File size: 23,179 Bytes
473b850
 
 
 
faa4cbd
473b850
 
 
 
 
b44d4b1
473b850
b44d4b1
 
 
473b850
b44d4b1
 
 
473b850
b554f48
 
 
 
 
 
 
 
 
 
473b850
b44d4b1
 
 
 
 
 
 
 
 
473b850
ef87020
b44d4b1
 
 
 
473b850
 
 
 
b44d4b1
 
faa4cbd
 
b44d4b1
473b850
b44d4b1
 
 
b554f48
 
f863ebd
 
b44d4b1
 
e8089b6
 
b44d4b1
e8089b6
 
 
 
b44d4b1
 
 
faa4cbd
 
 
b44d4b1
 
 
 
 
faa4cbd
 
 
 
 
 
 
 
 
b44d4b1
473b850
b44d4b1
473b850
 
cbc964f
 
faa4cbd
 
b44d4b1
e8089b6
 
 
faa4cbd
 
b44d4b1
473b850
 
b44d4b1
 
 
 
 
 
 
e8089b6
b44d4b1
 
 
 
 
 
 
 
 
 
d1c233a
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
b44d4b1
 
473b850
b44d4b1
 
473b850
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
b44d4b1
 
 
 
 
 
 
 
473b850
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
b44d4b1
 
 
 
473b850
b44d4b1
 
 
 
 
 
473b850
b44d4b1
473b850
b44d4b1
 
473b850
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
b44d4b1
 
 
 
 
 
 
 
 
4af6431
b44d4b1
 
 
473b850
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
b44d4b1
 
 
 
 
 
 
 
4af6431
b44d4b1
 
473b850
 
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4af6431
b44d4b1
 
473b850
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473b850
b44d4b1
 
 
 
 
 
 
 
 
 
4af6431
b44d4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
# This file is adapted from gradio_*.py in https://github.com/lllyasviel/ControlNet/tree/f4748e3630d8141d7765e2bd9b1e348f47847707
# The original license file is LICENSE.ControlNet in this repo.
from __future__ import annotations

import gc
import pathlib
import sys

import cv2
import numpy as np
import PIL.Image
import torch
from diffusers import (ControlNetModel, DiffusionPipeline,
                       StableDiffusionControlNetPipeline,
                       UniPCMultistepScheduler)

repo_dir = pathlib.Path(__file__).parent
submodule_dir = repo_dir / 'ControlNet'
sys.path.append(submodule_dir.as_posix())

try:
    from annotator.canny import apply_canny
    from annotator.hed import apply_hed, nms
    from annotator.midas import apply_midas
    from annotator.mlsd import apply_mlsd
    from annotator.openpose import apply_openpose
    from annotator.uniformer import apply_uniformer
    from annotator.util import HWC3, resize_image
except Exception:
    pass

CONTROLNET_MODEL_IDS = {
    'canny': 'lllyasviel/sd-controlnet-canny',
    'hough': 'lllyasviel/sd-controlnet-mlsd',
    'hed': 'lllyasviel/sd-controlnet-hed',
    'scribble': 'lllyasviel/sd-controlnet-scribble',
    'pose': 'lllyasviel/sd-controlnet-openpose',
    'seg': 'lllyasviel/sd-controlnet-seg',
    'depth': 'lllyasviel/sd-controlnet-depth',
    'normal': 'lllyasviel/sd-controlnet-normal',
}


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)

    def load_pipe(self, base_model_id: str, task_name) -> DiffusionPipeline:
        if self.device.type == 'cpu':
            return None
        if base_model_id == self.base_model_id and task_name == self.task_name and hasattr(
                self, 'pipe'):
            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)
        pipe.enable_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 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]:
        if seed == -1:
            seed = np.random.randint(0, np.iinfo(np.int64).max)
        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

    @staticmethod
    def preprocess_canny(
        input_image: np.ndarray,
        image_resolution: int,
        low_threshold: int,
        high_threshold: int,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        image = resize_image(HWC3(input_image), image_resolution)
        control_image = apply_canny(image, low_threshold, high_threshold)
        control_image = HWC3(control_image)
        vis_control_image = 255 - control_image
        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            vis_control_image)

