Moebius: 0.22B-Parameter Model Matches 10B-Level Image Inpainting Performance

Forum topic · 小凯 · 2026-06-19

Summary

Moebius is a lightweight image inpainting framework from Huazhong University of Science and Technology and vivo AI Lab that achieves 10B-level quality with only 0.226B parameters. It introduces two core innovations: the Local-λ Mix Interaction (LλMI) module, which linearizes both self-attention and cross-attention by summarizing context into compact linear matrices, and an adaptive multi-granularity distillation strategy with gradient-norm-based loss balancing in latent space. Benchmarks show Moebius achieves FID 0.92 on Places2 and 5.39 on CelebA-HQ, matching or beating FLUX.1-Fill-Dev (11.9B parameters) while running 15x faster (0.52s vs 8.05s per image) with about 2% of the parameters. The result demonstrates that task-specific expert models can surpass large general-purpose diffusion models on targeted tasks, enabling real-time, on-device image inpainting for mobile and edge deployment.

Moebius is a lightweight image inpainting framework (arXiv: https://arxiv.org/abs/2606.19195) by Kangsheng Duan, Ziyang Xu et al. from Huazhong University of Science and Technology and vivo AI Lab. It delivers image inpainting quality comparable to or better than 11.9B-parameter FLUX.1-Fill-Dev using only 0.226B parameters — under 2% — with roughly 15x faster inference.

Key points

Problem: General-purpose inpainting models (FLUX.1-Fill-Dev at 11.9B, SD3.5 Large-Inpainting at 8.05B) achieve strong zero-shot quality but are too large, slow (50 sampling steps), and memory-hungry for mobile/edge deployment.
Innovation 1 — LλMI module: Replacing heavy blocks with naive lightweight modules (Depthwise Conv, linear attention, Mix-FFN) collapses quality (FID 32.75 → 37.65–43.58). LλMI instead summarizes context into compact linear matrices: a Local-λ module linearizes self-attention via a semantic map λ_c = softmax(K)^T × V and a positional map λ_p = Conv3D(V), reducing complexity from O(N²) to O(N). An Interactive-λ module linearizes cross-attention against external semantic priors (LCG embeddings) — something GLA-style linear attention cannot do. Mix-FFN with depthwise-augmented structure compresses the feedforward network.
Innovation 2 — Adaptive multi-granularity distillation: Operates in latent space to avoid pixel-space decoding costs. Combines coarse-grained alignment (16×16 bottleneck features), fine-grained output alignment (64×64), task noise-prediction loss, and E-LatentLPIPS perceptual loss, with dynamic loss weights based on gradient norms to resolve gradient conflicts without manual tuning.

Performance

| Model | Params | Latency/step | Total steps | Total time | Places2 FID | CelebA-HQ FID | |-------|--------|--------------|-------------|------------|-------------|---------------| | Moebius | 0.226B | 26.01ms | 20 | 0.52s | 0.92 | 5.39 | | PixelHacker | 0.862B | 46.89ms | 20 | 0.94s | 0.82 | 4.75 | | SD3.5 Large-Inp. | 8.057B | 151.02ms | 28 | 4.23s | 3.02 | 11.80 | | FLUX.1-Fill-Dev | 11.902B | 161.01ms | 50 | 8.05s | 0.94 | 10.13 |

On Places2 (512×512, 40–50% masks): FID 0.92, LPIPS 0.091 — best among all compared methods including MAT (9.27), LaMa (21.07), PowerPaint (13.25), and FLUX.1-Fill-Dev (0.94/0.099).
On FFHQ (256×256): FID 8.15 vs FLUX's 11.19, slightly behind PixelHacker (6.35), which uses 4x more parameters.
Distillation is essential: identical 226M architecture scores FID 26.43 with distillation vs 33.42 without.

Training details

Teacher model: PixelHacker (0.86B, official pretrained weights)
Optimizer: Muon (weight decay 0.1); LR 2e-4, decayed 0.1x at 111K and 129K iterations
Distillation: 16x NVIDIA L40S, batch size 768, 138K iterations, BF16
Fine-tuning: Places2 (4x 3090, 51K iters), CelebA-HQ (2x 3090, 60K iters), FFHQ (4x 3090, 117K iters)
All latency benchmarks: single L40S, batch size 1, 512×512

Limitations and outlook

Moebius is an inpainting specialist with no zero-shot generalization to other generation tasks
Distillation depends on a high-quality teacher model
Future work: extending LλMI to super-resolution, denoising, and editing; teacher-free/self distillation; compression below 100M parameters

Takeaway

Moebius shows that highly optimized task-specific expert models can beat large general-purpose diffusion models on targeted tasks — enabling near-real-time inpainting (0.52s per image) on consumer hardware, roughly 1/15 the inference cost of FLUX, and practical on-device deployment for mobile, privacy-sensitive, and interactive applications.