src/audio2vid.py

import os
import ffmpeg
from datetime import datetime
from pathlib import Path
import numpy as np
import cv2
import torch
import spaces
from scipy.spatial.transform import Rotation as R
from scipy.interpolate import interp1d

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 save_videos_grid

from src.audio_models.model import Audio2MeshModel
from src.utils.audio_util import prepare_audio_feature
from src.utils.mp_utils  import LMKExtractor
from src.utils.draw_util import FaceMeshVisualizer
from src.utils.pose_util import project_points


def matrix_to_euler_and_translation(matrix):
    rotation_matrix = matrix[:3, :3]
    translation_vector = matrix[:3, 3]
    rotation = R.from_matrix(rotation_matrix)
    euler_angles = rotation.as_euler('xyz', degrees=True)
    return euler_angles, translation_vector


def smooth_pose_seq(pose_seq, window_size=5):
    smoothed_pose_seq = np.zeros_like(pose_seq)

    for i in range(len(pose_seq)):
        start = max(0, i - window_size // 2)
        end = min(len(pose_seq), i + window_size // 2 + 1)
        smoothed_pose_seq[i] = np.mean(pose_seq[start:end], axis=0)

    return smoothed_pose_seq

def get_headpose_temp(input_video):
    lmk_extractor = LMKExtractor()
    cap = cv2.VideoCapture(input_video)

    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = cap.get(cv2.CAP_PROP_FPS)  

    trans_mat_list = []
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break

        result = lmk_extractor(frame)
        trans_mat_list.append(result['trans_mat'].astype(np.float32))
    cap.release()

    trans_mat_arr = np.array(trans_mat_list)

    # 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

    # interpolate to 30 fps
    new_fps = 30
    old_time = np.linspace(0, total_frames / fps, total_frames)
    new_time = np.linspace(0, total_frames / fps, int(total_frames * new_fps / fps))

    pose_arr_interp = np.zeros((len(new_time), 6))
    for i in range(6):
        interp_func = interp1d(old_time, pose_arr[:, i])
        pose_arr_interp[:, i] = interp_func(new_time)

    pose_arr_smooth = smooth_pose_seq(pose_arr_interp)
    
    return pose_arr_smooth

@spaces.GPU
def audio2video(input_audio, ref_img, headpose_video=None, size=512, steps=25, length=150, seed=42):
    fps = 30
    cfg = 3.5

    config = OmegaConf.load('./configs/prompts/animation_audio.yaml')

    if config.weight_dtype == "fp16":
        weight_dtype = torch.float16
    else:
        weight_dtype = torch.float32
        
    audio_infer_config = OmegaConf.load(config.audio_inference_config)
    # prepare model
    a2m_model = Audio2MeshModel(audio_infer_config['a2m_model'])
    a2m_model.load_state_dict(torch.load(audio_infer_config['pretrained_model']['a2m_ckpt']), strict=False)
    a2m_model.cuda().eval()

    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)
    
    sample = prepare_audio_feature(input_audio, wav2vec_model_path=audio_infer_config['a2m_model']['model_path'])
    sample['audio_feature'] = torch.from_numpy(sample['audio_feature']).float().cuda()
    sample['audio_feature'] = sample['audio_feature'].unsqueeze(0)

    # inference
    pred = a2m_model.infer(sample['audio_feature'], sample['seq_len'])
    pred = pred.squeeze().detach().cpu().numpy()
    pred = pred.reshape(pred.shape[0], -1, 3)
    pred = pred + face_result['lmks3d']
    
    if headpose_video is not None:
        pose_seq = get_headpose_temp(headpose_video)
    else:
        pose_seq = np.load(config['pose_temp'])
    mirrored_pose_seq = np.concatenate((pose_seq, pose_seq[-2:0:-1]), axis=0)
    cycled_pose_seq = np.tile(mirrored_pose_seq, (sample['seq_len'] // len(mirrored_pose_seq) + 1, 1))[:sample['seq_len']]

    # project 3D mesh to 2D landmark
    projected_vertices = project_points(pred, face_result['trans_mat'], cycled_pose_seq, [height, width])

    pose_images = []
    for i, verts in enumerate(projected_vertices):
        lmk_img = vis.draw_landmarks((width, height), verts, normed=False)
        pose_images.append(lmk_img)

    pose_list = []
    pose_tensor_list = []

    pose_transform = transforms.Compose(
        [transforms.Resize((height, width)), transforms.ToTensor()]
    )
    args_L = len(pose_images) if length==0 or length > len(pose_images) else length
    for pose_image_np in pose_images[: args_L]:
        pose_image_pil = Image.fromarray(cv2.cvtColor(pose_image_np, cv2.COLOR_BGR2RGB))
        pose_tensor_list.append(pose_transform(pose_image_pil))
        pose_image_np = cv2.resize(pose_image_np,  (width, height))
        pose_list.append(pose_image_np)
    
    pose_list = np.array(pose_list)
    
    video_length = len(pose_tensor_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=fps,
    )
    
    stream = ffmpeg.input(save_path)
    audio = ffmpeg.input(input_audio)
    ffmpeg.output(stream.video, audio.audio, save_path.replace('_noaudio.mp4', '.mp4'), vcodec='copy', acodec='aac').run()
    os.remove(save_path)
    
    return save_path.replace('_noaudio.mp4', '.mp4')