File size: 7,698 Bytes
2e4e201
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import os
import shutil
import ffmpeg
from datetime import datetime
from pathlib import Path
import numpy as np
import cv2
import torch
# import spaces
from diffusers import AutoencoderKL, DDIMScheduler
from einops import repeat
from omegaconf import OmegaConf
from PIL import Image
from torchvision import transforms
from transformers import CLIPVisionModelWithProjection

from src.models.pose_guider import PoseGuider
from src.models.unet_2d_condition import UNet2DConditionModel
from src.models.unet_3d import UNet3DConditionModel
from src.pipelines.pipeline_pose2vid_long import Pose2VideoPipeline
from src.utils.util import get_fps, read_frames, save_videos_grid

from src.utils.mp_utils  import LMKExtractor
from src.utils.draw_util import FaceMeshVisualizer
from src.utils.pose_util import project_points_with_trans,  matrix_to_euler_and_translation
from src.audio2vid import smooth_pose_seq


def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, default='./configs/prompts/animation_facereenac.yaml')
    parser.add_argument("-W", type=int, default=512)
    parser.add_argument("-H", type=int, default=512)
    parser.add_argument("-L", type=int)
    parser.add_argument("--seed", type=int, default=42)
    parser.add_argument("--cfg", type=float, default=3.5)
    parser.add_argument("--steps", type=int, default=25)
    parser.add_argument("--fps", type=int)
    args = parser.parse_args()

    return args

# @spaces.GPU
def video2video(ref_img, source_video, size=512, steps=25, length=150, seed=42):
    cfg = 3.5
    
    config = OmegaConf.load('./configs/prompts/animation_facereenac.yaml')

    if config.weight_dtype == "fp16":
        weight_dtype = torch.float16
    else:
        weight_dtype = torch.float32

    vae = AutoencoderKL.from_pretrained(
        config.pretrained_vae_path,
    ).to("cuda", dtype=weight_dtype)

    reference_unet = UNet2DConditionModel.from_pretrained(
        config.pretrained_base_model_path,
        subfolder="unet",
    ).to(dtype=weight_dtype, device="cuda")

    inference_config_path = config.inference_config
    infer_config = OmegaConf.load(inference_config_path)
    denoising_unet = UNet3DConditionModel.from_pretrained_2d(
        config.pretrained_base_model_path,
        config.motion_module_path,
        subfolder="unet",
        unet_additional_kwargs=infer_config.unet_additional_kwargs,
    ).to(dtype=weight_dtype, device="cuda")

    pose_guider = PoseGuider(noise_latent_channels=320, use_ca=True).to(device="cuda", dtype=weight_dtype) # not use cross attention

    image_enc = CLIPVisionModelWithProjection.from_pretrained(
        config.image_encoder_path
    ).to(dtype=weight_dtype, device="cuda")

    sched_kwargs = OmegaConf.to_container(infer_config.noise_scheduler_kwargs)
    scheduler = DDIMScheduler(**sched_kwargs)

    generator = torch.manual_seed(seed)

    width, height = size, size

    # load pretrained weights
    denoising_unet.load_state_dict(
        torch.load(config.denoising_unet_path, map_location="cpu"),
        strict=False,
    )
    reference_unet.load_state_dict(
        torch.load(config.reference_unet_path, map_location="cpu"),
    )
    pose_guider.load_state_dict(
        torch.load(config.pose_guider_path, map_location="cpu"),
    )

    pipe = Pose2VideoPipeline(
        vae=vae,
        image_encoder=image_enc,
        reference_unet=reference_unet,
        denoising_unet=denoising_unet,
        pose_guider=pose_guider,
        scheduler=scheduler,
    )
    pipe = pipe.to("cuda", dtype=weight_dtype)

    date_str = datetime.now().strftime("%Y%m%d")
    time_str = datetime.now().strftime("%H%M")
    save_dir_name = f"{time_str}--seed_{seed}-{size}x{size}"

    save_dir = Path(f"output/{date_str}/{save_dir_name}")
    save_dir.mkdir(exist_ok=True, parents=True)


    lmk_extractor = LMKExtractor()
    vis = FaceMeshVisualizer(forehead_edge=False)
    


    ref_image_np = cv2.cvtColor(ref_img, cv2.COLOR_RGB2BGR)
    # TODO: 人脸检测+裁剪
    ref_image_np = cv2.resize(ref_image_np, (size, size))
    ref_image_pil = Image.fromarray(cv2.cvtColor(ref_image_np, cv2.COLOR_BGR2RGB))
    
    face_result = lmk_extractor(ref_image_np)
    if face_result is None: 
        return None
    
    lmks = face_result['lmks'].astype(np.float32)
    ref_pose = vis.draw_landmarks((ref_image_np.shape[1], ref_image_np.shape[0]), lmks, normed=True)
    