    @torch.inference_mode()
    def process_canny(
        self,
        input_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]:
        control_image, vis_control_image = self.preprocess_canny(
            input_image=input_image,
            image_resolution=image_resolution,
            low_threshold=low_threshold,
            high_threshold=high_threshold,
        )
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_hough(
        input_image: np.ndarray,
        image_resolution: int,
        detect_resolution: int,
        value_threshold: float,
        distance_threshold: float,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        input_image = HWC3(input_image)
        control_image = apply_mlsd(
            resize_image(input_image, detect_resolution), value_threshold,
            distance_threshold)
        control_image = HWC3(control_image)
        image = resize_image(input_image, image_resolution)
        H, W = image.shape[:2]
        control_image = cv2.resize(control_image, (W, H),
                                   interpolation=cv2.INTER_NEAREST)

        vis_control_image = 255 - cv2.dilate(
            control_image, np.ones(shape=(3, 3), dtype=np.uint8), iterations=1)

        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            vis_control_image)

    @torch.inference_mode()
    def process_hough(
        self,
        input_image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        detect_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        value_threshold: float,
        distance_threshold: float,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_hough(
            input_image=input_image,
            image_resolution=image_resolution,
            detect_resolution=detect_resolution,
            value_threshold=value_threshold,
            distance_threshold=distance_threshold,
        )
        self.load_controlnet_weight('hough')
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_hed(
        input_image: np.ndarray,
        image_resolution: int,
        detect_resolution: int,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        input_image = HWC3(input_image)
        control_image = apply_hed(resize_image(input_image, detect_resolution))
        control_image = HWC3(control_image)
        image = resize_image(input_image, image_resolution)
        H, W = image.shape[:2]
        control_image = cv2.resize(control_image, (W, H),
                                   interpolation=cv2.INTER_LINEAR)
        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            control_image)

    @torch.inference_mode()
    def process_hed(
        self,
        input_image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        detect_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_hed(
            input_image=input_image,
            image_resolution=image_resolution,
            detect_resolution=detect_resolution,
        )
        self.load_controlnet_weight('hed')
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_scribble(
        input_image: np.ndarray,
        image_resolution: int,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        image = resize_image(HWC3(input_image), image_resolution)
        control_image = np.zeros_like(image, dtype=np.uint8)
        control_image[np.min(image, axis=2) < 127] = 255
        vis_control_image = 255 - control_image
        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            vis_control_image)

    @torch.inference_mode()
    def process_scribble(
        self,
        input_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,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_scribble(
            input_image=input_image,
            image_resolution=image_resolution,
        )
        self.load_controlnet_weight('scribble')
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_scribble_interactive(
        input_image: np.ndarray,
        image_resolution: int,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        image = resize_image(HWC3(input_image['mask'][:, :, 0]),
                             image_resolution)
        control_image = np.zeros_like(image, dtype=np.uint8)
        control_image[np.min(image, axis=2) > 127] = 255
        vis_control_image = 255 - control_image
        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            vis_control_image)

    @torch.inference_mode()
    def process_scribble_interactive(
        self,
        input_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,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_scribble_interactive(
            input_image=input_image,
            image_resolution=image_resolution,
        )
        self.load_controlnet_weight('scribble')
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_fake_scribble(
        input_image: np.ndarray,
        image_resolution: int,
        detect_resolution: int,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        input_image = HWC3(input_image)
        control_image = apply_hed(resize_image(input_image, detect_resolution))
        control_image = HWC3(control_image)
        image = resize_image(input_image, image_resolution)
        H, W = image.shape[:2]

        control_image = cv2.resize(control_image, (W, H),
                                   interpolation=cv2.INTER_LINEAR)
        control_image = nms(control_image, 127, 3.0)
        control_image = cv2.GaussianBlur(control_image, (0, 0), 3.0)
        control_image[control_image > 4] = 255
        control_image[control_image < 255] = 0

        vis_control_image = 255 - control_image

        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            vis_control_image)

