LipForcing

Lip Forcing is an AI research project from KAIST AI and AIPARK focused on real-time lip synchronization. Lip synchronization means adjusting the lip movements shown in a video so they match a different audio track, which is useful for dubbing videos into other languages, generating synthetic speech for virtual presenters, or creating talking-head video from audio alone. The project's full name is "Lip Forcing: Few-Step Autoregressive Diffusion for Real-time Lip Synchronization."

The technical approach described in the title involves two ideas: autoregressive diffusion and few-step generation. Diffusion-based models generate images or video by starting with noise and gradually refining it into something realistic, frame by frame or chunk by chunk. "Autoregressive" means each output feeds into the next, so the model generates video in sequence rather than all at once. "Few-step" means the method reduces the number of refinement passes needed compared to standard diffusion approaches, which is what makes real-time speed achievable. Standard diffusion can require dozens or hundreds of steps per frame, which is far too slow for live use.

The repository is essentially a placeholder at this stage. The README consists of the paper title, the list of co-authors from KAIST and AIPARK, and the notice that code is coming soon. There is a link to a project page but no code, model weights, or instructions have been published in the repository yet.

The research is authored by a team of nine people, with two listed as equal contributors and one designated corresponding author. The paper itself is listed but not yet linked.