hustvl/Moebius

What it solves

Moebius addresses the "scale-at-all-costs" trend in AI image inpainting, where high-quality results typically require massive foundation models (10B+ parameters) that are computationally expensive and slow. It provides a lightweight alternative that maintains industrial-grade quality while drastically reducing the parameter count and inference time, making high-quality inpainting viable for consumer-grade and edge devices.

How it works

Moebius uses a combination of architectural innovation and knowledge distillation to achieve its efficiency:

• LΙMI Block: A custom architecture that reformulates self- and cross-attention by condensing spatial context and global semantic priors into fixed-size linear matrices, eliminating the quadratic computational overhead usually associated with attention mechanisms.
• Adaptive Multi-Granularity Distillation: The model is trained as a "student" that learns from a larger "teacher" model (PixelHacker). It aligns supervision across multiple levels—from microscopic intermediate features to macroscopic diffusion trajectories—strictly within the latent space to avoid expensive pixel-space decoding.
• Synergistic Balancing: The framework optimizes the relationship between the compact model structure and the distillation process to ensure the student model absorbs maximum semantic reasoning without hitting a performance ceiling.

Who it’s for

Moebius is designed for developers and researchers who need high-performance image inpainting and AI object removal but are constrained by hardware limits or require real-time inference speeds on edge devices.

Highlights

• Extreme Efficiency: Uses only 0.22B parameters, which is less than 2% of the size of models like FLUX.1-Fill-Dev.
• High Speed: Delivers over 15— inference acceleration with a latency of approximately 26ms per step.
• Competitive Quality: Performs on par with or better than 10B-level SOTA models across six benchmarks for both natural and portrait scenes.
• Specialized Focus: Optimized specifically for inpainting rather than being a bloated general-purpose model.