LeapTalk generates real-time talking-head videos in one step at up to 200 FPS
Shanghai-Jiao-Tong-University-SAI · hf · 2026-08-04
LeapTalk breaks the real-time talking-head trade-off
The paper proposes LeapTalk, a talking-head generation framework that aims to keep quality stable while generating in real time with a single forward step.
- It reframes the task as a data-to-data transport problem using a Brownian bridge anchored by a persistent reference, which helps reduce identity drift.
- To transfer knowledge from a diffusion teacher to the bridge student, it uses heterogeneous distillation with an SNR-aligned time transform.
- An audio-driven classifier-free guidance mechanism is added to preserve lip sync even under extreme step reduction.
- The authors report high-fidelity, temporally consistent output at up to 200 FPS and say it outperforms prior methods on both efficiency and stability.
More from Multimodal
- MiniMax H3 Open Weights Released with ComfyUI Integration — petewoodbridge · 2026-08-04
- MiniMax Exec Reviews Hailuo AI's Growth, Announces Push for Open Source — VoidAsuka · 2026-08-04
- Creating Poster Animations with Hailuo AI: Prompts and Results — LudovicCreator · 2026-08-04
- Qwen3.8-Max Tested: Generates Photorealistic Bugatti Engine in Three.js — cedric_chee · 2026-08-04
- RTX 3060 12GB Test: Generates 10s Portrait Video Locally in 17 Minutes — merica420_69 · 2026-08-04
- No More Messy Wires: Visual Debugger Tool for ComfyUI Released — niknah · 2026-08-04