|
620 |
+
@misc{ma2023peadiffusion,
|
621 |
+
title={PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation},
|
622 |
+
author={Jian Ma and Chen Chen and Qingsong Xie and Haonan Lu},
|
623 |
+
year={2023},
|
624 |
+
eprint={2311.17086},
|
625 |
+
archivePrefix={arXiv},
|
626 |
+
primaryClass={cs.CV}
|
627 |
+
}
|
628 |
+
```
|