File size: 14,509 Bytes
a660631
 
 
 
 
 
a9864eb
a660631
 
f521e88
 
 
 
 
 
a660631
 
 
e4d01a7
a660631
 
b5515fe
da031b9
b5515fe
da031b9
b5515fe
da031b9
b5515fe
da031b9
b5515fe
a660631
 
 
 
 
 
 
 
 
f521e88
 
 
 
a660631
 
 
 
f521e88
 
 
 
 
 
a660631
 
654fc61
a660631
654fc61
f521e88
 
3da112e
 
a660631
2e8f4d7
a660631
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f521e88
a660631
 
 
 
654fc61
a660631
 
 
 
 
 
 
 
 
 
f521e88
a660631
 
f521e88
a660631
 
 
 
 
 
 
 
 
 
c763397
 
a660631
f521e88
 
 
 
 
 
 
 
 
a660631
 
c763397
a660631
 
 
 
 
 
 
 
 
 
 
 
 
 
89efab8
 
e4d01a7
 
 
 
 
f521e88
 
 
 
a660631
f521e88
a660631
 
 
 
 
 
 
 
 
8fad46e
 
 
 
 
 
a660631
 
c763397
a660631
 
 
 
 
 
 
 
 
 
 
 
 
 
89efab8
 
e4d01a7
 
 
 
 
f521e88
a660631
 
 
f521e88
 
 
a660631
 
 
 
 
 
f521e88
 
 
a660631
 
 
 
 
 
 
 
f521e88
a660631
 
 
 
 
 
 
 
 
8fad46e
 
 
 
 
 
 
 
 
a660631
 
c763397
a660631
 
 
 
 
 
 
 
 
 
 
 
 
 
89efab8
 
e4d01a7
 
 
 
 
f521e88
a660631
 
 
 
 
 
 
 
 
 
f521e88
a660631
 
 
 
 
 
 
 
 
8fad46e
 
 
 
 
 
 
 
 
a660631
 
c763397
a660631
 
 
 
 
 
 
 
 
 
 
 
 
 
89efab8
 
e4d01a7
 
 
 
 
f521e88
a660631
 
 
 
 
 
 
 
 
 
f521e88
a660631
 
 
 
 
 
 
 
 
8fad46e
 
 
 
 
 
 
 
a660631
 
c763397
a660631
 
 
 
 
 
 
 
 
 
 
 
 
 
89efab8
 
e4d01a7
 
 
 
 
f521e88
cce8954
a660631
 
f521e88
 
 
a660631
 
 
 
 
 
f521e88
 
a660631
 
 
 
 
da031b9
f521e88
 
a660631
f521e88
a660631
 
 
 
 
 
 
 
 
cce8954
8fad46e
 
 
 
 
 
 
cce8954
 
 
 
8fad46e
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
from __future__ import annotations

import gc

import numpy as np
import PIL.Image
import spaces
import torch
from controlnet_aux.util import HWC3
from diffusers import (
    ControlNetModel,
    DiffusionPipeline,
    StableDiffusionControlNetPipeline,
    UniPCMultistepScheduler,
)

from cv_utils import resize_image
from preprocessor import Preprocessor
from settings import MAX_IMAGE_RESOLUTION, MAX_NUM_IMAGES

CONTROLNET_MODEL_IDS = {
    "Canny": "checkpoints/canny/controlnet",

    "softedge": "checkpoints/hed/controlnet",

    "segmentation": "checkpoints/seg/controlnet",

    "depth": "checkpoints/depth/controlnet",

    "lineart": "checkpoints/lineart/controlnet",
}


def download_all_controlnet_weights() -> None:
    for model_id in CONTROLNET_MODEL_IDS.values():
        ControlNetModel.from_pretrained(model_id)


class Model:
    def __init__(self, base_model_id: str = "runwayml/stable-diffusion-v1-5", task_name: str = "Canny"):
        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
        self.base_model_id = ""
        self.task_name = ""
        self.pipe = self.load_pipe(base_model_id, task_name)
        self.preprocessor = Preprocessor()

    def load_pipe(self, base_model_id: str, task_name) -> DiffusionPipeline:
        if (
            base_model_id == self.base_model_id
            and task_name == self.task_name
            and hasattr(self, "pipe")
            and self.pipe is not None
        ):
            return self.pipe
        model_id = CONTROLNET_MODEL_IDS[task_name]
        controlnet = ControlNetModel.from_pretrained(model_id, torch_dtype=torch.float16)
        pipe = StableDiffusionControlNetPipeline.from_pretrained(
            base_model_id, safety_checker=None, controlnet=controlnet, torch_dtype=torch.float16
        )
        pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
        # if self.device.type == "cuda":
        #     pipe.disable_xformers_memory_efficient_attention()
        pipe.to(self.device)

