Moebius: a 0.2B inpainting model that keeps up with 12B ones

A team from Huazhong University of Science and Technology and VIVO AI Lab has released Moebius, an image inpainting model with 226 million parameters that performs on par with systems roughly fifty times its size. Inpainting is the task of filling a masked region of an image so the result looks seamless, and the strong open option until now has been FLUX.1-Fill-Dev at 11.9 billion parameters. Moebius is less than 2 percent of that size. Across six benchmarks covering natural scenes (Places2) and faces (CelebA-HQ, FFHQ), the authors report quality that matches or beats both FLUX.1-Fill-Dev and SD3.5 Large-Inpainting.

The speed gap is the part that matters in practice. Moebius runs a step in about 26 milliseconds on a single GPU and is more than 15 times faster end to end than the 10B-class models. Two ideas carry most of the weight. The first replaces standard attention with what the authors call Local-λ Mix Interaction blocks, which compress spatial and global context into fixed-size linear matrices and so avoid the quadratic cost of attention. The second is a distillation recipe that transfers knowledge from a large teacher entirely in latent space, with a loss that adapts its weighting across different levels of supervision.

The claim the authors make is narrow and worth keeping: scale is not the only path to quality when the task is well defined. A specialized model trained carefully can match a general one many times larger. The project page has the benchmark tables and qualitative comparisons.

Why it matters

If you ship inpainting or photo editing on phones or cheap GPUs, a 226M model at FLUX-level quality changes what fits on device. It is also a concrete case for the wider bet that task-specific small models can replace large general ones where the job is narrow.