File size: 20,728 Bytes
a0bcaae
 
 
 
 
 
 
 
 
 
 
 
bb0f5a9
a0bcaae
 
 
 
 
 
 
 
bb0f5a9
 
 
 
 
 
 
 
 
 
 
 
a0bcaae
bb0f5a9
a0bcaae
bb0f5a9
 
 
 
 
 
 
a0bcaae
 
 
bb0f5a9
 
 
 
 
 
 
a0bcaae
 
bb0f5a9
 
 
 
a0bcaae
 
 
 
bb0f5a9
a0bcaae
 
 
 
 
 
 
 
 
 
 
 
bb0f5a9
a0bcaae
 
bb0f5a9
a0bcaae
 
 
bb0f5a9
 
a0bcaae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb0f5a9
a0bcaae
 
bb0f5a9
 
a0bcaae
 
 
bb0f5a9
 
 
 
a0bcaae
 
 
 
bb0f5a9
a0bcaae
bb0f5a9
a0bcaae
 
 
 
 
 
 
 
 
bb0f5a9
a0bcaae
 
 
 
 
bb0f5a9
 
a0bcaae
 
 
 
bb0f5a9
 
 
a0bcaae
bb0f5a9
 
a0bcaae
 
 
 
 
 
 
bb0f5a9
 
 
a0bcaae
 
bb0f5a9
 
 
a0bcaae
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
 
 
a0bcaae
bb0f5a9
a0bcaae
bb0f5a9
 
 
a0bcaae
bb0f5a9
a0bcaae
 
 
 
 
 
 
bb0f5a9
a0bcaae
 
 
bb0f5a9
a0bcaae
 
 
 
bb0f5a9
 
 
 
a0bcaae
bb0f5a9
 
a0bcaae
bb0f5a9
 
a0bcaae
 
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
 
 
a0bcaae
 
 
 
 
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
 
 
 
 
 
 
 
 
 
a0bcaae
 
 
 
 
 
bb0f5a9
 
a0bcaae
bb0f5a9
a0bcaae
 
 
 
 
bb0f5a9
a0bcaae
 
 
 
bb0f5a9
 
a0bcaae
 
 
bb0f5a9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a0bcaae
 
 
bb0f5a9
 
a0bcaae
 
bb0f5a9
 
a0bcaae
 
 
 
 
bb0f5a9
a0bcaae
bb0f5a9
a0bcaae
 
 
 
 
 
 
 
 
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
 
 
a0bcaae
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
 
 
 
 
a0bcaae
 
 
 
 
bb0f5a9
a0bcaae
 
 
 
 
 
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
 
 
 
 
 
 
a0bcaae
 
 
 
 
 
 
 
 
 
 
 
bb0f5a9
 
a0bcaae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb0f5a9
a0bcaae
 
 
 
 
 
 
 
bb0f5a9
a0bcaae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb0f5a9
 
a0bcaae
 
 
 
 
bb0f5a9
 
a0bcaae
 
 
 
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
a0bcaae
bb0f5a9
a0bcaae
 
 
 
 
 
bb0f5a9
a0bcaae
 
 
bb0f5a9
 
a0bcaae
bb0f5a9
 
 
 
a0bcaae
 
 
 
bb0f5a9
a0bcaae
bb0f5a9
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
# Copyright (c) SenseTime Research. All rights reserved.


import torch
import torch.nn.functional as F
from tqdm import tqdm
from lpips import LPIPS
import numpy as np
from torch_utils.models import Generator as bodyGAN
from torch_utils.models_face import Generator as FaceGAN
import dlib
from utils.face_alignment import align_face_for_insetgan
from utils.util import visual, tensor_to_numpy, numpy_to_tensor
import legacy
import os
import click


class InsetGAN(torch.nn.Module):
    def __init__(self, stylebody_ckpt, styleface_ckpt):
        super().__init__()

