Meta unveiled its media generation models Muse Image and Muse Video on July 7, 2026. Both were built by Meta Superintelligence Labs, with Muse Image launched and Muse Video in an early preview phase. Beyond following instructions precisely and composing from multiple reference images, Muse Image writes and executes its own code to produce accurate plots, QR codes, and animated GIFs, and searches the web to ground images in factual, real-time information. On Arena as of July 5, 2026, Muse Image ranked #2 in text-to-image, single-image editing, and multi-image editing, while Muse Video ranked #3 in text-to-video human-preference Elo. ASAP summarizes it from Meta's official announcement and primary reporting by TechCrunch.

What Muse Image does: compose, self-refine, use tools

What Meta foregrounds in Muse Image is not fidelity but that an image model behaves like a tool-using agent. Muse Image follows instructions faithfully and edits with precision, and it pulls elements from multiple reference images to compose a single scene. It works as an agent with self-refinement, polishing results through local edits or full regeneration. It supports inline prompts that mix text and images, and at the planning stage it works jointly with Muse Spark. The tool-use axis stands out most. To produce accurate plots, QR codes, and animated GIFs, Muse Image writes and runs its own code, and it searches the web to ground images in real-time facts.

Where it stands: Arena #2, Muse Video #3

Muse Image ranked #2 on Arena as of July 5, 2026 across text-to-image, single-image editing, and multi-image editing, while Muse Video ranked #3 in text-to-video human-preference Elo. Meta presented Muse's performance through these public leaderboard rankings rather than through internal benchmark scores. Distribution differs by surface. Muse Image is available now in the Meta AI app and at meta.ai, in Instagram Stories in the US only, and in WhatsApp in a limited set of countries, with Facebook coming later. Muse Video is coming soon to creators and Meta AI.

Why it matters now: image models shift to tool-using agents

Muse Image's design shows the baseline for generation models shifting from pure image synthesis to tool-using agents. The weakness of earlier image generators was clear. They could not render text accurately, got numbers in charts wrong, and painted plausible fiction into scenes that needed facts. Running code to build charts and QR codes and checking facts via web search takes direct aim at that weakness. Instead of painting all the pixels in one pass, it hands the needed computation and verification to external tools to raise the reliability of the result. The self-refinement axis points the same way. Not ending with a single generation but polishing through local edits or regeneration means image generation is moving from "one lucky shot" to "iterative convergence." The competitive axis is moving from how pretty an image is to how accurate and verifiable it is.

The real weapon is distribution, and provenance as the response

As much as performance, what deserves attention in Muse is its distribution path and its provenance mechanism. Muse Image did not ship only as a standalone app; it was placed directly into Meta's large user surfaces such as Instagram Stories and WhatsApp. When rival models are of similar quality, where and how widely you land decides actual usage, and Meta uses that distribution network as a weapon in itself. At the same time, Meta attached an invisible watermarking system called Content Seal so the provenance of generated media can be verified, and it offers a detection tool at meta.ai/identification. Shipping mass distribution and provenance tracking as one set is no accident. The more generated images are pushed into feeds seen by hundreds of millions, the larger the authenticity and deepfake concerns grow, so watermarking is closer to a required safeguard bolted onto that spread. That said, how well an invisible watermark survives editing, re-compression, and screen capture remains an open question this announcement alone does not settle.

What practitioners should check now

The first thing to check is which access paths are open. Muse Image is available in Instagram Stories in the US only and in WhatsApp in a limited set of countries, so whether users outside those surfaces can use it through official channels has to be checked individually in the Meta AI app and at meta.ai. Country-by-country rollout creates gaps in access timing. Performance rankings, too, should be validated on your own work rather than taken at face value. Arena #2 is a human-preference Elo ranking, not an absolute metric of chart accuracy or text rendering, so whether code execution and web search actually cut errors in real deliverables as advertised has to be re-measured with your own prompts. If you plan to use Muse output in content, deciding in advance how Content Seal watermarking and labeling policy mesh with your own source-disclosure rules is the practical task to take from this launch.

Source: Meta official announcement (2026-07-07, "Introducing Muse Image and Muse Video"); reporting by TechCrunch.