| from infer_utils import load_poses_whole_video, resizecrop |
| import os |
| import cv2 |
| import json |
| import argparse |
| import numpy as np |
| import glob |
| from PIL import Image |
| from functools import partial |
|
|
| import decord |
| from unit_test.io_utils import save_video |
| from concurrent.futures import ThreadPoolExecutor |
| from tqdm import tqdm |
|
|
| |
| cache_base_dir="../input_cache" |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| task_list = [ |
| ("../examples/img.png","../examples/vid.mp4",1,True,"ref","","",True,False,False), |
| ("../examples/joker2_resize.png","../examples/douyinvid5_v2.mp4",1,False,"pose","","",True,True,False), |
| ("../examples/musk.jpg","../examples/vid2.mp4",1,True,"ref","","",True,False,False), |
| ("../examples/maodie.png","../examples/vid2.mp4",1,True,"ref","","",True,True,False), |
| ] |
|
|
| def process_one(reference_path, video_path, frame_interval, do_align, alignmode, h=None, w=None, face_change=True, head_change=True, without_face=False): |
|
|
| if not h or not w: |
| ref_img_tmp = Image.open(reference_path).convert("RGB") |
| w, h = ref_img_tmp.size |
| else: |
| h = int(h) |
| w = int(w) |
|
|
|
|
| max_short = 768 |
| if min(h, w) > max_short: |
| if h < w: |
| scale = max_short / h |
| h, w = max_short, int(w * scale) |
| else: |
| scale = max_short / w |
| w, h = max_short, int(h * scale) |
| new_h = (h // 16) * 16 |
| new_w = (w // 16) * 16 |
| transform = partial(resizecrop, th=new_h, tw=new_w) |
| anchor_idx = 0 |
|
|
| pose_tensor, image_input, pose_input, mask_input = load_poses_whole_video( |
| video_path=video_path, |
| reference=reference_path, |
| frame_interval=frame_interval, |
| do_align=do_align, |
| transform=transform, |
| alignmode=alignmode, |
| anchor_idx=anchor_idx, |
| face_change=face_change, |
| head_change=head_change, |
| without_face=without_face, |
| ) |
|
|
|
|
| if os.path.isdir(video_path): |
| fps = 30 |
| print(f"⚠ {video_path} isdir, use default fps={fps}") |
| else: |
| vr = decord.VideoReader(video_path) |
| fps = vr.get_avg_fps() |
|
|
| output_fps = fps / frame_interval |
|
|
| |
| video_tensor = pose_tensor / 255 * 2 - 1 |
|
|
|
|
| vidname = os.path.splitext(os.path.basename(video_path))[0] |
| refname = os.path.splitext(os.path.basename(reference_path))[0] |
| cache_dir = os.path.join(cache_base_dir, f"ref_{refname}_driven_{vidname}_align_{alignmode}_h_{new_h}_w_{new_w}_fps_{output_fps}") |
| os.makedirs(cache_dir, exist_ok=True) |
|
|
|
|
| save_video(os.path.join(cache_dir, "pose.mp4"), video_tensor, fps=output_fps) |
| image_input.save(os.path.join(cache_dir, "image_input.png")) |
| pose_input.save(os.path.join(cache_dir, "pose_input.png")) |
| mask_input.save(os.path.join(cache_dir, "mask_input.png")) |
|
|
| if __name__ == "__main__": |
| with ThreadPoolExecutor(max_workers=1) as executor: |
| futures = [] |
| for ref_path, video_path , interval, align, mode, h, w, face_change, head_change, without_face in task_list: |
| futures.append(executor.submit(process_one, ref_path, video_path, interval, align, mode, h, w, face_change, head_change, without_face)) |
| |
| for f in tqdm(futures): |
| f.result() |
|
|
|
|
|
|
