# Moebius — 0.22B image inpainting matches FLUX.1-Fill-Dev's 11.9B

> A 226M-parameter inpainting model that keeps up with 11.9B systems and runs 15x faster.

- Org: HUST + VIVO AI Lab
- Date: 2026-06-18
- Category: paper, algorithm, model
- Source: https://arxiv.org/abs/2606.19195
- Canonical page: https://ai-tldr.dev/releases/hustvl-moebius-jun18/

## In short

Moebius is a 0.22B image inpainting model from HUST and VIVO AI Lab that matches the 11.9B FLUX.1-Fill-Dev across six benchmarks while running over 15x faster, with code and weights now on GitHub.

## What it is

Moebius is a lightweight image inpainting framework from HUST and VIVO AI Lab. The 0.22B model is under 2% the size of FLUX.1-Fill-Dev (11.9B) and matches or beats it on six benchmarks across Places2, CelebA-HQ, and FFHQ. Moebius's training and inference code plus model weights shipped on June 18, 2026.

## How it works

Moebius introduces a Local-lambda Mix Interaction block to fight the representation bottleneck of extreme compression, paired with adaptive multi-granularity distillation that transfers knowledge from a 10B-class teacher. The architecture targets local context at multiple scales, which lets a tiny network preserve detail in masked regions. Per-step inference runs in 26.01 ms.

## Why it matters

Moebius shows that high-quality inpainting does not need a billion-parameter generator. Teams that could not afford to deploy FLUX.1-Fill-Dev now have an Apache-2.0 model 50x smaller that runs on commodity GPUs. The 15x speedup also opens up interactive editing flows where each user click triggers a new inpaint.

## Who it's for

ML researchers building image editing tools, product teams adding inpainting to apps, students studying model distillation.

## Try it

git clone https://github.com/hustvl/Moebius

## Links

- [arXiv Paper](https://arxiv.org/abs/2606.19195)
- [Project Page](https://hustvl.github.io/Moebius/)
- [GitHub Repo](https://github.com/hustvl/Moebius)
- [HF Paper Page](https://huggingface.co/papers/2606.19195)