    @torch.inference_mode()
    def process_fake_scribble(
        self,
        input_image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        detect_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_fake_scribble(
            input_image=input_image,
            image_resolution=image_resolution,
            detect_resolution=detect_resolution,
        )
        self.load_controlnet_weight('scribble')
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_pose(
        input_image: np.ndarray,
        image_resolution: int,
        detect_resolution: int,
        is_pose_image: bool,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        input_image = HWC3(input_image)
        if not is_pose_image:
            control_image, _ = apply_openpose(
                resize_image(input_image, detect_resolution))
            control_image = HWC3(control_image)
            image = resize_image(input_image, image_resolution)
            H, W = image.shape[:2]
            control_image = cv2.resize(control_image, (W, H),
                                       interpolation=cv2.INTER_NEAREST)
        else:
            control_image = resize_image(input_image, image_resolution)

        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            control_image)

    @torch.inference_mode()
    def process_pose(
        self,
        input_image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        detect_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        is_pose_image: bool,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_pose(
            input_image=input_image,
            image_resolution=image_resolution,
            detect_resolution=detect_resolution,
            is_pose_image=is_pose_image,
        )
        self.load_controlnet_weight('pose')
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_seg(
        input_image: np.ndarray,
        image_resolution: int,
        detect_resolution: int,
        is_segmentation_map: bool,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        input_image = HWC3(input_image)
        if not is_segmentation_map:
            control_image = apply_uniformer(
                resize_image(input_image, detect_resolution))
            image = resize_image(input_image, image_resolution)
            H, W = image.shape[:2]
            control_image = cv2.resize(control_image, (W, H),
                                       interpolation=cv2.INTER_NEAREST)
        else:
            control_image = resize_image(input_image, image_resolution)
        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            control_image)

    @torch.inference_mode()
    def process_seg(
        self,
        input_image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        detect_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        is_segmentation_map: bool,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_seg(
            input_image=input_image,
            image_resolution=image_resolution,
            detect_resolution=detect_resolution,
            is_segmentation_map=is_segmentation_map,
        )
        self.load_controlnet_weight('seg')
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_depth(
        input_image: np.ndarray,
        image_resolution: int,
        detect_resolution: int,
        is_depth_image: bool,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        input_image = HWC3(input_image)
        if not is_depth_image:
            control_image, _ = apply_midas(
                resize_image(input_image, detect_resolution))
            control_image = HWC3(control_image)
            image = resize_image(input_image, image_resolution)
            H, W = image.shape[:2]
            control_image = cv2.resize(control_image, (W, H),
                                       interpolation=cv2.INTER_LINEAR)
        else:
            control_image = resize_image(input_image, image_resolution)
        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            control_image)

    @torch.inference_mode()
    def process_depth(
        self,
        input_image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        detect_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        is_depth_image: bool,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_depth(
            input_image=input_image,
            image_resolution=image_resolution,
            detect_resolution=detect_resolution,
            is_depth_image=is_depth_image,
        )
        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,
        )
        return [vis_control_image] + results

    @staticmethod
    def preprocess_normal(
        input_image: np.ndarray,
        image_resolution: int,
        detect_resolution: int,
        bg_threshold: float,
        is_normal_image: bool,
    ) -> tuple[PIL.Image.Image, PIL.Image.Image]:
        input_image = HWC3(input_image)
        if not is_normal_image:
            _, control_image = apply_midas(resize_image(
                input_image, detect_resolution),
                                           bg_th=bg_threshold)
            control_image = HWC3(control_image)
            image = resize_image(input_image, image_resolution)
            H, W = image.shape[:2]
            control_image = cv2.resize(control_image, (W, H),
                                       interpolation=cv2.INTER_LINEAR)
        else:
            control_image = resize_image(input_image, image_resolution)
        return PIL.Image.fromarray(control_image), PIL.Image.fromarray(
            control_image)

    @torch.inference_mode()
    def process_normal(
        self,
        input_image: np.ndarray,
        prompt: str,
        additional_prompt: str,
        negative_prompt: str,
        num_images: int,
        image_resolution: int,
        detect_resolution: int,
        num_steps: int,
        guidance_scale: float,
        seed: int,
        bg_threshold: float,
        is_normal_image: bool,
    ) -> list[PIL.Image.Image]:
        control_image, vis_control_image = self.preprocess_normal(
            input_image=input_image,
            image_resolution=image_resolution,
            detect_resolution=detect_resolution,
            bg_threshold=bg_threshold,
            is_normal_image=is_normal_image,
        )
        self.load_controlnet_weight('normal')
        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,
        )
        return [vis_control_image] + results