scripts/inference.py

# Copyright (c) 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
from omegaconf import OmegaConf
import torch
from diffusers import AutoencoderKL, DDIMScheduler
from latentsync.models.unet import UNet3DConditionModel
from latentsync.pipelines.lipsync_pipeline import LipsyncPipeline
#from diffusers.utils.import_utils import is_xformers_available
from accelerate.utils import set_seed
from latentsync.whisper.audio2feature import Audio2Feature

def main(video_path, audio_path, video_out_path="./outputs/outvideo.mp4",unet_ckpt_path="./checkpoints/latentsync/latentsync_unet.pt",vae_path="./checkpoints/sd-vae-ft-mse",unet_config_path="configs/unet/second_stage.yaml", guidance_scale=1.0, seed=1247):
    print(f"Input video path: {video_path}")
    print(f"Input audio path: {audio_path}")
    print(f"Loaded unet checkpoint path: {unet_ckpt_path}")
    config = OmegaConf.load(unet_config_path)
    scheduler = DDIMScheduler.from_pretrained("configs")

    if config.model.cross_attention_dim == 768:
        whisper_model_path = "checkpoints/whisper/small.pt"
    elif config.model.cross_attention_dim == 384:
        whisper_model_path = "checkpoints/whisper/tiny.pt"
    else:
        raise NotImplementedError("cross_attention_dim must be 768 or 384")

    audio_encoder = Audio2Feature(model_path=whisper_model_path, device="cuda", num_frames=config.data.num_frames)

    vae = AutoencoderKL.from_pretrained(vae_path, torch_dtype=torch.float16)
    vae.config.scaling_factor = 0.18215
    vae.config.shift_factor = 0

    unet, _ = UNet3DConditionModel.from_pretrained(
        OmegaConf.to_container(config.model),
        unet_ckpt_path,  # load checkpoint
        device="cpu",
    )

    unet = unet.to(dtype=torch.float16)
    
    pipeline = LipsyncPipeline(
        vae=vae,
        audio_encoder=audio_encoder,
        unet=unet,
        scheduler=scheduler,
    ).to("cuda")

    if seed != -1:
        set_seed(seed)
    else:
        torch.seed()

    print(f"Initial seed: {torch.initial_seed()}")

    pipeline(
        video_path=video_path,
        audio_path=audio_path,
        video_out_path=video_out_path,
        video_mask_path=video_out_path.replace(".mp4", "_mask.mp4"),
        num_frames=config.data.num_frames,
        num_inference_steps=config.run.inference_steps,
        guidance_scale=guidance_scale,
        weight_dtype=torch.float16,
        width=config.data.resolution,
        height=config.data.resolution,
    )


if __name__ == "__main__":
    main("./assets/demo2_video.mp4","./assets/demo1_audio.wav")