Meta AI starts U.S. rollout of Muse Image and Muse Video

• Muse Image shipped in Meta AI as the first image model from Meta Superintelligence Labs, and Muse Video was previewed in chrisfirst's launch thread.
• The workflow is agentic: chrisfirst's feature list says Muse Image can search the web, write and run code, and work with Muse Spark.
• Self-critique emerged during reinforcement learning, according to chrisfirst's self-correction clip.
• The rollout starts in the U.S. and expands gradually, with Instagram Stories support named in ozansihay's rollout note.
• Meta is pairing generation with Content Seal watermarking, which chrisfirst's watermarking note says survives cropping, compression, resizing, and screenshots.

Muse can search the web and write code for artifacts where image models usually fail, charts, QR codes, and figures, according to chrisfirst's feature list. chrisfirst's RL note says self-critique appeared because it improved reward. chrisfirst pointed readers to a read-more link after listing Content Seal and Muse Video.

Muse Image is described as the first image model from Meta Superintelligence Labs. The same thread says Meta previewed Muse Video and frames the image model as a generator that can search, code, and refine its output.

Posts framed video two ways: chrisfirst said Meta previewed Muse Video, while ozansihay's rollout note said Muse Image and Muse Video would be usable in Meta AI starting today.

ozansihay said distribution starts in the U.S. and expands gradually. The same note said Turkey did not have access yet and that the models will also appear inside Instagram Stories.

The mechanics chrisfirst listed are:

1. Web search to ground images in current facts.
2. Code writing and execution for charts, QR codes, and figures.
3. Collaboration with Muse Spark to make GIFs, websites, and visual games.

For creators, the sharp bit is code execution: the model is being pitched for visual outputs that need exact structure, not only prettier pixels.

Muse Image critiques drafts mid-generation, according to chrisfirst. The thread says the behavior emerged during RL training because self-correction produced higher reward.

Muse Image accepts several references at once:

• person
• product
• outfit
• style
• environment

chrisfirst described the output as a single blended scene. That is the reference-stack workflow art directors already use, compressed into one generation pass.

chrisfirst said Muse Image ranks #2 by human-preference Elo in all three Arena image categories. His test set included a Monaco F1 cockpit, a perfume product shot, a pink interior, and a restaurant scene.

AIandDesign posted another batch and called the new Meta image model "pretty awesome."

Every Muse image made in Meta AI carries Content Seal, an invisible watermark. chrisfirst says it survives:

• cropping
• compression
• resizing
• screenshots

A public detector is in preview, and video support is next.

LLMJunky liked the look of the model but used the launch to pressure Meta to return to open weights. In a follow-up, LLMJunky argued open-sourcing the model would be cooler than ranking second.