What Meta actually shipped
Meta Superintelligence Labs put out its first in-house media models on July 7, 2026: Muse Image and Muse Video, followed two days later by an upgraded reasoning model, Muse Spark 1.1. Meta calls Muse Image its most advanced image model, built to follow instructions faithfully, edit with precision, and compose a scene from several reference pictures at once.
Muse Video sits on the same pretraining base and adds native audio, with access coming soon to creators and Meta AI. Muse Image is live today in the Meta AI app and on meta.ai, in Instagram Stories in the US, and in WhatsApp in a limited set of countries, with Facebook listed as coming soon. For anyone in the EU, that availability map matters: at launch the reach is US and limited, so most European teams will meet the model later than the headlines suggest.
Why an agentic image model is different
The real shift is that Muse Image behaves like an agent, not a straight prompt-to-pixel generator. Instead of guessing what a chart or a QR code should look like, it writes and runs code to produce the real thing, searches the web to ground an image in factual and real-time references, and self-refines by spending more compute at generation time until the output holds up. It keeps coherence across several editing turns, and because it integrates with Muse Spark, the image and reasoning models share tools and plan the job together.
That buys real accuracy, correct logos, working QR codes, charts that actually add up, but it also blurs where the pixels came from. A picture that was partly assembled from a live web search and partly from training data is harder to trace than one drawn from a single prompt. The useful question stops being does it look right and becomes what did it read to make this, because the two are no longer the same thing.
What it means for a brand using it
Before you push brand assets through a model like this, the question to ask is not is the picture good but what data did it touch and who can see it. On launch day users pushed back over Meta drawing on their own photos, a training-data and consent concern reported by TechCrunch on July 7, and that backlash makes the abstract point concrete on day one. Under EU rules the same consent question lands harder, so a European team has more reason, not less, to be precise about it.
For a business the practical move is to treat brand images, logos and customer material as inputs with a paper trail, not as free fuel for a generator. Ask where a reference came from, whether the model can retain it, and who downstream can see the result before it becomes a public asset. The output can be genuinely more factual than older tools, and that is exactly why the provenance of the inputs deserves the same scrutiny as the quality of the picture.
Read next: Meta's Ad AI Launches Around Europe, Not In It | Even Realities Hit a Billion Without a Camera
