LeapTalk: Breaking the Latency-Quality Trade-off in Talking Head Generation
LeapTalk claims one-step talking-head generation at up to 200 FPS while holding identity stable over long videos.
The paper targets the gap between slow diffusion methods and faster autoregressive systems that drift over time. Its main device is a single-step bridge distillation scheme using a Brownian bridge anchored to a persistent reference. The authors also add audio-driven classifier-free guidance to preserve lip sync under the heavy step reduction. They report better efficiency and stability than existing approaches in experiments. HF Daily Papers' note
The paper targets the gap between slow diffusion methods and faster autoregressive systems that drift over time. Its main device is a single-step bridge distillation scheme using a Brownian bridge anchored to a persistent reference. The authors also add audio-driven classifier-free guidance to preserve lip sync under the heavy step reduction. They report better efficiency and stability than existing approaches in experiments. HF Daily Papers' note
score 5