        # convert pkl to pth
        if not os.path.exists(stylebody_ckpt.replace('.pkl', '.pth')):
            legacy.convert(
                stylebody_ckpt, stylebody_ckpt.replace('.pkl', '.pth'))
        stylebody_ckpt = stylebody_ckpt.replace('.pkl', '.pth')

        if not os.path.exists(styleface_ckpt.replace('.pkl', '.pth')):
            legacy.convert(
                styleface_ckpt, styleface_ckpt.replace('.pkl', '.pth'))
        styleface_ckpt = styleface_ckpt.replace('.pkl', '.pth')

        # dual generator
        config = {"latent": 512, "n_mlp": 8, "channel_multiplier": 2}
        self.body_generator = bodyGAN(
            size=1024,
            style_dim=config["latent"],
            n_mlp=config["n_mlp"],
            channel_multiplier=config["channel_multiplier"]
        )
        self.body_generator.load_state_dict(
            torch.load(stylebody_ckpt)['g_ema'])
        self.body_generator.eval().requires_grad_(False).cuda()

        self.face_generator = FaceGAN(
            size=1024,
            style_dim=config["latent"],
            n_mlp=config["n_mlp"],
            channel_multiplier=config["channel_multiplier"]
        )
        self.face_generator.load_state_dict(
            torch.load(styleface_ckpt)['g_ema'])
        self.face_generator.eval().requires_grad_(False).cuda()
        # crop function
        self.dlib_predictor = dlib.shape_predictor(
            './pretrained_models/shape_predictor_68_face_landmarks.dat')
        self.dlib_cnn_face_detector = dlib.cnn_face_detection_model_v1(
            "pretrained_models/mmod_human_face_detector.dat")

        # criterion
        self.lpips_loss = LPIPS(net='alex').cuda().eval()
        self.l1_loss = torch.nn.L1Loss(reduction='mean')

    def loss_coarse(self, A_face, B, p1=500, p2=0.05):
        A_face = F.interpolate(A_face, size=(64, 64), mode='area')
        B = F.interpolate(B, size=(64, 64), mode='area')
        loss_l1 = p1 * self.l1_loss(A_face, B)
        loss_lpips = p2 * self.lpips_loss(A_face, B)
        return loss_l1 + loss_lpips

    @staticmethod
    def get_border_mask(A, x, spec):
        mask = torch.zeros_like(A)
        mask[:, :, :x, ] = 1
        mask[:, :, -x:, ] = 1
        mask[:, :, :, :x] = 1
        mask[:, :, :, -x:] = 1
        return mask

    @staticmethod
    def get_body_mask(A, crop, padding=4):
        mask = torch.ones_like(A)
        mask[:, :, crop[1]-padding:crop[3]+padding,
             crop[0]-padding:crop[2]+padding] = 0
        return mask

    def loss_border(self, A_face, B, p1=10000, p2=2, spec=None):
        mask = self.get_border_mask(A_face, 8, spec)
        loss_l1 = p1 * self.l1_loss(A_face*mask, B*mask)
        loss_lpips = p2 * self.lpips_loss(A_face*mask, B*mask)
        return loss_l1 + loss_lpips

    def loss_body(self, A, B, crop, p1=9000, p2=0.1):
        padding = int((crop[3] - crop[1]) / 20)
        mask = self.get_body_mask(A, crop, padding)
        loss_l1 = p1 * self.l1_loss(A*mask, B*mask)
        loss_lpips = p2 * self.lpips_loss(A*mask, B*mask)
        return loss_l1+loss_lpips

    def loss_face(self, A, B, crop, p1=5000, p2=1.75):
        mask = 1 - self.get_body_mask(A, crop)
        loss_l1 = p1 * self.l1_loss(A*mask, B*mask)
        loss_lpips = p2 * self.lpips_loss(A*mask, B*mask)
        return loss_l1+loss_lpips

    def loss_reg(self, w, w_mean, p1, w_plus_delta=None, p2=None):
        return p1 * torch.mean(((w - w_mean) ** 2)) + p2 * torch.mean(w_plus_delta ** 2)

