News · 2026-07-07

Meta unveils Muse Image and Muse Video, its first in-house media generators

Meta Superintelligence Labs previewed Muse Image and Muse Video on July 7, 2026, its first media-generation models built fully in-house. Meta calls Muse Image its "most advanced image generation model yet," and - the detail that matters most for builders - it embeds "agentic tool use capabilities" directly into image generation instead of treating a picture as the one-shot output of a single prompt.

Key facts

• Two models: Muse Image (create and edit images) and Muse Video (video with native audio), both previewed by the official @AIatMeta account.
• Muse Image is pitched on multi-reference composition, faithful instruction following, and agentic tool use; it integrates with Meta's newly announced Muse Spark model.
• Muse Video is trained on the same pretraining base as Muse Image and ships with native audio rather than a separate sound pipeline.
• Announced July 7, 2026 by Meta Superintelligence Labs, the reorganized AI group Meta stood up in 2025.

For two years, the image-generation race has mostly been a quality race: sharper faces, better text rendering, fewer mangled hands. Meta is entering that race late with its own stack, but the framing of the announcement suggests it is trying to change the game rather than just match it. The headline capability is not resolution - it is that Muse Image can act like an agent while it draws.

Here is what that means in plain terms. A normal image model takes your words, runs them through the network once, and hands back a picture. If the result is wrong - the logo is in the wrong corner, the product shot doesn't match your reference photo - you rewrite the prompt and roll the dice again. "Agentic tool use" flips that loop inward. The model can take intermediate steps: pull in reference images, call helper tools, check its own work against your instruction, and revise, all before it shows you a final frame. Think of the difference between handing a brief to an artist who mails back one drawing versus one who sketches, compares it to your references, and fixes it before delivery. Meta says Muse Image is strong at "multi-reference composition" - stitching several source images into one coherent scene - which is exactly the kind of task that benefits from that step-by-step approach.

Muse Video is the more ambitious piece. Meta built it on the same pretraining foundation as Muse Image, so the two share a visual understanding of the world, and it generates sound as part of the model rather than as an afterthought. Most video generators today produce silent clips that get a separately generated audio track glued on, which is why the footsteps never quite line up with the feet. Native audio means the model learns picture and sound together, so a door slam or a spoken line is meant to be synchronized from the start.

Why does this matter beyond Meta's product roadmap? Because it signals where the whole field is heading. Media generation is converging with the agent paradigm that has dominated coding and web automation all year. Instead of a passive "prompt in, picture out" box, the image model becomes a small worker that plans, uses tools, and iterates - the same shift we have watched turn chatbots into coding agents. For the practitioners who build on these systems, an image model that can call tools is a different integration surface entirely: it can fit into an automated pipeline as an actor, not just a renderer.

There is a real caveat: this is a preview, and previews are marketing. Meta showed capabilities and framing, not an independent benchmark, a public model card with hard numbers, or - at announcement time - general availability. "Most advanced yet" is Meta grading its own homework, and the company is playing catch-up in a field where OpenAI, Google, and a wall of open-weight Chinese models already ship strong image and video systems. The interesting claim - agentic tool use inside generation - is also the hardest to evaluate from a launch tweet, because its value only shows up when developers can wire it into real workflows and see whether the extra steps actually produce more faithful images or just more latency. Until Muse Image and Muse Video are in enough hands to test against the incumbents, the honest read is that Meta has staked out an interesting position, not that it has won it.

Primary source, verified: read the paper →

Ground Truth. (2026, July 7). Meta unveils Muse Image and Muse Video, its first in-house media generators. Ground Truth. https://groundtruth.day/news/meta-previews-muse-image-and-muse-video.html

@misc{groundtruth:meta-previews-muse-image-and-muse-video,
  title  = {Meta unveils Muse Image and Muse Video, its first in-house media generators},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/meta-previews-muse-image-and-muse-video.html}
}

Topics: meta · image-generation · video-generation · product-launch · agents