Meta built an AI model that is already the second most powerful image model in the world. Meta's Superintelligence Labs just dropped Muse Image and Muse Video, their first media models, and Muse Image debuted at number two on the Image Arena leaderboard.
The headline feature was a complete accident.
It Checks Its Own Work
During training, Muse Image started to review its own images and redo the weak parts. Meta says they never programmed this. It emerged on its own, because better images yield a higher reward.
That is the part worth sitting with. The self-correction loop was not a designed feature. The training incentive made the model discover that critiquing and revising its own output produced better results, so it started doing exactly that.
It Looks Things Up Instead of Guessing
Most image models hallucinate text and numbers. Ask for a chart and you get plausible-looking nonsense. Ask for a QR code and you get a decorative pattern that scans to nothing.
Muse takes a different path. It can search the web to get a fact right, and it can run actual code to draw an exact chart or a working QR code. Instead of painting a picture of data, it computes the data and renders it.
Every Image Is Secretly Tagged
Meta bakes an invisible content seal into every output. It survives screenshots, cropping, and compression, and you can verify any image at meta.ai. As AI images get harder to spot, a watermark that survives the usual laundering steps matters.
What Comes Next
Muse Video ships next, with sound built in.
This is where AI image tools are heading: models that check their own work. The interesting shift is not prettier pictures. It is generation systems that verify facts, run code, and critique their own outputs before you ever see them.