Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance

2026-06-17

AI Summary

This work addresses the limitations of existing lightweight image inpainting models, which suffer from insufficient representational capacity, while large-scale models incur prohibitive computational costs and deployment challenges. To bridge this gap, the authors propose Moebius, an efficient and lightweight framework featuring a novel Local-λ Mix Interaction module that compresses spatial context and global semantic priors into fixed linear matrices. The approach further integrates latent-space adaptive multi-granularity knowledge distillation with a parameter-efficient diffusion architecture. With only 0.22 billion parameters—less than 2% of FLUX.1-Fill-Dev—the method achieves over 15× faster inference while delivering inpainting quality comparable to or surpassing that of industrial-scale models with 10 billion parameters, thereby significantly overcoming the performance bottleneck of lightweight inpainting systems.

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.

Problem

• image inpainting
• lightweight model
• computational efficiency
• representation bottleneck
• model compression

Innovation

• lightweight inpainting
• diffusion backbone compression
• Local-λ Mix Interaction
• latent-space distillation
• parameter-efficient generation