LeapTalk Achieves 200 FPS Talking-Head Video Generation in a Single Inference Step

Researchers from Shanghai Jiao Tong University and Harbin Institute of Technology have introduced LeapTalk, a framework that generates real-time talking-head video in a single forward step — reaching up to 200 FPS on a single H200 GPU. The method sidesteps the traditional latency-quality trade-off that plagues existing approaches: multi-step diffusion models are too slow for streaming, while autoregressive systems accumulate errors and suffer identity drift over long sequences.

LeapTalk's core innovation is a reformulation of generation as reference-anchored data-to-data transport using a Brownian bridge, rather than the conventional noise-to-data reconstruction cycle. A heterogeneous distillation framework with SNR-aligned time transformation transfers knowledge from a pre-trained diffusion teacher to the single-step student model, and an audio-driven classifier-free guidance mechanism preserves lip synchronization even under extreme step reduction. The result is temporally stable, high-fidelity video that scales to arbitrarily long outputs.