- Home
- AI Video Lip Sync Tools
- Lip Forcing
On the recordRead 8 Oct 202624 facts
Lip Forcing
On the record
- Automatic processing
- The inference pipeline automatically detects and aligns faces to 512×512 crops and pastes the generated result back into the video.github.com · read 4 Oct 2026
- Availability
- The Hugging Face model card says the model is not deployed by an Inference Provider.huggingface.co · read 8 Oct 2026
- Dependencies
- Inference requires external components including the Wan VAE, wav2vec audio encoder, text encoder or precomputed embeddings, decoder, and mouth mask.github.com · read 8 Oct 2026
- External components
- Inference also requires external components including the Wan VAE, wav2vec audio encoder, UMT5-XXL text encoder or precomputed embeddings, TAEW decoder, and mouth mask.github.com · read 4 Oct 2026
- Generation
- Its causal student models generate each chunk in two denoising steps without classifier-free guidance at inference.github.com · read 4 Oct 2026
- Hardware
- The repository reports testing on Ubuntu 24.04 with an NVIDIA H200 and says inference runs on a single GPU.github.com · read 4 Oct 2026
- Hosting
- The checkpoint page states that the model is not deployed by any Hugging Face Inference Provider.huggingface.co · read 4 Oct 2026
- Inference setup
- The documented inference command takes a reference video and speech audio and writes an output video.github.com · read 4 Oct 2026
- Input handling
- Inference accepts a reference video and speech audio, and automatically detects, aligns, and composites the face.github.com · read 8 Oct 2026
- Integrations
- The checkpoint is hosted on Hugging Face, and the README also identifies external model components from Wan, wav2vec2, TAEW, LatentSync, and SyncNet sources.github.com · read 8 Oct 2026
- License
- The repository states that Lip Forcing is released under the Apache License 2.0.github.com · read 8 Oct 2026
- Memory requirement
- The 14B model uses about 37 GB peak GPU memory with precomputed text embeddings, or about 50 GB when encoding text at runtime.github.com · read 8 Oct 2026
- Model sizes
- The release includes a 14B student checkpoint; the 1.3B student weights are listed as coming soon.github.com · read 8 Oct 2026
- Performance
- The project reports that the 1.3B student reaches 31 FPS and the 14B student runs 39.8 times faster than its teacher at comparable reference fidelity.cvlab-kaist.github.io · read 4 Oct 2026
- Platform
- The documented installation was tested on Ubuntu 24.04 with Python 3.12, PyTorch 2.10.0, CUDA 12.8, and an NVIDIA H200; inference runs on a single GPU.github.com · read 8 Oct 2026
- Project affiliation
- The project lists KAIST AI and AIPARK affiliations for its authors.github.com · read 4 Oct 2026
- Purpose
- Lip Forcing is an autoregressive diffusion method for video-to-video lip synchronization that animates a reference video from audio.github.com · read 8 Oct 2026
- Real-time performance
- The project reports 31 FPS for the 1.3B student on a single H100 GPU.cvlab-kaist.github.io · read 8 Oct 2026