        torch.cuda.empty_cache()
        gc.collect()
        self.base_model_id = base_model_id
        self.task_name = task_name
        return pipe

    def set_base_model(self, base_model_id: str) -> str:
        if not base_model_id or base_model_id == self.base_model_id:
            return self.base_model_id
        del self.pipe
        torch.cuda.empty_cache()
        gc.collect()
        try:
            self.pipe = self.load_pipe(base_model_id, self.task_name)
        except Exception:
            self.pipe = self.load_pipe(self.base_model_id, self.task_name)
        return self.base_model_id

    def load_controlnet_weight(self, task_name: str) -> None:
        if task_name == self.task_name:
            return
        if self.pipe is not None and hasattr(self.pipe, "controlnet"):
            del self.pipe.controlnet
        torch.cuda.empty_cache()
        gc.collect()
        model_id = CONTROLNET_MODEL_IDS[task_name]
        controlnet = ControlNetModel.from_pretrained(model_id, torch_dtype=torch.float16)
        controlnet.to(self.device)
        torch.cuda.empty_cache()
        gc.collect()
        self.pipe.controlnet = controlnet
        self.task_name = task_name

    def get_prompt(self, prompt: str, additional_prompt: str) -> str:
        if not prompt:
            prompt = additional_prompt
        else:
            prompt = f"{prompt}, {additional_prompt}"
        return prompt

    @torch.autocast("cuda")
    def run_pipe(
        self,
        prompt: str,
        negative_prompt: str,
        control_image: PIL.Image.Image,
        num_images: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
    ) -> list[PIL.Image.Image]:
        self.pipe.to(self.device)
        self.pipe.controlnet.to(self.device)
        generator = torch.Generator().manual_seed(seed)
        return self.pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            guidance_scale=guidance_scale,
            num_images_per_prompt=num_images,
            num_inference_steps=num_steps,
            generator=generator,
            image=control_image,
        ).images

    @torch.inference_mode()
    @spaces.GPU(enable_queue=True)
    def process_canny(
        self,
        image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        low_threshold: int,
        high_threshold: int,
    ) -> list[PIL.Image.Image]:
        if image is None:
            raise ValueError
        if image_resolution > MAX_IMAGE_RESOLUTION:
            raise ValueError
        if num_images > MAX_NUM_IMAGES:
            raise ValueError

        self.preprocessor.load("Canny")
        control_image = self.preprocessor(
            image=image, low_threshold=low_threshold, high_threshold=high_threshold, detect_resolution=image_resolution
        )

        self.load_controlnet_weight("Canny")
        results = self.run_pipe(
            prompt=self.get_prompt(prompt, additional_prompt),
            negative_prompt=negative_prompt,
            control_image=control_image,
            num_images=num_images,
            num_steps=num_steps,
            guidance_scale=guidance_scale,
            seed=seed,
        )
        conditions_of_generated_imgs = [
            self.preprocessor(
                image=x, low_threshold=low_threshold, high_threshold=high_threshold, detect_resolution=image_resolution
            ) for x in results
        ]
        return [control_image] * num_images + results + conditions_of_generated_imgs

    @torch.inference_mode()
    @spaces.GPU(enable_queue=True)
    def process_softedge(
        self,
        image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        preprocess_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        preprocessor_name: str,
    ) -> list[PIL.Image.Image]:
        if image is None:
            raise ValueError
        if image_resolution > MAX_IMAGE_RESOLUTION:
            raise ValueError
        if num_images > MAX_NUM_IMAGES:
            raise ValueError

        if preprocessor_name == "None":
            image = HWC3(image)
            image = resize_image(image, resolution=image_resolution)
            control_image = PIL.Image.fromarray(image)
        elif preprocessor_name in ["HED", "HED safe"]:
            safe = "safe" in preprocessor_name
            self.preprocessor.load("HED")
            control_image = self.preprocessor(
                image=image,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
                scribble=safe,
            )
        elif preprocessor_name in ["PidiNet", "PidiNet safe"]:
            safe = "safe" in preprocessor_name
            self.preprocessor.load("PidiNet")
            control_image = self.preprocessor(
                image=image,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
                safe=safe,
            )
        else:
            raise ValueError
        self.load_controlnet_weight("softedge")
        results = self.run_pipe(
            prompt=self.get_prompt(prompt, additional_prompt),
            negative_prompt=negative_prompt,
            control_image=control_image,
            num_images=num_images,
            num_steps=num_steps,
            guidance_scale=guidance_scale,
            seed=seed,
        )
        conditions_of_generated_imgs = [
            self.preprocessor(
                image=x,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
                scribble=safe,
            ) for x in results
        ]
        return [control_image] * num_images + results + conditions_of_generated_imgs

