Files changed (1) hide show
  1. app.py +286 -0
app.py ADDED
@@ -0,0 +1,286 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from huggingface_hub import notebook_login
2
+ import cv2
3
+ from google.colab.patches import cv2_imshow
4
+ import tempfile
5
+
6
+ notebook_login()
7
+
8
+ import inspect
9
+ from typing import List, Optional, Union
10
+
11
+ import numpy as np
12
+ import torch
13
+
14
+ import PIL
15
+ from diffusers import AutoencoderKL, DDIMScheduler, DiffusionPipeline, PNDMScheduler, UNet2DConditionModel
16
+ from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
17
+ from tqdm.auto import tqdm
18
+ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
19
+
20
+
21
+ def preprocess_image(image):
22
+ w, h = image.size
23
+ w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
24
+ image = image.resize((w, h), resample=PIL.Image.LANCZOS)
25
+ image = np.array(image).astype(np.float32) / 255.0
26
+ image = image[None].transpose(0, 3, 1, 2)
27
+ image = torch.from_numpy(image)
28
+ return 2.0 * image - 1.0
29
+
30
+ def preprocess_mask(mask):
31
+ mask=mask.convert("L")
32
+ w, h = mask.size
33
+ w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
34
+ mask = mask.resize((w//8, h//8), resample=PIL.Image.NEAREST)
35
+ mask = np.array(mask).astype(np.float32) / 255.0
36
+ mask = np.tile(mask,(4,1,1))
37
+ mask = mask[None].transpose(0, 1, 2, 3)#what does this step do?
38
+ mask = 1 - mask #repaint white, keep black
39
+ mask = torch.from_numpy(mask)
40
+ return mask
41
+
42
+
43
+ class StableDiffusionInpaintingPipeline(DiffusionPipeline):
44
+ def __init__(
45
+ self,
46
+ vae: AutoencoderKL,
47
+ text_encoder: CLIPTextModel,
48
+ tokenizer: CLIPTokenizer,
49
+ unet: UNet2DConditionModel,
50
+ scheduler: Union[DDIMScheduler, PNDMScheduler],
51
+ safety_checker: StableDiffusionSafetyChecker,
52
+ feature_extractor: CLIPFeatureExtractor,
53
+ ):
54
+ super().__init__()
55
+ scheduler = scheduler.set_format("pt")
56
+ self.register_modules(
57
+ vae=vae,
58
+ text_encoder=text_encoder,
59
+ tokenizer=tokenizer,
60
+ unet=unet,
61
+ scheduler=scheduler,
62
+ safety_checker=safety_checker,
63
+ feature_extractor=feature_extractor,
64
+ )
65
+
66
+ @torch.no_grad()
67
+ def __call__(
68
+ self,
69
+ prompt: Union[str, List[str]],
70
+ init_image: torch.FloatTensor,
71
+ mask_image: torch.FloatTensor,
72
+ strength: float = 0.8,
73
+ num_inference_steps: Optional[int] = 50,
74
+ guidance_scale: Optional[float] = 7.5,
75
+ eta: Optional[float] = 0.0,
76
+ generator: Optional[torch.Generator] = None,
77
+ output_type: Optional[str] = "pil",
78
+ ):
79
+
80
+ if isinstance(prompt, str):
81
+ batch_size = 1
82
+ elif isinstance(prompt, list):
83
+ batch_size = len(prompt)
84
+ else:
85
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
86
+
87
+ if strength < 0 or strength > 1:
88
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
89
+
90
+ # set timesteps
91
+ accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())
92
+ extra_set_kwargs = {}
93
+ offset = 0
94
+ if accepts_offset:
95
+ offset = 1
96
+ extra_set_kwargs["offset"] = 1
97
+
98
+ self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
99
+
100
+ #preprocess image
101
+ init_image = preprocess_image(init_image).to(self.device)
102
+
103
+ # encode the init image into latents and scale the latents
104
+ init_latents = self.vae.encode(init_image).sample()
105
+ init_latents = 0.18215 * init_latents
106
+
107
+ # prepare init_latents noise to latents
108
+ init_latents = torch.cat([init_latents] * batch_size)
109
+ init_latents_orig = init_latents
110
+
111
+ # preprocess mask
112
+ mask = preprocess_mask(mask_image).to(self.device)
113
+ mask = torch.cat([mask] * batch_size)
114
+
115
+ #check sizes
116
+ if not mask.shape == init_latents.shape:
117
+ raise ValueError(f"The mask and init_image should be the same size!")
