Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance
Kangsheng Duan, Ziyang Xu, Wenyu Liu, Xiaohu Ruan, Xiaoxin Chen, Xinggang Wang
arXiv:2606.19195 · 139▲ · cs.CV
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
Moebius is a 0.22B-parameter image inpainting model built on a latent diffusion U-Net whose transformer blocks are replaced with a Local-Lambda Mix Interaction (LLambdaMI) block that summarizes spatial context and semantic priors into fixed-size linear matrices instead of quadratic attention. It is trained with an adaptive multi-granularity knowledge distillation scheme that aligns the small student to the larger PixelHacker teacher entirely in latent space using gradient-norm-balanced losses.
Why it matters
Industrial inpainting models like FLUX.1-Fill-Dev have ~11.9B parameters and need many sampling steps, making them impractical on edge devices or latency-sensitive apps. Moebius reaches comparable quality with under 2% of the parameters and about 15x faster total inference, which matters if you want to run inpainting locally or cheaply.
Practical takeaway
Watch for lightweight, task-specific inpainting specialists that can be fine-tuned per benchmark and deployed on modest hardware (they trained on L40S and fine-tuned on RTX 3090s) rather than reaching for a 10B generalist. The core lesson is that extreme architecture compression only works when paired with latent-space distillation from a bigger teacher.
Key result
On Places2 (Small), Moebius scores FID 0.92 / LPIPS 0.091 versus FLUX.1-Fill-Dev's 0.94 / 0.099 and PixelHacker's 0.82 / 0.088, at 0.22B params, 0.154 TFLOPs, and 0.52s total inference versus FLUX's 8.05s. Note these are fine-tuned in-domain results on academic benchmarks, not zero-shot; on out-of-distribution sets Moebius (FID 17.81 natural) trails both PixelHacker (13.84) and FLUX (14.52).
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.