We are pleased to launch Muse Image and preview Muse Video, the first media generation models developed by Meta Superintelligence Labs.
Muse Image is the most advanced image generation model ever. Follow instructions closely, edit with precision, compose from multiple references, and use Instagram in a social context. It also provides the ability to use agent tools and integrates with Muse Spark. Built on the same pre-training base, Muse Video provides superior visual fidelity with native audio support.
Muse Image is currently available in the Meta AI app and meta eyeavailable on Instagram Stories in the US, WhatsApp in select countries, and coming to Facebook soon. Muse Video will be available to creators and meta AIs soon.
Muse image: Generating an agent image
Muse Image acts as an agent rather than mapping prompts directly to images. Muse Image invokes search and coding tools to improve accuracy, self-tune its own generation, and improve through scaling of calculations during testing. Muse Image is also integrated with Muse Spark, allowing the two models to share tools and jointly plan powerful agent media generation.
Using tools
Muse Image provides access to tools to enhance agent functionality.
coding. During reinforcement learning, Muse Image learns to write and run code that generates accurate plots and QR codes, and conditions the rendered diagrams to improve the accuracy of the generated images. You can also integrate Muse Spark and Muse Image to combine code and media generation to create animated GIFs, image-embedded websites, and interactive visual games.
search. Muse Image learns how to search the web and find generated images based on factual and real-time information and visual references. Enabling search improves the factual accuracy of knowledge-intensive prompts, especially prompts about current events and real-world facts.
Muse images can be improved using the search tool. Win rate with internal ablation.
self-improvement
Muse Image reflects on and improves its own work during its train of thought. This self-regulatory behavior can take many forms. For example, use other tactics such as editing the current image draft locally if a small detail is wrong, generating a new image from scratch if a large part is wrong, or using tools to generate a more factual image. This behavior was not designed by us. Rather, it emerged during RL training simply because self-refinement resulted in a better image and therefore a higher reward.
The muse image is improved through emergent self-refinement. Win rate with internal ablation.
self-improvement
Search for reference images
self-improvement
Assembling the spread
It brings together glossy magazine pages, weaving proofing instructions, headlines, and fashion portraits into a cohesive layout, while checking out details in typography, spacing, and gold accents for a sophisticated editorial feel.
self-improvement
Refine the expression
Correct the formula to include the missing division slash, update the layout so that the formula is S = n(n + 1) / 2, and double-check the surrounding text for clarity and consistency.
self-improvement
Check the image
We review generated magazine images for macro details, glossy layouts, and accurate proofs before sharing. You’re ready to tweak the typography or swap out the portrait if you wish. What do you want to adjust next?
Scaling compute during testing
Like language models, Muse Images get better the more you think about them during inference. With more compute during testing, the model makes more inferences, uses more tool calls, and uses more self-refinement steps to improve generation. Increasing inference strength (and test-time computation) improves human preference Elo scores, showing an approximately log-linear scaling relationship. Notably, although this computation spans two very different kinds of work (textual tokens for inference and visual tokens for generation), the quality is a function of the total computation.
We found that using your token budget wisely is just as important to effectively scale your test time. Best-of-N (BoN), where the model generates multiple images and keeps the best one, improves quality early but quickly saturates. Spending the same computation on intentional inference scales much better. When you combine deduction and tools, it becomes complex. Tools allow you to go beyond what the model already knows and fill in gaps that inference alone cannot by searching for missing references or writing code to retrieve accurate details.
Muse Image improves by scaling the compute during testing. Eroticism from internal ablation.
image editing
Muse Image precisely edits images and changes them to exactly what you want. As the example shows, you can follow various instructions.
Muse Image maintains consistency across editing turns, supporting iterative refinement and free brainstorming toward your desired results.
Composition of multiple reference images
Muse Image can compose elements such as people, objects, clothing, styles, environments, etc. from many input reference images in a prompt. Supports inline interleaving of text and images with prompts for complex image compositions.
Image benchmark
Muse Image holds the No. 2 spot in the arena for text-to-image conversion, single-image editing, and multi-image editing, according to the Elo ranking of human preferences at the time of this writing.
Arena Elo rankings as of July 5, 2026.
Arena Elo rankings as of July 5, 2026.
Arena Elo rankings as of July 5, 2026.
Muse video preview
To coincide with the release of Muse Image, we’re sharing an early preview of Muse Video. It provides superior performance in rapid compliance, visual fidelity, and temporal consistency. We are investing in areas where there are current performance gaps, such as audio and video synchronization and physically accurate high-speed motion. Muse Video is coming soon to creators and meta AIs.
At Arena, Muse Video ranks #3 in human preference Elo for text to video conversion as of this writing.
Arena Elo ranking as of July 5, 2026
Contents sticker
To help you verify whether an image is generated by AI, Muse Image includes Content Seal, an invisible watermarking system. Images created by Muse Image, such as in the Meta AI app meta eye Even if you crop, compress, resize, or take a screenshot, it remains intact and conveys hidden provenance signals. We plan to extend content seals to videos soon. Currently in preview detection tool This allows you to see if an image contains the Content Seal watermark and is the first way to better understand if an image was created with Meta AI.
Metaproduct Muse image
Muse Image is deeply connected to the meta ecosystem. Combined with Meta AI’s social tools, users can create images with friends and reimagine their Instagram photos. Our continued investment in image and video generation enables creators and businesses to generate dynamic content across Meta products.
Marketing assets for small businesses like @averyandme
Image generated by Meta AI from @mentioning a public Instagram account.
Personalized presets directly on Instagram
Explore additional resources
Try out Muse images with Meta AI
Read more Meta AI Update
author:
Meta Super Intelligence Research Institute
Latest updates delivered to your inbox
Subscribe to our newsletter to receive Meta AI news, events, research breakthroughs, and more.
