Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance (hustvl.github.io)

Researchers have unveiled Moebius, a groundbreaking lightweight image inpainting framework boasting a mere 0.2 billion parameters while achieving performance levels comparable to models with 10 billion parameters. Utilizing the Latent Diffusion Model (LDM) integrated with Latent Categories Guidance (LCG), Moebius significantly enhances the efficiency of image processing tasks through a restructured denoising U-Net architecture. This restructuring employs innovative LλM I blocks, which are designed to maintain high levels of performance without the typical increase in resource consumption. This development is notable for the AI/ML community as it demonstrates that significant computational and memory savings do not necessarily compromise quality. Moebius employs an adaptive multi-granularity distillation strategy during training, which aligns the lightweight specialist network with a high-capacity teacher model. This approach adeptly mitigates the performance drops associated with reducing model size, opening new avenues for deploying sophisticated image inpainting techniques in resource-constrained environments like mobile devices or real-time applications.