Lip Forcing

What Lip Forcing is

Lip Forcing is a free, self-hosted method for generating lip-synchronized video from a reference video and speech audio. Its inference pipeline detects and aligns faces, processes 512×512 crops, and composites the generated result into the video. Causal student models generate each chunk in two denoising steps without classifier-free guidance during inference. Streaming inference handles chunks as they arrive; the project reports sub-millisecond time to first frame and constant GPU memory use with clip length under its stated setup. Reported performance includes 31 FPS for the 1.3B student on a single H100 GPU and a 14B student running 39.8 times faster than its teacher at comparable reference fidelity. The 14B checkpoint is available as a merged, self-contained file; the 1.3B weights are listed as coming soon. Inference also requires external components for video, audio and text processing, decoding, and mouth masking. The project reports testing on Ubuntu 24.04 with an NVIDIA H200 and says inference runs on a single GPU. Peak VRAM for the 14B model is reported at about 37 GB with precomputed text embeddings or about 50 GB with runtime text encoding. The code is released under Apache License 2.0, and the model is not deployed by a Hugging Face Inference Provider.