Paper 2606.19195
Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance
- Published
- Jun 2026
- Research lab
- Independent
- Citations
- 0
- GitHub
- 508 stars
01 In brief
Summary
Moebius is a highly efficient lightweight image inpainting framework that rivals the generation quality of 10B-level industrial models like FLUX.1-Fill-Dev while using only 0.22B parameters (less than 2% of FLUX's 11.9B) and delivering over 15× faster total inference time.
To overcome the representation bottleneck from extreme structural compression, the authors introduce the Local-λ Mix Interaction (LλMI) block, which uses Local-λ and Interactive-λ modules to summarize spatial contexts and global semantic priors into fixed-size linear matrices, enabling efficient self- and cross-attention equivalents.
They also propose an adaptive multi-granularity distillation strategy that operates strictly in latent space, dynamically balancing gradient-based losses (coarse-grained, fine-grained, task, and perceptual) to align the student with a high-capacity teacher (PixelHacker).
Experiments on Places2, CelebA-HQ, and FFHQ show Moebius achieves competitive or superior FID and LPIPS scores compared to 10B-level generalists, with a latency of 26.01 ms/step and 0.154 TFLOPs.
A user study confirms Moebius matches its teacher's preference and outperforms industrial models, especially in portrait scenes.
The framework also demonstrates strong out-of-distribution generalization and real-world object removal capability, setting a new efficiency standard for high-fidelity inpainting.
02 From the paper
Abstract
While 10B-level industrial foundation models have pushed the boundaries of image inpainting, their prohibitive computational costs severely hinder practical deployment. Constructing a highly optimized task-specific specialist offers a promising solution; however, extreme structural compression inevitably triggers a severe representation bottleneck. To conquer this, we propose Moebius, a highly efficient lightweight inpainting framework. We systematically reconstruct the diffusion backbone by introducing the Local-$λ$ Mix Interaction ($LλMI$) block. Comprising Local-$λ$ and Interactive-$λ$ modules, it elegantly summarizes spatial contexts and global semantic priors into fixed-size linear matrices, preserving complex latent interactions while drastically shedding parameters. Furthermore, to unlock the full representational capacity of this highly compact architecture, we synergistically pair it with an adaptive multi-granularity distillation strategy. Operating strictly within the latent space to avoid expensive pixel-space decoding, this strategy dynamically balances multiple gradient-based losses to achieve high-fidelity alignment. Extensive experiments across natural and portrait benchmarks demonstrate that this optimal synergy enables Moebius to rival or even surpass the generation quality of the 10B-level industrial generalist FLUX.1-Fill-Dev. Remarkably, Moebius achieves this using less than 2\% of the parameters (0.22B vs. 11.9B) while delivering a $>15\times$ acceleration in total inference time, setting a new efficiency standard for high-fidelity inpainting. Project page at https://hustvl.github.io/Moebius.