118
+
119
+
120
+ # get the original timestep using init_timestep
121
+ init_timestep = int(num_inference_steps * strength) + offset
122
+ init_timestep = min(init_timestep, num_inference_steps)
123
+ timesteps = self.scheduler.timesteps[-init_timestep]
124
+ timesteps = torch.tensor([timesteps] * batch_size, dtype=torch.long, device=self.device)
125
+
126
+ # add noise to latents using the timesteps
127
+ noise = torch.randn(init_latents.shape, generator=generator, device=self.device)
128
+ init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)
129
+
130
+ # get prompt text embeddings
131
+ text_input = self.tokenizer(
132
+ prompt,
133
+ padding="max_length",
134
+ max_length=self.tokenizer.model_max_length,
135
+ truncation=True,
136
+ return_tensors="pt",
137
+ )
138
+ text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
139
+
140
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
141
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
142
+ # corresponds to doing no classifier free guidance.
143
+ do_classifier_free_guidance = guidance_scale > 1.0
144
+ # get unconditional embeddings for classifier free guidance
145
+ if do_classifier_free_guidance:
146
+ max_length = text_input.input_ids.shape[-1]
147
+ uncond_input = self.tokenizer(
148
+ [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
149
+ )
150
+ uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
151
+
152
+ # For classifier free guidance, we need to do two forward passes.
153
+ # Here we concatenate the unconditional and text embeddings into a single batch
154
+ # to avoid doing two forward passes
155
+ text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
156
+
157
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
158
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
159
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
160
+ # and should be between [0, 1]
161
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
162
+ extra_step_kwargs = {}
163
+ if accepts_eta:
164
+ extra_step_kwargs["eta"] = eta
165
+
166
+ latents = init_latents
167
+ t_start = max(num_inference_steps - init_timestep + offset, 0)
168
+ for i, t in tqdm(enumerate(self.scheduler.timesteps[t_start:])):
169
+ # expand the latents if we are doing classifier free guidance
170
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
171
+
172
+ # predict the noise residual
173
+ noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]
174
+
175
+ # perform guidance
176
+ if do_classifier_free_guidance:
177
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
178
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
179
+
180
+ # compute the previous noisy sample x_t -> x_t-1
181
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs)["prev_sample"]
182
+
183
+ #masking
184
+ init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, t)
185
+ latents = ( init_latents_proper * mask ) + ( latents * (1-mask) )
186
+
187
+ # scale and decode the image latents with vae
188
+ latents = 1 / 0.18215 * latents
189
+ image = self.vae.decode(latents)
190
+
191
+ image = (image / 2 + 0.5).clamp(0, 1)
192
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
193
+
194
+ # run safety checker
195
+ safety_cheker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(self.device)
196
+ image, has_nsfw_concept = self.safety_checker(images=image, clip_input=safety_cheker_input.pixel_values)
197
+
198
+ if output_type == "pil":
199
+ image = self.numpy_to_pil(image)
200
+
201
+ return {"sample": image, "nsfw_content_detected": has_nsfw_concept}
202
+
