LeapTalk

LeapTalk is a research framework for audio-driven talking-head generation, described in a 2026 paper by Rongxiang Zhang and Songhua Liu. It takes a reference image and speech audio and generates video in a streaming, chunk-by-chunk fashion, with the aim of removing the usual trade-off between latency and quality.

The method distills a pre-trained flow-matching diffusion teacher into a single-step student. The student uses a Brownian-bridge, data-to-data formulation anchored to a persistent reference image, which is intended to limit identity drift and error accumulation in long videos. Distillation uses a heterogeneous DMD setup with an SNR-aligned time transformation, plus audio-driven classifier-free guidance to keep lip sync under extreme step reduction.

The project page reports up to 200 FPS with one inference step per chunk on a single H200 GPU. A YouTube reviewer described the output motion as rigid but very fast, and said it was released openly. No code, weights or license were visible in the provided material, so release status and terms should be checked against the project page and the arXiv paper.