A little image editing model the size of your palm actually caught up with a beast fifty times its size.

There’s this model called Moebius that’s built specifically for inpainting—you know, filling in missing or erased parts of an image naturally. The crazy thing is it’s only about 1/50th the size of mainstream inpainting models like FLUX, runs way faster, but supposedly matches their quality. It’s totally counterintuitive because everyone’s been assuming you need to stack up parameters for quality—if you want to catch up to a big model, you gotta build another big model. And inpainting is especially unforgiving; if you mess up even a tiny bit, the human eye instantly spots that blurry edge or warped texture.

From what I’ve heard, their approach uses compression techniques to cram the work into very few parameters, while training directly on the output of large models. Kind of like an apprentice copying the master’s finished work, practicing until they can replicate the result with a fraction of the effort. As someone who’s written code, I buy into this distillation idea—small models cribbing from big models’ homework is a legit path. Just wondering if it’ll fall apart on extremely complex textures.

Distillation should really be played this way from the start.

One in fifty and still tied? That’s actually impressive.

Inpainting is the one you absolutely can’t miss, but actually doing it well is really tough.

Small models are fast as hell, but man, I’m scared they’ll mess up on edge cases.

“Apprentice copying the master” — that’s a pretty spot-on analogy.