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

▤ Summary

This article introduces Moebius, a lightweight image inpainting framework built on latent diffusion. The model uses a restructured denoising U-Net and latent categories guidance to improve efficiency. It also applies adaptive multi-granularity distillation to preserve performance despite extreme compression. The work positions Moebius as a compact specialist model that reaches strong results relative to much larger systems.

▦ Linked Companies