One video diffusion model to handle 30 different tasks
UniVidX is a unified video diffusion framework from HKUST, Stanford, and Tsinghua, just published at SIGGRAPH 2026. It handles 30 different tasks — matting, normal estimation, relighting, inpainting — from a single pretrained backbone using stochastic condition masking and decoupled gated LoRAs. And it does this training on fewer than 1,000 videos per domain.