Moebius: High-Performance Lightweight Image Inpainting Framework

Moebius achieves 10B-level inpainting quality with 0.22B parameters

Moebius is a lightweight image inpainting framework that delivers generation quality on par with, and occasionally surpassing, industrial generalist models like FLUX.1-Fill-Dev and SD3.5 Large-Inpainting, despite using less than 2% of the parameters. By utilizing a compact 0.22B (226M) parameter architecture, Moebius reduces the computational overhead associated with 10B-level foundation models, making high-fidelity inpainting viable for consumer-grade and edge devices.

Extreme efficiency and inference acceleration

Moebius provides a significant reduction in latency and resource requirements compared to large-scale industrial models:

  • Parameter Reduction: Moebius operates with 0.22B parameters, compared to the 11.9B parameters of FLUX.1-Fill-Dev.
  • Inference Speed: The model achieves an inference latency of 26.01 ms per step on a single GPU.
  • Total Runtime: Moebius delivers a >15x total runtime acceleration over 10B-level models.

Core technical innovations

Moebius overcomes the representation bottleneck typically caused by extreme structural compression through two primary synergistic innovations:

Local-λ Mix Interaction (LλMI) Block

To avoid the quadratic computational overhead of standard attention mechanisms, Moebius reconstructs the diffusion backbone using the LλMI block. This architecture summarizes spatial contexts and global semantic priors into fixed-size linear matrices, allowing the model to preserve complex latent interactions while drastically reducing the total parameter count.

Adaptive Multi-Granularity Distillation

Moebius uses a distillation strategy to transfer representational capacity from a teacher model (PixelHacker) to the student model (Moebius). This process occurs strictly within the latent space to avoid the high cost of pixel-space decoding. The strategy employs a gradient norm adaptive loss weighting mechanism to balance multiple gradient-based losses, aligning supervision across various scales—from microscopic intermediate features to macroscopic diffusion trajectories.

Performance across benchmarks

Moebius was tested across six comprehensive benchmarks covering both natural scenes (using the Places2 dataset) and portrait scenes (using CelebA-HQ and FFHQ). The results indicate that the model maintains high fidelity in complex textures and facial plausibility, effectively serving as a task-specific specialist rather than a bloated generalist model.

Community insights and practical observations

While the technical specifications show significant gains, community feedback highlights several practical considerations regarding the model's real-world application:

  • Output Quality: Some users noted that inpainted regions can appear visibly smoother than the surrounding original image, lacking the high-frequency detail of natural photos. One user observed that the model "performed very badly on novel objects."
  • LImitations: Current practical use may be limited by a 512x512 output resolution.
  • Deployment: The model's efficiency has already enabled community-driven browser-based implementations. For example, developer Simon Willison ported the model to ONNX for an interactive browser demo, and another developer created a browser-based experimentation tool at inpaintlab.com.

"The inpainted areas are, as usual, very smooth compared to the detailed, 'high frequency' look of natural photos. Barely useful enough to erase things in thumbnails."

"It did work reasonably well with natural images but inpainted regions were visibly smoother than surroundings, and performed very badly on novel objects."

Sources