Moebius: 0.2B image inpainting model with 10B-level performance
5 hours ago
- #Image Inpainting
- #Lightweight Architecture
- #Diffusion Models
- Moebius uses the Latent Diffusion Model (LDM) framework with Latent Categories Guidance (LCG).
- The denoising U-Net is restructured using LλM I blocks for extreme architectural efficiency.
- An adaptive multi-granularity distillation strategy aligns the lightweight specialist with a high-capacity teacher to mitigate capacity drop from compression.