DigiNews

Summary

Moebius presents a 0.22B parameter lightweight image inpainting framework that achieves 10B-level performance through architecture innovations (Local-λ Mix Interaction blocks) and an adaptive multi-granularity distillation strategy. The approach operates primarily in latent space to maintain high fidelity while delivering over 15x faster inference than larger 10B-scale models, with strong results on natural and portrait benchmarks and open-source releases for code and BibTeX details.