Moebius: A 0.2B Lightweight Image Inpainting Framework Achieving 10B-Class Performance

Existing 10B(10 billion)-scale industrial foundation models deliver excellent image Inpainting performance, but their massive computational cost makes real-world service deployment difficult. Moebius, proposed to solve this problem, focuses on overcoming the representational degradation that occurs during structural compression.

Moebius reconstructs the Diffusion backbone by introducing the LλMI(Local-λ Mix Interaction) block. Composed of the Local-λ and Interactive-λ modules, this block summarizes spatial context and global semantic information into fixed-size linear matrices, dramatically reducing the parameter count while preserving complex latent interactions.

It also combines an Adaptive Multi-granularity Distillation strategy to maximize the model's representational capacity. This strategy operates solely within the Latent Space to avoid costly pixel-space decoding, and dynamically balances multiple Gradient-based loss functions to achieve high-precision alignment.

Experimental results show that Moebius matches or even surpasses the 10B-scale FLUX.1-Fill-Dev model in generation quality on natural image and portrait benchmarks. In particular, using only 0.22B parameters, it achieves over 15x inference speed compared to existing models, setting a new efficiency benchmark for high-precision inpainting.