One repo, five papers, infinite pixels

What it does

NÜWA is a family of Microsoft Research generative models that create and manipulate images and videos. The original model handles eight visual synthesis tasks—generation and editing alike—while later papers stretch into infinite-resolution images, language-guided inpainting, 3D photography from single stills, and extremely long video generation. This repository acts as a central bulletin board linking to the papers and project pages.

The interesting bit

Instead of a monolithic toolkit, the repo is a curated anthology of five distinct conference papers—ECCV, NeurIPS, CVPR, IJCAI, and ACL—bound together by the NUWA brand and a shared obsession with visual world creation. Think of it as a research mixtape with GIFs.

Key highlights

• Aggregates five peer-reviewed papers under one organizational umbrella
• NÜWA covers eight distinct visual synthesis tasks in a unified multimodal framework
• NUWA-Infinity generates arbitrarily-sized high-resolution images or long-duration videos
• NUWA-XL uses a “diffusion over diffusion” architecture for extremely long video generation
• NUWA-LIP performs language-guided image inpainting with a defect-free VQGAN

Caveats

• The README lists papers and animated demos, but shows no code, file tree, or implementation details
• It is unclear from the sources which (if any) of the five papers share a common codebase

Verdict

Academic researchers tracing the lineage of generative video and image models should bookmark this; practitioners hunting for a unified synthesis SDK should look elsewhere.