---
title: Lip Forcing
emoji: 🗣️
colorFrom: indigo
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
python_version: "3.10"
short_description: Few-step autoregressive diffusion for real-time lip sync
startup_duration_timeout: 60m
---

# Lip Forcing

Gradio demo for **Lip Forcing: Few-Step Autoregressive Diffusion for Real-time
Lip Synchronization** (KAIST AI · AIPARK), running the released self-contained
**14B student** checkpoint.

Given a talking-head reference video and a driving audio clip, the model detects
and aligns the face to 512×512, then regenerates the mouth region to match the
audio using a causal 2-step autoregressive diffusion student, and composites the
result back into the original frames.

- Paper: https://arxiv.org/abs/2606.11180
- Project page: https://cvlab-kaist.github.io/LipForcing/
- Code: https://github.com/cvlab-kaist/LipForcing
- Weights: https://huggingface.co/JinhyukJang/lipforcing

This demo reproduces the official streaming inference pipeline
(`scripts/inference/inference_streaming.py`) 1:1 on ZeroGPU. Driving audio is
capped to the first few seconds per request to keep a single call within the GPU
time budget.