Browse Teach Curate

Unified Generalized Video Face Restoration with Stable Video Diffusion

A single framework that jointly performs video blind face restoration, inpainting, and colorization by conditioning a Stable Video Diffusion backbone on a learnable task embedding. A Unified Latent Regularization encourages a shared feature representation across the subtasks, and facial-prior learning together with self-referred refinement improve restoration quality and temporal stability across video frames.

Rosie Text

SVFR: A Unified Framework for Generalized Video Face Restoration Zhiyao Wang1,2, Xu Chen2,*

Face restoration reconstructs high-quality faces from degraded inputs; the video case is harder than the image case because of temporal consistency, motion artifacts, and scarce high-quality video da…