    

    source_images = read_frames(source_video)
    src_fps = get_fps(source_video)
    pose_transform = transforms.Compose(
        [transforms.Resize((height, width)), transforms.ToTensor()]
    )
    
    step = 1
    if src_fps == 60:
        src_fps = 30
        step = 2
    
    pose_trans_list = []
    verts_list = []
    bs_list = []
    src_tensor_list = []
    args_L = len(source_images) if length==0 or length*step > len(source_images) else length*step
    for src_image_pil in source_images[: args_L: step]:
        src_tensor_list.append(pose_transform(src_image_pil))
        src_img_np = cv2.cvtColor(np.array(src_image_pil), cv2.COLOR_RGB2BGR)
        frame_height, frame_width, _ = src_img_np.shape
        src_img_result = lmk_extractor(src_img_np)
        if src_img_result is None:
            break
        pose_trans_list.append(src_img_result['trans_mat'])
        verts_list.append(src_img_result['lmks3d'])
        bs_list.append(src_img_result['bs'])

    
    # pose_arr = np.array(pose_trans_list)
    trans_mat_arr = np.array(pose_trans_list)
    verts_arr = np.array(verts_list)
    bs_arr = np.array(bs_list)
    min_bs_idx = np.argmin(bs_arr.sum(1))
    
    # compute delta pose
    trans_mat_inv_frame_0 = np.linalg.inv(trans_mat_arr[0])
    pose_arr = np.zeros([trans_mat_arr.shape[0], 6])

    for i in range(pose_arr.shape[0]):
        pose_mat = trans_mat_inv_frame_0 @ trans_mat_arr[i]
        euler_angles, translation_vector = matrix_to_euler_and_translation(pose_mat)
        pose_arr[i, :3] =  euler_angles
        pose_arr[i, 3:6] =  translation_vector

    pose_arr = smooth_pose_seq(pose_arr)

    # face retarget
    verts_arr = verts_arr - verts_arr[min_bs_idx] + face_result['lmks3d']
    # project 3D mesh to 2D landmark
    projected_vertices = project_points_with_trans(verts_arr, pose_arr, [frame_height, frame_width])
    
    pose_list = []
    for i, verts in enumerate(projected_vertices):
        lmk_img = vis.draw_landmarks((frame_width, frame_height), verts, normed=False)
        pose_image_np = cv2.resize(lmk_img,  (width, height))
        pose_list.append(pose_image_np)
    
    pose_list = np.array(pose_list)
    
    video_length = len(pose_list)

    video = pipe(
        ref_image_pil,
        pose_list,
        ref_pose,
        width,
        height,
        video_length,
        steps,
        cfg,
        generator=generator,
    ).videos

    save_path = f"{save_dir}/{size}x{size}_{time_str}_noaudio.mp4"
    save_videos_grid(
        video,
        save_path,
        n_rows=1,
        fps=src_fps,
    )
    
    audio_output = f'{save_dir}/audio_from_video.aac'
    # extract audio
    try:
        ffmpeg.input(source_video).output(audio_output, acodec='copy').run()
        # merge audio and video
        stream = ffmpeg.input(save_path)
        audio = ffmpeg.input(audio_output)
        ffmpeg.output(stream.video, audio.audio, save_path.replace('_noaudio.mp4', '.mp4'), vcodec='copy', acodec='aac').run()
    
        os.remove(save_path)
        os.remove(audio_output)
    except:
        shutil.move(
            save_path,
            save_path.replace('_noaudio.mp4', '.mp4')
        )
    
    return save_path.replace('_noaudio.mp4', '.mp4')