Meta launches Muse Image: what it is, how it works, and what it means for AI builders

Meta launched Muse Image, its first image-generation model from Meta Superintelligence Labs, integrated into Meta AI, Instagram, and WhatsApp. The model can generate images from text prompts, edit existing photos, and create images featuring friends based on publicly available Instagram posts. It also includes coding and web search capabilities, with advertiser access and a potential API pathway on the horizon.

What happened

Meta released Muse Image, the first image-generation model from its Superintelligence Labs division led by Chief AI Officer Alexandr Wang. The model is rolling out inside the Meta AI chatbot and across Instagram and WhatsApp. Users can generate images from text prompts, alter existing pictures, and create images featuring friends or creators based on their publicly available Instagram posts. Those who do not want their content reused or remixed by AI can opt out in the settings menu. All images made with Muse Image include an invisible watermark, and the system has safety precautions to prevent violations of Meta's terms of service, including CSAM.

Advertisers will soon have access to the model for creating marketing materials through Meta's Advantage+ suite. Meta plans to sell access to its AI models to outside developers via a cloud offering, allowing the company to monetize its data center infrastructure and expensive AI chips. The company is also building a video generation model expected in the coming months.

Why AI builders should care

Muse Image is not just another image generator. Under the hood, it shares several similarities with reasoning-optimized LLMs. The model features a tool-use capability: if a prompt doesn't contain all the details required to generate a file, it can use a search tool to retrieve the needed information from the web. Another built-in tool enables the model to generate code, such as writing a Python script to turn spreadsheet data into a graph. Muse Image can also loop in Muse Spark, Meta's reasoning model, when its coding capabilities are insufficient to process a user request.

Unlike traditional best-of-N (BoN) approaches that generate multiple images and pick the best one, Muse Image engages in "deliberate reasoning" before generating an image. The model supports a test-time compute feature, meaning increasing the amount of hardware at its disposal boosts output quality. The model also reviews the images it generates before displaying them and makes tweaks when necessary, a self-refining behavior that emerged on its own during the reinforcement learning phase of training.

Practical implications

For AI builders, the most immediate takeaway is that Meta is building a multimodal reasoning system, not just a standalone image model. The ability to search the web for missing prompt details and write Python scripts to generate accurate visualizations means Muse Image can handle complex, multi-step tasks that other image generators cannot.

The integration with Instagram and WhatsApp also creates new possibilities for social media automation, content creation, and advertising workflows. Builders who work with Meta's ecosystem should watch for the advertiser access rollout and the potential API pathway, which could open up programmatic access to the model.

Caveats

On launch, Muse Image is available via the Meta AI chatbot in a limited number of markets. It also powers a set of new image effects in Instagram stories. Meta plans to bring the model to Facebook, Messenger, and more parts of Instagram in the future. The company has not announced pricing for developer access or API usage, and the timeline for broader rollout remains unclear. The model's self-refining behavior and test-time compute scaling are vendor claims that have not been independently verified.

FAQs

What is Muse Image and who developed it?

Muse Image is Meta's first image-generation model from Meta Superintelligence Labs, led by Chief AI Officer Alexandr Wang. It is integrated with Meta AI and can generate and edit images based on prompts, with safety safeguards including an invisible watermark and opt-out options for users who do not want their content reused.

How can Muse Image be used within Instagram and WhatsApp?

Muse Image is embedded across Meta AI, Instagram, and WhatsApp. Users can generate images from text prompts, edit existing photos, and create images featuring friends based on their publicly available Instagram posts. The generated images can be shared directly to chats, stories, or feeds.

Can Muse Image edit existing photos and remove fog or adjust camera angles?

Yes, Muse Image can edit existing photos, including removing fog and changing camera angles. Users can also refine the model's output by uploading a sketch that explains what edits it should make.

Will Muse Image be available to developers via an API?

Meta has suggested a potential API pathway via its AI infrastructure strategy, with future developer access possible. The company plans to sell access to its AI models to outside developers through a cloud offering, which could include Muse Image.