Meta announced the launch of Muse Image and previewed Muse Video, the first media generation models developed by Meta Superintelligence Labs. These new tools represent a significant advance in the creation of visual and audiovisual content.
Muse Image, considered the company’s most sophisticated model to date, stands out for its ability to follow instructions precisely, perform detailed edits, compose elements from multiple references and use Instagram for social context. Additionally, it offers interactive tooling capabilities and integrates with Muse Spark, while Muse Video, built on the same pre-training foundation, promises exceptional visual fidelity with native audio support.
In terms of availability, Muse Image can now be accessed on the Meta AI app, on the meta.ai platform, on Instagram Stories for users in the United States and on WhatsApp in some countries. Soon, it will also be available on Facebook. Muse Video is coming soon to content creators and on Meta AI.
Generation of agency images with Muse Image
Instead of simply converting instructions into images, Muse Image works as an intelligent agent. It invokes search and coding tools to refine accuracy, improves its own generations, and optimizes its performance by scaling computational power during testing. Its integration with Muse Spark allows both models to share tools and plan together, providing a powerful agency media generation capability.
Ability to use tools
Muse Image has access to several tools that enhance its performance capabilities. He learned, through reinforcement learning, to write and execute code to produce accurate graphics and QR codes, conditioning on the rendered figures to improve the accuracy of the generated images. Combining with Muse Spark allows you to create animated GIFs, websites with embedded images, and interactive visual games.
A notable example of his programming prowess is creating a readable QR code within a complex Korean manhwa-style illustration, ensuring the code functions correctly while artistically integrating into the scene. The model is also capable of generating detailed fractal graphics, such as Julia sets and Sierpinski triangles, and applying them to style-specific poster layouts.
In the area of animation, Muse Image demonstrates the creation of flipbook sequences. It enhances character details, adjusts their positions to generate dynamic interactions, such as a punch and a dodge movement, while maintaining consistency of scenery and lighting in each frame. Another application involves creating different life stages of a pet, such as a cream Persian cat transformed into a newborn kitten or a slender teenager, for use in visual games, converting the images into data URIs for instant loading.
In addition to programming, search functionality allows Muse Image to query the web to substantiate generated images with real-time, factual information as well as visual references. This ability is crucial for improving accuracy in requests that require prior knowledge, especially about current events and real-world facts.
The research tool was used to identify fashion trends for summer 2026, analyzing catwalk reports and product catalogs to generate purchase links and outfit suggestions. In the field of science, the model researched diagrams and reliable sources on the giant impact hypothesis to create accurate and visually consistent infographics about the formation of the Moon. In another demonstration, Muse Image planned and selected unbranded vintage items from Facebook Marketplace in San Francisco, combining them with the cozy and simple decor style of a bedroom.
Model self-improvement and reflection
One of the emerging characteristics of Muse Image is its ability to self-improve. The model reflects on its own work and improves it within its reasoning, whether making local edits for small errors or generating an entirely new image when there are major flaws. This behavior emerged during reinforcement learning training, as it resulted in better quality images and, consequently, greater rewards.
In examples of self-improvement, the model demonstrated the ability to put together a magazine page, integrating review, title and fashion portrait steps with a cohesive layout and impeccable details. Another illustration involves correcting mathematical formulas, adjusting the layout and checking the clarity of the text. The process also included final review of generated images, confirming macro details, glossy finish and accuracy before sharing.
Scalability and computational efficiency
As with language models, Muse Image progressively improves the more information it processes during inference. With greater computational power at test time, the model deepens its reasoning, makes more tool calls and employs more self-improvement steps to refine its generations. Increasing reasoning ability and computational power results in improvements in human preference Elo scores, demonstrating an approximately log-linear scaling relationship.
Studies have revealed that judicious use of token budgeting is essential for effective scaling. While the Best-of-N method (which generates multiple images and selects the best one) improves quality initially, it saturates quickly. On the other hand, investing that same computational power in deliberate reasoning provides significantly better scalability. Reasoning and the use of tools mutually enhance each other, allowing the model to go beyond its previous knowledge, whether searching for references or writing codes for precise details, filling gaps that reasoning alone would not be able to fill.
Precise image editing
Muse Image demonstrates a precision image editing capability, changing exactly what the user requests in a variety of instructions. It can, for example, dissipate fog from an image to reveal a valley, modify the petals of a flower to form a rainbow gradient, or change texts and numbers on a sign, in addition to adjusting the zoom to contextualize a scene.
The model also maintains coherence throughout the editing stages, facilitating iterative refinement and open-ended experimentation to achieve a desired result. In one demonstration, he was able to transform a living room into the Japandi style, using specific elements from other images. Another complex example included composing a picnic between a cat and a dog, then imagining it as a framed photo in a cafe, and finally generating the establishment’s menu with a themed item and animal illustration.
Creation of compositions with multiple references
Muse Image’s capabilities extend to composing elements from multiple reference images inserted into the prompt, including people, objects, clothing, styles and environments. The template supports interleaving text and images directly at the prompt for complex visual creations.
This is demonstrated by composing an image of a person riding a bicycle, wearing clothes and sitting on a park bench, all with a specific drawing style, using multiple visual references. In another case, the model managed to change the color of a living room to a specific pattern on the wall, adding a coffee table, sideboard and a 65-inch television with a specific image on the screen, and inserting a person sitting on the sofa with specific items.
Muse Image currently ranks second in Arena for text-to-image conversion, single-image editing, and multi-image editing, as per the latest Human Preference Elo ratings.