    # FFHQ type
    def detect_face_dlib(self, img):
        # tensor to numpy array rgb uint8
        img = tensor_to_numpy(img)
        aligned_image, crop, rect = align_face_for_insetgan(img=img,
                                                            detector=self.dlib_cnn_face_detector,
                                                            predictor=self.dlib_predictor,
                                                            output_size=256)

        aligned_image = np.array(aligned_image)
        aligned_image = numpy_to_tensor(aligned_image)
        return aligned_image, crop, rect

    # joint optimization
    def dual_optimizer(self,
                       face_w,
                       body_w,
                       joint_steps=500,
                       face_initial_learning_rate=0.02,
                       body_initial_learning_rate=0.05,
                       lr_rampdown_length=0.25,
                       lr_rampup_length=0.05,
                       seed=None,
                       output_path=None,
                       video=0):
        '''
        Given a face_w, optimize a body_w with suitable body pose & shape for face_w
        '''
        def visual_(path, synth_body, synth_face, body_crop, step, both=False, init_body_with_face=None):
            tmp = synth_body.clone().detach()
            tmp[:, :, body_crop[1]:body_crop[3],
                body_crop[0]:body_crop[2]] = synth_face
            if both:
                tmp = torch.cat([synth_body, tmp], dim=3)
            save_path = os.path.join(path, f"{step:04d}.jpg")
            visual(tmp, save_path)

        def forward(face_w_opt,
                    body_w_opt,
                    face_w_delta,
                    body_w_delta,
                    body_crop,
                    update_crop=False
                    ):
            if face_w_opt.shape[1] != 18:
                face_ws = (face_w_opt).repeat([1, 18, 1])
            else:
                face_ws = face_w_opt.clone()
            face_ws = face_ws + face_w_delta
            synth_face, _ = self.face_generator(
                [face_ws], input_is_latent=True, randomize_noise=False)

            body_ws = (body_w_opt).repeat([1, 18, 1])
            body_ws = body_ws + body_w_delta
            synth_body, _ = self.body_generator(
                [body_ws], input_is_latent=True, randomize_noise=False)

            if update_crop:
                old_r = (body_crop[3]-body_crop[1]
                         ) // 2, (body_crop[2]-body_crop[0]) // 2
                _, body_crop, _ = self.detect_face_dlib(synth_body)
                center = (body_crop[1] + body_crop[3]
                          ) // 2, (body_crop[0] + body_crop[2]) // 2
                body_crop = (center[1] - old_r[1], center[0] - old_r[0],
                             center[1] + old_r[1], center[0] + old_r[0])

            synth_body_face = synth_body[:, :, body_crop[1]:body_crop[3], body_crop[0]:body_crop[2]]

            if synth_face.shape[2] > body_crop[3]-body_crop[1]:
                synth_face_resize = F.interpolate(synth_face, size=(
                    body_crop[3]-body_crop[1], body_crop[2]-body_crop[0]), mode='area')

            return synth_body, synth_body_face, synth_face, synth_face_resize, body_crop

        def update_lr(init_lr, step, num_steps, lr_rampdown_length, lr_rampup_length):
            t = step / num_steps
            lr_ramp = min(1.0, (1.0 - t) / lr_rampdown_length)
            lr_ramp = 0.5 - 0.5 * np.cos(lr_ramp * np.pi)
            lr_ramp = lr_ramp * min(1.0, t / lr_rampup_length)
            lr = init_lr * lr_ramp
            return lr

        # update output_path
        output_path = os.path.join(output_path, seed)
        os.makedirs(output_path, exist_ok=True)

