ORIGINAL_README.md

LatentSync: Audio Conditioned Latent Diffusion Models for Lip Sync

📖 Abstract

We present LatentSync, an end-to-end lip sync framework based on audio conditioned latent diffusion models without any intermediate motion representation, diverging from previous diffusion-based lip sync methods based on pixel space diffusion or two-stage generation. Our framework can leverage powerful capabilities of Stable Diffusion to directly model complex audio-visual correlations. Additionally, we found that diffusion-based lip sync methods exhibit inferior temporal consistency due to inconsistency in diffusion process across different frames. We propose Temporal REPresentation Alignment (TREPA) to enhance temporal consistency while preserving lip-sync accuracy. TREPA uses temporal representations extracted by large-scale self-supervised video models to align generated frames with ground truth frames.

🏗️ Framework

LatentSync uses Whisper to convert melspectrogram into audio embeddings, which are then integrated into U-Net via cross-attention layers. The reference and masked frames are channel-wise concatenated with noised latents as input of U-Net. In training process, we use one-step method to get estimated clean latents from predicted noises, which are then decoded to obtain the estimated clean frames. The TREPA, LPIPS and SyncNet loss are added in the pixel space.

🎬 Demo

Original video	Lip-synced video
demo2_input.mp4	demo2_output_v1.6.mp4
demo3_input.mp4	demo3_output_v1.6.mp4
demo4_input.mp4	demo4_output_v1.6.mp4
demo5_input.mp4	demo5_output_v1.6.mp4
demo4_video.mp4	demo4_output.mp4

(Photorealistic videos are filmed by contracted models, and anime videos are from VASA-1 and EMO)

📑 Open-source Plan

• Inference code and checkpoints
• Data processing pipeline
• Training code

🔧 Setting up the Environment

Install the required packages and download the checkpoints via:

source setup_env.sh

If the download is successful, the checkpoints should appear as follows:

./checkpoints/
|-- latentsync_unet.pt
|-- latentsync_syncnet.pt
|-- whisper
|   `-- tiny.pt
|-- auxiliary
|   |-- 2DFAN4-cd938726ad.zip
|   |-- i3d_torchscript.pt
|   |-- koniq_pretrained.pkl
|   |-- s3fd-619a316812.pth
|   |-- sfd_face.pth
|   |-- syncnet_v2.model
|   |-- vgg16-397923af.pth
|   `-- vit_g_hybrid_pt_1200e_ssv2_ft.pth

These already include all the checkpoints required for latentsync training and inference. If you just want to try inference, you only need to download latentsync_unet.pt and tiny.pt from our HuggingFace repo

🚀 Inference

Run the script for inference, which requires about 6.5 GB GPU memory.

./inference.sh

You can try adjusting the following inference parameters to achieve better results:

• inference_steps [20-50]: A higher value improves visual quality but slows down the generation speed.
• guidance_scale [1.0-3.0]: A higher value improves lip-sync accuracy but may cause the video distortion or jitter.

🔄 Data Processing Pipeline

The complete data processing pipeline includes the following steps:

1. Remove the broken video files.
2. Resample the video FPS to 25, and resample the audio to 16000 Hz.
3. Scene detect via PySceneDetect.
4. Split each video into 5-10 second segments.
5. Remove videos where the face is smaller than 256 $\times$ 256, as well as videos with more than one face.
6. Affine transform the faces according to the landmarks detected by face-alignment, then resize to 256 $\times$ 256.
7. Remove videos with sync confidence score lower than 3, and adjust the audio-visual offset to 0.
8. Calculate hyperIQA score, and remove videos with scores lower than 40.

Run the script to execute the data processing pipeline:

./data_processing_pipeline.sh

You should change the parameter input_dir in the script to specify the data directory to be processed. The processed data will be saved in the same directory. Each step will generate a new directory to prevent the need to redo the entire pipeline in case the process is interrupted by an unexpected error.

🏋️‍♂️ Training U-Net

Before training, you must process the data as described above and download all the checkpoints. We released a pretrained SyncNet with 94% accuracy on VoxCeleb2 dataset for the supervision of U-Net training. Note that this SyncNet is trained on affine transformed videos, so when using or evaluating this SyncNet, you need to perform affine transformation on the video first (the code of affine transformation is included in the data processing pipeline).

If all the preparations are complete, you can train the U-Net with the following script:

./train_unet.sh

You should change the parameters in the U-Net config file to specify the data directory, checkpoint save path, and other training hyperparameters.

🏋️‍♂️ Training SyncNet

In case you want to train SyncNet on your own datasets, you can run the following script. The data processing pipeline for SyncNet is the same as for U-Net.

./train_syncnet.sh

After validations_steps training, the loss charts will be saved in train_output_dir. They contain both the training and validation loss.