Moebius matches 10B inpainting giants with under 2% of the parameters

Researchers behind Moebius argue that image inpainting doesn’t need an 11.9-billion-parameter behemoth like FLUX.1-Fill-Dev to produce high-quality results. Their model packs just 226 million parameters — under 2% of the size — yet matches or beats those industrial generalists across six benchmarks spanning natural scenes (Places2) and portraits (CelebA-HQ, FFHQ), reportedly pulling ahead on complex textures and facial plausibility. The payoff is speed: roughly 26ms per denoising step on a single GPU and a claimed 15x-plus reduction in total inference time.

Two ideas carry the compression. The first is a redesigned diffusion backbone built around a Local-λ Mix Interaction block, which replaces quadratic self- and cross-attention with fixed-size linear matrices that fold spatial context and global semantics together — cutting parameters without losing the latent interactions that drive quality. The second is a distillation strategy that transfers knowledge from a larger teacher model (PixelHacker) entirely within latent space, avoiding costly pixel decoding, and dynamically reweights its losses by gradient norm to keep training balanced across fine features and overall diffusion trajectories.

The broader pitch is a bet on specialists over bloated generalists: when a task is narrowly defined, a small, purpose-built model can be smarter, lighter, and faster than scaling up. If the benchmark claims hold, the practical upshot is high-fidelity inpainting and AI object removal that can plausibly run on consumer-grade and edge hardware rather than datacenter GPUs.