        # define optimized params
        body_w_mean = self.body_generator.mean_latent(10000).detach()
        face_w_opt = face_w.clone().detach().requires_grad_(True)
        body_w_opt = body_w.clone().detach().requires_grad_(True)
        face_w_delta = torch.zeros_like(
            face_w.repeat([1, 18, 1])).requires_grad_(True)
        body_w_delta = torch.zeros_like(
            body_w.repeat([1, 18, 1])).requires_grad_(True)
        # generate ref face & body
        ref_body, _ = self.body_generator(
            [body_w.repeat([1, 18, 1])], input_is_latent=True, randomize_noise=False)
        # for inversion
        ref_face, _ = self.face_generator(
            [face_w.repeat([1, 18, 1])], input_is_latent=True, randomize_noise=False)
        # get initilized crop
        _, body_crop, _ = self.detect_face_dlib(ref_body)
        # NOTE: this is face rect only. no FFHQ type.
        _, _, face_crop = self.detect_face_dlib(ref_face)
        # create optimizer
        face_optimizer = torch.optim.Adam([face_w_opt, face_w_delta], betas=(
            0.9, 0.999), lr=face_initial_learning_rate)
        body_optimizer = torch.optim.Adam([body_w_opt, body_w_delta], betas=(
            0.9, 0.999), lr=body_initial_learning_rate)

        global_step = 0
        # Stage1: remove background of face image
        face_steps = 25
        pbar = tqdm(range(face_steps))
        for step in pbar:
            face_lr = update_lr(face_initial_learning_rate / 2, step,
                                face_steps, lr_rampdown_length, lr_rampup_length)
            for param_group in face_optimizer.param_groups:
                param_group['lr'] = face_lr
            synth_body, synth_body_face, synth_face_raw, synth_face, body_crop = forward(face_w_opt,
                                                                                         body_w_opt,
                                                                                         face_w_delta,
                                                                                         body_w_delta,
                                                                                         body_crop)
            loss_face = self.loss_face(
                synth_face_raw, ref_face, face_crop, 5000, 1.75)
            loss_coarse = self.loss_coarse(
                synth_face, synth_body_face, 50, 0.05)
            loss_border = self.loss_border(
                synth_face, synth_body_face, 1000, 0.1)
            loss = loss_coarse + loss_border + loss_face
            face_optimizer.zero_grad()
            loss.backward()
            face_optimizer.step()
            # visualization
            if video:
                visual_(output_path, synth_body,
                        synth_face, body_crop, global_step)
            pbar.set_description(
                (
                    f"face: {step:.4f}, lr: {face_lr}, loss: {loss.item():.2f}, loss_coarse: {loss_coarse.item():.2f};"
                    f"loss_border: {loss_border.item():.2f}, loss_face: {loss_face.item():.2f};"
                )
            )
            global_step += 1

        # Stage2: find a suitable body
        body_steps = 150
        pbar = tqdm(range(body_steps))
        for step in pbar:
            body_lr = update_lr(body_initial_learning_rate, step,
                                body_steps, lr_rampdown_length, lr_rampup_length)
            update_crop = True if (step % 50 == 0) else False
            # update_crop = False
            for param_group in body_optimizer.param_groups:
                param_group['lr'] = body_lr
            synth_body, synth_body_face, synth_face_raw, synth_face, body_crop = forward(face_w_opt,
                                                                                         body_w_opt,
                                                                                         face_w_delta,
                                                                                         body_w_delta,
                                                                                         body_crop,
                                                                                         update_crop=update_crop)
            loss_coarse = self.loss_coarse(
                synth_face, synth_body_face, 500, 0.05)
            loss_border = self.loss_border(
                synth_face, synth_body_face, 2500, 0)
            loss_body = self.loss_body(
                synth_body, ref_body, body_crop, 9000, 0.1)
            loss_reg = self.loss_reg(
                body_w_opt, body_w_mean, 15000, body_w_delta, 0)
            loss = loss_coarse + loss_border + loss_body + loss_reg
            body_optimizer.zero_grad()
            loss.backward()
            body_optimizer.step()

