Browse Teach Curate

Shifted Window Attention Diffusion Transformer for Video Restoration

A diffusion-transformer design for restoring real-world video of arbitrary length and resolution, in which full self-attention is replaced by large non-overlapping shifted-window attention over a compressed latent. Variable-sized boundary windows remove the requirement that input dimensions be multiples of the window size, letting the model restore long, high-resolution video efficiently while avoiding the slow overlapping patch-based sampling of earlier diffusion methods.

KITT Text

SeedVR: Seeding Infinity in Diffusion Transformer Towards Generic Video Restoration Jianyi Wang1,2∗

SeedVR is a diffusion transformer (DiT) for generic real-world video restoration, the task of reconstructing high-quality video from low-quality input suffering unknown, in-the-wild degradations, at …