    @torch.inference_mode()
    @spaces.GPU(enable_queue=True)
    def process_segmentation(
        self,
        image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        preprocess_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        preprocessor_name: str,
    ) -> list[PIL.Image.Image]:
        if image is None:
            raise ValueError
        if image_resolution > MAX_IMAGE_RESOLUTION:
            raise ValueError
        if num_images > MAX_NUM_IMAGES:
            raise ValueError

        if preprocessor_name == "None":
            image = HWC3(image)
            image = resize_image(image, resolution=image_resolution)
            control_image = PIL.Image.fromarray(image)
        else:
            self.preprocessor.load(preprocessor_name)
            control_image = self.preprocessor(
                image=image,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
            )
        self.load_controlnet_weight("segmentation")
        results = self.run_pipe(
            prompt=self.get_prompt(prompt, additional_prompt),
            negative_prompt=negative_prompt,
            control_image=control_image,
            num_images=num_images,
            num_steps=num_steps,
            guidance_scale=guidance_scale,
            seed=seed,
        )
        self.preprocessor.load('UPerNet')
        conditions_of_generated_imgs = [
            self.preprocessor(
                image=np.array(x),
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
            ) for x in results
        ]
        return [control_image] * num_images + results + conditions_of_generated_imgs

    @torch.inference_mode()
    @spaces.GPU(enable_queue=True)
    def process_depth(
        self,
        image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        preprocess_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        preprocessor_name: str,
    ) -> list[PIL.Image.Image]:
        if image is None:
            raise ValueError
        if image_resolution > MAX_IMAGE_RESOLUTION:
            raise ValueError
        if num_images > MAX_NUM_IMAGES:
            raise ValueError

        if preprocessor_name == "None":
            image = HWC3(image)
            image = resize_image(image, resolution=image_resolution)
            control_image = PIL.Image.fromarray(image)
        else:
            self.preprocessor.load(preprocessor_name)
            control_image = self.preprocessor(
                image=image,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
            )
        self.load_controlnet_weight("depth")
        results = self.run_pipe(
            prompt=self.get_prompt(prompt, additional_prompt),
            negative_prompt=negative_prompt,
            control_image=control_image,
            num_images=num_images,
            num_steps=num_steps,
            guidance_scale=guidance_scale,
            seed=seed,
        )
        conditions_of_generated_imgs = [
            self.preprocessor(
                image=x,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
            ) for x in results
        ]
        return [control_image] * num_images + results + conditions_of_generated_imgs

    @torch.inference_mode()
    @spaces.GPU(enable_queue=True)
    def process_lineart(
        self,
        image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        preprocess_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        preprocessor_name: str,
    ) -> list[PIL.Image.Image]:
        if image is None:
            raise ValueError
        if image_resolution > MAX_IMAGE_RESOLUTION:
            raise ValueError
        if num_images > MAX_NUM_IMAGES:
            raise ValueError

        if preprocessor_name in ["None", "None (anime)"]:
            image = 255 - HWC3(image)
            image = resize_image(image, resolution=image_resolution)
            control_image = PIL.Image.fromarray(image)
        elif preprocessor_name in ["Lineart", "Lineart coarse"]:
            coarse = "coarse" in preprocessor_name
            self.preprocessor.load("Lineart")
            control_image = self.preprocessor(
                image=image,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
                coarse=coarse,
            )
        elif preprocessor_name == "Lineart (anime)":
            self.preprocessor.load("LineartAnime")
            control_image = self.preprocessor(
                image=image,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
            )
        # NOTE: We still use the general lineart model
        if "anime" in preprocessor_name:
            self.load_controlnet_weight("lineart_anime")
        else:
            self.load_controlnet_weight("lineart")
        results = self.run_pipe(
            prompt=self.get_prompt(prompt, additional_prompt),
            negative_prompt=negative_prompt,
            control_image=control_image,
            num_images=num_images,
            num_steps=num_steps,
            guidance_scale=guidance_scale,
            seed=seed,
        )
        self.preprocessor.load("Lineart")
        conditions_of_generated_imgs = [
            self.preprocessor(
                image=x,
                image_resolution=image_resolution,
                detect_resolution=preprocess_resolution,
            ) for x in results
        ]

        control_image = PIL.Image.fromarray((255 - np.array(control_image)).astype(np.uint8))
        conditions_of_generated_imgs = [PIL.Image.fromarray((255 - np.array(x)).astype(np.uint8)) for x in conditions_of_generated_imgs]

        return [control_image] * num_images + results + conditions_of_generated_imgs