HighSync: High-Quality Lip Synchronization via Latent Diffusion Models
HighSync is an end-to-end diffusion model for generating photorealistic talking-face videos aligned with audio. It extends Stable Diffusion 1.5 to handle sequences of 12 frames, using reference images for identity and audio embeddings to drive lip movement. The method successfully resolves data leakage issues related to face cropping and muscle coupling, achieving state-of-the-art visual quality and synchronization accuracy.