Kangsheng Duan, Ziyang Xu, Wenyu Liu, Xiaohu Ruan, Xiaoxin Chen, Xinggang Wang
Moebius is a lightweight inpainting framework with 0.22B parameters that matches the performance of 10B-level models (FLUX.1-Fill-Dev) while achieving over 15x inference speedup.
10B-level industrial foundation models excel at image inpainting but are computationally prohibitive for deployment. Extreme compression for task-specific specialists leads to a severe representation bottleneck.
The diffusion backbone is systematically reconstructed using the Local-λ Mix Interaction (LλMI) block, which summarizes spatial contexts and global semantic priors into fixed-size linear matrices, preserving complex latent interactions while drastically reducing parameters. An adaptive multi-granularity distillation strategy operating in latent space dynamically balances gradient-based losses for high-fidelity alignment.
Experiments on natural and portrait benchmarks show Moebius (0.22B parameters) rivals or surpasses FLUX.1-Fill-Dev (11.9B parameters) in generation quality, with over 15x faster total inference time.