Moebius: 0.22B image-inpainting model matching 10B-class quality

Category: ML research
Status: developing
Sources: project page, arXiv 2606.19195, discussion
Summary: Researchers from Huazhong University of Science and Technology and the VIVO AI Lab describe Moebius, a 226M-parameter image-inpainting model they report matches or exceeds FLUX.1-Fill-Dev (11.9B) and SD3.5 Large-Inpainting across six Places2/CelebA-HQ/FFHQ benchmarks at under 2% of the size. The method uses an LλMI block that condenses spatial context into fixed-size linear matrices to avoid quadratic attention cost, plus multi-granularity distillation from a PixelHacker teacher; the authors report 26ms per step and over 15x runtime speedup.
Why it matters: A sub-billion-parameter model claiming parity with 10B-class inpainting would cut serving cost and latency for an editing workload if the result reproduces.
Follow-up: Confirm the license and released weights, and watch for independent benchmark reproduction.