Top AI Product

Most state-of-the-art image inpainting runs on 10-billion-parameter models like FLUX.1-Fill-Dev. Moebius, open-sourced June 18 by researchers at Huazhong University of Science and Technology and vivo AI Lab, does the same job with 0.22B parameters — under 2% of the size — and in some cases does it better.

## What Moebius does

Inpainting means filling a masked region of an image so the result looks seamless: removing an object, reconstructing a face, repairing a texture. Moebius matches and sometimes beats FLUX.1-Fill-Dev and SD3.5 Large-Inpainting on complex textures and facial plausibility, while running at 26ms per step on a single GPU. That works out to a >15x total runtime speedup over the 10B-class models it competes with.

## How it gets there

The trick is a redesigned diffusion backbone built around a Local-λ Mix Interaction block, which compresses spatial context and global semantic priors into fixed-size linear matrices instead of dragging them through billions of parameters. Code, weights, and the ECCV 2026 paper are all public. The point isn’t just speed — at 226M parameters it runs where a 12B model simply can’t.