            # visualization
            if video:
                visual_(output_path, synth_body,
                        synth_face, body_crop, global_step)
            pbar.set_description(
                (
                    f"body: {step:.4f}, lr: {body_lr}, loss: {loss.item():.2f}, loss_coarse: {loss_coarse.item():.2f};"
                    f"loss_border: {loss_border.item():.2f}, loss_body: {loss_body.item():.2f}, loss_reg: {loss_reg:.2f}"
                )
            )
            global_step += 1

        # Stage3: joint optimization
        interval = 50
        joint_face_steps = joint_steps // 2
        joint_body_steps = joint_steps // 2
        face_step = 0
        body_step = 0
        pbar = tqdm(range(joint_steps))
        flag = -1
        for step in pbar:
            if step % interval == 0:
                flag += 1
            text_flag = 'optimize_face' if flag % 2 == 0 else 'optimize_body'
            synth_body, synth_body_face, synth_face_raw, synth_face, body_crop = forward(face_w_opt,
                                                                                         body_w_opt,
                                                                                         face_w_delta,
                                                                                         body_w_delta,
                                                                                         body_crop)
            if text_flag == 'optimize_face':
                face_lr = update_lr(face_initial_learning_rate, face_step,
                                    joint_face_steps, lr_rampdown_length, lr_rampup_length)
                for param_group in face_optimizer.param_groups:
                    param_group['lr'] = face_lr
                loss_face = self.loss_face(
                    synth_face_raw, ref_face, face_crop, 5000, 1.75)
                loss_coarse = self.loss_coarse(
                    synth_face, synth_body_face, 500, 0.05)
                loss_border = self.loss_border(
                    synth_face, synth_body_face, 25000, 0)
                loss = loss_coarse + loss_border + loss_face
                face_optimizer.zero_grad()
                loss.backward()
                face_optimizer.step()
                pbar.set_description(
                    (
                        f"face: {step}, lr: {face_lr:.4f}, loss: {loss.item():.2f}, loss_coarse: {loss_coarse.item():.2f};"
                        f"loss_border: {loss_border.item():.2f}, loss_face: {loss_face.item():.2f};"
                    )
                )
                face_step += 1
            else:
                body_lr = update_lr(body_initial_learning_rate, body_step,
                                    joint_body_steps, lr_rampdown_length, lr_rampup_length)
                for param_group in body_optimizer.param_groups:
                    param_group['lr'] = body_lr
                loss_coarse = self.loss_coarse(
                    synth_face, synth_body_face, 500, 0.05)
                loss_border = self.loss_border(
                    synth_face, synth_body_face, 2500, 0)
                loss_body = self.loss_body(
                    synth_body, ref_body, body_crop, 9000, 0.1)
                loss_reg = self.loss_reg(
                    body_w_opt, body_w_mean, 25000, body_w_delta, 0)
                loss = loss_coarse + loss_border + loss_body + loss_reg
                body_optimizer.zero_grad()
                loss.backward()
                body_optimizer.step()
                pbar.set_description(
                    (
                        f"body: {step}, lr: {body_lr:.4f}, loss: {loss.item():.2f}, loss_coarse: {loss_coarse.item():.2f};"
                        f"loss_border: {loss_border.item():.2f}, loss_body: {loss_body.item():.2f}, loss_reg: {loss_reg:.2f}"
                    )
                )
                body_step += 1
            if video:
                visual_(output_path, synth_body,
                        synth_face, body_crop, global_step)
            global_step += 1
        return face_w_opt.repeat([1, 18, 1])+face_w_delta, body_w_opt.repeat([1, 18, 1])+body_w_delta, body_crop


"""
Jointly combine and optimize generated faces and bodies .
Examples:

\b
# Combine the generate human full-body image from the provided StyleGAN-Human pre-trained model
# and the generated face image from FFHQ model, optimize both latent codes to produce the coherent face-body image
python insetgan.py --body_network=pretrained_models/stylegan_human_v2_1024.pkl --face_network=pretrained_models/ffhq.pkl \\
    --body_seed=82 --face_seed=43  --trunc=0.6 --outdir=outputs/insetgan/ --video 1 
"""


