Moebius: 0.2B image inpainting model with 10B-level performance

304 points · 77 comments on HN · read original →

Points and comments are a snapshot, not live.

A 0.22B parameter image inpainting model matches 10B-level quality with over 15× speedup.

Moebius is a 226M parameter image inpainting framework that rivals the 11.9B parameter FLUX.1-Fill-Dev across six benchmarks. Its efficiency comes from the Local-λ Mix Interaction block, which compresses spatial and semantic information into fixed-size linear matrices, and an adaptive multi-granularity distillation strategy that trains within latent space to avoid expensive pixel decoding. Inference runs at 26 ms per step on a single GPU.

Developed by HUST and VIVO AI Lab, Moebius offers a specialist alternative to bloated generalist models, delivering high-fidelity inpainting for both natural scenes and portraits with less than 2% of the parameters and a >15× total runtime improvement.

What commenters are saying

Commenters praise the model's sample gallery but express frustration over unclear access. The project is available on Hugging Face and has a live demo.

One thread discusses practical inpainting for e-commerce, with comparisons to proprietary models like GPT-image-2 and NB2, as well as locally hostable options like Flux.2 Klein and Boogu-Image. Several note that additive modifications like awnings might not require full inpainting if approached with classical photogrammetry. Another commenter highlights that local purpose-built models are often more practical than cloud-based generalists.