Moebius, a lightweight image inpainting framework from Huazhong University and VIVO AI Lab, drops a 10B‑level model’s size to just 0.22 B parameters. Its core, the Local‑λ Mix Interaction block, collapses spatial and semantic cues into fixed‑size matrices, cutting quadratic attention overhead while keeping rich latent interactions for photo editing tasks across consumer devices today.

By pairing the compact backbone with an adaptive multi‑granularity distillation strategy, Moebius matches or outperforms FLUX.1‑Fill‑Dev on six benchmarks, including Places2 and CelebA‑HQ. The training keeps all operations in latent space, avoiding costly pixel‑space decoding and dynamically balancing gradient losses for high‑fidelity inpainting results that scale to edge devices without heavy compute burden today.

Moebius achieves a 15× speedup, running at 26 ms per step on a single GPU. With under 2% of FLUX’s parameters, it delivers comparable quality while cutting inference time and memory usage, setting a new standard for specialist models that prioritize efficiency without sacrificing realism in real‑world applications across photography and video editing tasks today and and.

The launch follows a trend of compressing large foundation models for edge deployment. By demonstrating that extreme structural pruning, coupled with smart distillation, can preserve high‑fidelity generation, Moebius challenges the assumption that only massive models can produce studio‑grade results. This work signals a shift toward task‑specific, resource‑aware AI in modern creative workflows today and beyond.