You can already pull a stranger out of a vacation photo and let software paint in the wall, sand, or sky that should sit behind them. The catch is that the best tools for this lean on huge AI models that need a beefy server to run. A new model says it can do the same trick at a tiny fraction of that size.
Moebius does one job: wipe an object out of a picture and fill the hole with new content that blends in. It comes from researchers at Huazhong University of Science and Technology and VIVO AI Lab. Its main rival, a model named FLUX.1-Fill-Dev, runs on 11.9 billion internal settings, the dials that a model tunes while it learns.
Moebius uses 226 million, under 2 percent as many. The team says that makes it more than 15 times faster on the same hardware.
Small model, big-model results
Smaller usually means worse. The team says that didn't happen here. Across six tests, Moebius beat its larger rival on portrait shots and on one everyday-scene test, while the bigger model stayed ahead on some other scenes. The authors call the overall quality close to that of the larger model. Still, these are lab numbers from the people who built it, so outside checks are worth waiting for.
If the claims hold, the payoff is simple. This kind of photo editing could run on more than just big servers, maybe on the hardware in your own home. The code and a research paper are both public, so other developers can test it.
For now it is research code and model files, not a finished app, though one browser demo from outside the team already exists. The team did not publish results from an independent lab. But if the figures survive, the gap between server-grade editing and home machines just got smaller.
