Meta Superintelligence Labs released Muse Image on Tuesday, its first in-house image generation model, and previewed a companion video model called Muse Video. The release, detailed on Meta's technical blog, is live now in the Meta AI app and on meta.ai, in Instagram Stories in the US, and in WhatsApp in a handful of countries.
An image model that argues with itself
Most image generators map a prompt to pixels and hand you the result. Muse Image does something stranger. It runs as an agent, calling search and code execution mid-generation, editing photos on instruction, and stitching together several reference images at once. It also plugs into Muse Spark, Meta's language model from April, so the two can share tools.
The genuinely interesting part is the self-correction. When a draft comes out wrong, the model can patch a small detail, regenerate from scratch, or switch tactics and reach for a tool. Meta is upfront that nobody built this in. It emerged during reinforcement learning because fixing its own drafts scored higher reward. That is the kind of claim that deserves a raised eyebrow, since "we didn't design it, it just appeared" is a convenient story for any lab, but Meta at least frames it as a byproduct of reward optimization rather than magic.
The compute math
Here's where Meta makes a sharper argument. Pairing self-refinement with a bigger thinking budget produces what the company describes as roughly log-linear gains in human-preference Elo. The old trick, Best-of-N, where you spit out several images and keep the winner, improves things early then flattens fast. Spending that same compute on deliberate reasoning, Meta says, scales better.
Whether that holds outside Meta's own internal ablations is the open question. Every number in the blog comes from Meta measuring Meta.
Where it ranks
On Arena, Muse Image sits at No. 2 for text-to-image, single-image editing, and multi-image editing by human-preference Elo, as of July 5. Muse Video lands at No. 3 for text-to-video. Second place is respectable. It is also not first, and Meta's own benchmarks reportedly show the model trailing OpenAI's latest image model while beating Google's Nano Banana 2. So the pitch is less "we won" and more "we're finally in the room."
Every image carries Content Seal, Meta's invisible watermark, which the company says survives cropping, compression, resizing, and screenshots. A detection tool is in preview.
Muse Video is coming to creators and Meta AI, with no firm date. Meta says Content Seal extends to video soon, and advertisers get Muse Image through Advantage+ in the coming weeks.
