Meta just shipped its first image generation model built in-house by Meta Superintelligence Labs. Muse Image is rolling out free across the Meta AI app, meta.ai, Instagram Stories in the US, and WhatsApp in select countries, positioning it as Meta’s answer to GPT Image and Gemini’s native image tools. The pitch that comes with it is a confident one. This isn’t just another prompt-to-picture generator, Meta says, it’s an agent that searches the web, writes code, and refines its own output before handing you a final image. Bold claims deserve testing, not trust. So I ran six prompts, each built to isolate one specific claim, and checked the results against what Meta actually promised.
Also read: Tencent’s Hy3 is free, open and almost as good as GPT-5.5: USA should be worried
I asked for a Simpsons-style illustration of someone scanning a QR code at a home appliances launch event, with the code linking to Digit’s homepage. Muse Image is supposed to write and execute code during generation, specifically to get things like QR codes right instead of drawing a fake-looking grid, and the output backed that up. The code came out with all three finder squares intact and a dense, structurally coherent pattern, sitting naturally inside a fully illustrated scene and if you scan the QR it will work perfectly taking you to the Digit homepage. This is one of Meta’s more testable claims, and it held up. Verified.
The graph requested was a bar chart representing market share for India’s smartphone sector in Q1 2026 by brands, according to IDC. Muse Image provided the information with market shares of 19.6 percent for Vivo, 17.1 percent for Samsung, 9.4 percent for Apple, and 8.4 percent for Xiaomi. When checking whether the figures coincide with IDC’s real Q1 2026 India’s figures, all matches perfectly, down to the percentage point. Vivo actually had 19.6 percent market share, Samsung was really second with 17.1 percent market share, and figures of Apple and Xiaomi also match perfectly. It is not by chance but an illustration of how genuine the search and ground method used by Muse Image is. The only mistake made is ignoring the styling requirements of a chart and providing a white flat design instead. Verified, and it is the highlight of the entire test.
Also read: Nothing Phone 4b vs Moto Edge 70 Fusion: Specs, design, price and more compared
I requested an image that depicts the weather conditions in Mumbai currently. I was provided with an appropriately rainy-looking Bandra-Worli Sea Link photo, rough sea, gloomy skies, people with umbrellas. However, it’s the monsoon time in Mumbai, so “rainy” weather can be assumed without conducting the live search at all by a model – not very good test at all. There are two more facts that indicate that it used generic visuals instead of something location specific: there are people strolling along the Sea Link that does not have a pedestrian way at all, and the water is depicted as a typical stormy ocean but not the characteristic grey-brown color of Mumbai waters with high concentration of sediments during the monsoon season.
I requested an illustration of my birthday party with 7 people, 12 candles on the cake, a dog wearing a party hat, and the time on the wall clock to be precisely 3:47 p.m. Muse Image is supposed to spot and fix any mistake it makes while creating the image, and that is when the accuracy claim was thoroughly tested. The number of people in the illustration was correctly 7, and the time on the clock was accurately 3:47 p.m. However, the candle count exceeded 12 quite noticeably. Failure to adhere to one out of three requested constraints is definitely not what I expected to see. Partial.
I uploaded a picture of myself and requested that my image be transformed into an entire movie poster in a sci-fi style, with a title and tagline, as well as some text at the bottom indicating the movie studio. While text generation has been one of the weaknesses of most image generators to date, in this case, the generated title and tagline and even the small text credits at the bottom were clearly readable. However, the face became distorted with all the heavy processing done to it. It is still recognizable, but several aspects have changed.
I tasked Muse Image with the job of inserting a leather sofa that had been uploaded by someone into a tiny apartment in Mumbai with proper proportions and lighting. All these factors, in addition to the direction of the shadows, as well as the soft afternoon light filtering through sheer curtains, were very convincing indeed. That’s the simplest case among all.
For the most part, in the easy-to-verifiably-test-and-hard-to-fake areas. The India smartphone image graph above is the one to note here; getting real-world data accurate, verified and represented correctly is a real capability, rather than buzzword BS. Product visualization and QR code creation held up as well.
It’s not that Muse Image failed on the easier tasks. Where it trips up is in the tricky business of precise accuracy of the real world when dealing with several factors at once; spatial accuracy that’s not simply “that looks about right,” pedestrians where there shouldn’t be any, coastal water that’s not quite how the coast actually appears, candle counts that aren’t quite right.
Also read: 5 best AI music generation tools in 2026