@click.command()
@click.pass_context
@click.option('--face_network', default="./pretrained_models/ffhq.pkl", help='Network pickle filename', required=True)
@click.option('--body_network', default='./pretrained_models/stylegan2_1024.pkl', help='Network pickle filename', required=True)
@click.option('--face_seed', type=int, default=82, help='selected random seed')
@click.option('--body_seed', type=int, default=43, help='selected random seed')
@click.option('--joint_steps', type=int, default=500, help='num steps for joint optimization')
@click.option('--trunc', 'truncation_psi', type=float, help='Truncation psi', default=0.6, show_default=True)
@click.option('--outdir', help='Where to save the output images', default="outputs/insetgan/", type=str, required=True, metavar='DIR')
@click.option('--video', help="set to 1 if want to save video", type=int, default=0)
def main(
        ctx: click.Context,
        face_network: str,
        body_network: str,
        face_seed: int,
        body_seed: int,
        joint_steps: int,
        truncation_psi: float,
        outdir: str,
        video: int):
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    insgan = InsetGAN(body_network, face_network)
    os.makedirs(outdir, exist_ok=True)
    face_z = np.random.RandomState(face_seed).randn(1, 512).astype(np.float32)
    face_mean = insgan.face_generator.mean_latent(3000)
    face_w = insgan.face_generator.get_latent(
        torch.from_numpy(face_z).to(device))  # [N, L, C]
    face_w = truncation_psi * face_w + (1-truncation_psi) * face_mean
    face_img, _ = insgan.face_generator([face_w], input_is_latent=True)

    body_z = np.random.RandomState(body_seed).randn(1, 512).astype(np.float32)
    body_mean = insgan.body_generator.mean_latent(3000)
    body_w = insgan.body_generator.get_latent(
        torch.from_numpy(body_z).to(device))  # [N, L, C]
    body_w = truncation_psi * body_w + (1-truncation_psi) * body_mean
    body_img, _ = insgan.body_generator([body_w], input_is_latent=True)

    _, body_crop, _ = insgan.detect_face_dlib(body_img)
    face_img = F.interpolate(face_img, size=(
        body_crop[3]-body_crop[1], body_crop[2]-body_crop[0]), mode='area')
    cp_body = body_img.clone()
    cp_body[:, :, body_crop[1]:body_crop[3],
            body_crop[0]:body_crop[2]] = face_img

    optim_face_w, optim_body_w, crop = insgan.dual_optimizer(
        face_w,
        body_w,
        joint_steps=joint_steps,
        seed=f'{face_seed:04d}_{body_seed:04d}',
        output_path=outdir,
        video=video
    )

    if video:
        ffmpeg_cmd = f"ffmpeg -hide_banner -loglevel error -i ./{outdir}/{face_seed:04d}_{body_seed:04d}/%04d.jpg -c:v libx264 -vf fps=30 -pix_fmt yuv420p ./{outdir}/{face_seed:04d}_{body_seed:04d}.mp4"
        os.system(ffmpeg_cmd)
    new_face_img, _ = insgan.face_generator(
        [optim_face_w], input_is_latent=True)
    new_shape = crop[3] - crop[1], crop[2] - crop[0]
    new_face_img_crop = F.interpolate(
        new_face_img, size=new_shape, mode='area')
    seamless_body, _ = insgan.body_generator(
        [optim_body_w], input_is_latent=True)
    seamless_body[:, :, crop[1]:crop[3], crop[0]:crop[2]] = new_face_img_crop
    temp = torch.cat([cp_body, seamless_body], dim=3)
    visual(temp, f"{outdir}/{face_seed:04d}_{body_seed:04d}.png")


if __name__ == "__main__":
    main()