A strong image generator is no longer just a prompt box. Meta says Muse Image can search the web, write and run code, and revise its own output, while the company has also offered an early look at Muse Video.
Muse Image works beyond the prompt
Meta describes Muse Image as an agent that can call tools while creating an image. It can search for references and current information, use code to make details such as plots and QR codes more accurate, and refine an image when its first attempt misses the mark.
That revision can mean making a small local edit, starting a new image, or changing tactics and using a tool. Meta says the model learned to self-refine during reinforcement learning because better images earned higher rewards. The company says it did not explicitly design that behavior.
Muse Image can also work with Muse Spark. Meta says the two models can plan together and combine code with media generation to make animated GIFs, websites with embedded images, and interactive visual games.
More thinking, not just more samples. Meta says Muse Image improves when given more computation at inference time. Its internal tests found that generating several candidates and keeping the best, a method known as Best-of-N, improves results at first but soon levels off. Spending the same budget on deliberate reasoning scales better, and combining reasoning with tool use produces further gains.
Meta describes the relationship between additional test-time compute and human-preference Elo scores as approximately log-linear. That compute includes both text tokens used for reasoning and visual tokens used to generate images. The company also says that how the model spends its token budget matters, which is a less glamorous finding than “more compute,” but generally the more useful one.
Editing, references, and rankings
Muse Image can make targeted edits, maintain consistency over multiple editing turns, and combine details from several reference images. Those references can include people, objects, clothing, styles, and environments. Prompts can mix text and images inline.
Meta says Muse Image holds the No. 2 position on Arena for text-to-image, single-image editing, and multi-image editing, based on human-preference Elo rankings at the time of its announcement. These are rankings reported by Meta, not a guarantee that every prompt will produce the desired result.
Muse Video is a preview
Meta also introduced Muse Video as an early preview. The model shares Muse Image’s pretraining base and supports native audio. Meta says it performs competitively on prompt adherence, visual fidelity, and temporal consistency, while identifying audio-video synchronization and physically accurate fast motion as areas still needing work.
The company says Muse Video ranks No. 3 on Arena for text-to-video human-preference Elo at the time of the announcement. It is expected to come to creators and Meta AI, though Meta did not give a date in its announcement.
Where the models are available
Meta says Muse Image is available in the Meta AI app and on meta.ai, in Instagram Stories in the United States, and in WhatsApp in a limited number of countries. It is expected to come to Facebook. Muse Video is not yet broadly available.
Images made with Muse Image in the Meta AI app and on meta.ai carry Content Seal, Meta’s invisible watermarking system. The company says the hidden provenance signal survives cropping, compression, resizing, and screenshots. Meta is also previewing a detection tool that can check whether an image has a Content Seal watermark, and plans to extend the system to video.
The practical pitch is not simply that Meta can generate images. Muse Image is built to look things up, try again, and use tools when a prompt calls for more than a plausible guess. That may make it more capable, though the usual small print still applies: better process is not the same thing as a perfect picture.
Sources
- Introducing Muse Image and Muse Video — Meta AI