203
+ device = "cuda"
204
+ model_path = "CompVis/stable-diffusion-v1-4"
205
+
206
+ pipe = StableDiffusionInpaintingPipeline.from_pretrained(
207
+ model_path,
208
+ revision="fp16",
209
+ torch_dtype=torch.float16,
210
+ use_auth_token=True
211
+ ).to(device)
212
+
213
+ import gdown
214
+ def download_gdrive_url():
215
+ url = 'https://drive.google.com/u/0/uc?id=1PPO2MCttsmSqyB-vKh5C7SumwFKuhgyj&export=download'
216
+ output = 'haarcascade_frontalface_default.xml'
217
+ gdown.download(url, output, quiet=False)
218
+
219
+ from torch import autocast
220
+ def inpaint(p, init_image, mask_image=None, strength=0.75, guidance_scale=7.5, generator=None, num_samples=1, n_iter=1):
221
+ all_images = []
222
+ for _ in range(n_iter):
223
+ with autocast("cuda"):
224
+ images = pipe(
225
+ prompt=[p] * num_samples,
226
+ init_image=init_image,
227
+ mask_image=mask_image,
228
+ strength=strength,
229
+ guidance_scale=guidance_scale,
230
+ generator=generator,
231
+ num_inference_steps=75
232
+ )["sample"]
233
+ all_images.extend(images)
234
+ print(len(all_images))
235
+ return all_images[0]
236
+
237
+ def identify_face(user_image):
238
+ img = cv2.imread(user_image.name) # read the resized image in cv2
239
+ print(img.shape)
240
+ gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) # convert to grayscale
241
+ download_gdrive_url() #download the haarcascade face recognition stuff
242
+ haar_cascade = cv2.CascadeClassifier('haarcascade_frontalface_default.xml')
243
+ faces_rect = haar_cascade.detectMultiScale(gray_img, scaleFactor=1.1, minNeighbors=9)
244
+ for (x, y, w, h) in faces_rect[:1]:
245
+ mask = np.zeros(img.shape[:2], dtype="uint8")
246
+ print(mask.shape)
247
+ cv2.rectangle(mask, (x, y), (x+w, y+h), 255, -1)
248
+ print(mask.shape)
249
+ inverted_image = cv2.bitwise_not(mask)
250
+ return inverted_image
251
+
252
+ def sample_images(init_image, mask_image):
253
+ p = "4K UHD professional profile picture of a person wearing a suit for work"
254
+ strength=0.65
255
+ guidance_scale=10
256
+ num_samples = 1
257
+ n_iter = 1
258
+
259
+ generator = torch.Generator(device="cuda").manual_seed(random.randint(0, 1000000)) # change the seed to get different results
260
+ all_images = inpaint(p, init_image, mask_image, strength=strength, guidance_scale=guidance_scale, generator=generator, num_samples=num_samples, n_iter=n_iter)
261
+ return all_images
262
+
263
+ import gradio as gr
264
+ import random
265
+ # accept an image input
266
+ # trigger the set of functions to occur => identify face, generate mask, save the inverted face mask, sample for the inverted images
267
+ # output the sampled images
268
+ def main(user_image):
269
+ # accept the image as input
270
+ init_image = PIL.Image.open(user_image).convert("RGB")
271
+ # # resize the image to be (512, 512)
272
+ newsize = (512, 512)
273
+ init_image = init_image.resize(newsize)
274
+ init_image.save(user_image.name) # save the resized image
275
+ ## identify the face + save the inverted mask
276
+ inverted_mask = identify_face(user_image)
277
+ fp = tempfile.NamedTemporaryFile(mode='wb', suffix=".png")
278
+ cv2.imwrite(fp.name, inverted_mask) # save the inverted image mask
279
+ pil_inverted_mask = PIL.Image.open(fp.name).convert("RGB")
280
+ print("type(init_image): ", type(init_image))
281
+ print("type(pil_inverted_mask): ", type(pil_inverted_mask))
282
+ # sample the new images
283
+ return sample_images(init_image, pil_inverted_mask)
284
+
285
+ demo = gr.Interface(main, gr.Image(type="file"), "image")
286
+ demo.launch(debug=True)