Meta Muse: Image and Video Generation, In-House

Meta Muse arrived on July 7, 2026, and it changes who owns the model behind “Meta AI” images. Meta Superintelligence Labs (MSL) shipped Muse Image, its first in-house image generation model, and previewed Muse Video — ending years in which Meta licensed outside models like Midjourney and Black Forest Labs to power the picture-making inside its apps.

The headline number everyone repeated was “#2 on Arena.” The more interesting story sits underneath it: what Meta discloses, what the Arena leaderboard discloses, where the models are actually free to use today, and a consent design that turned into a talking point within hours of launch. For brands and social teams, those details matter more than the rank.

This guide walks through what launched, the ranking claims from both Meta and Arena side by side, how the model generates images, exactly where you can and cannot use it right now, the Advantage+ advertising angle, and the privacy mechanic every social manager should understand before a brand account gets remixed. Every fact below is drawn from Meta’s own announcements and independent coverage from the launch.

Meta Superintelligence Labs, the group led by Chief AI Officer Alexandr Wang, released Muse Image as its first artificial intelligence model for image creation, alongside a preview of Muse Video. The image model was reportedly code-named “Mango” internally. The significance is less about a single benchmark and more about ownership: Meta previously relied on third-party generative models — Midjourney and Black Forest Labs among them — to power the image and video features inside the Meta AI app. Muse marks the first time that entire stack is built and controlled in-house.

That shift changes the vendor-risk calculus for any brand building a workflow on top of “Meta AI” outputs. The roadmap, the usage policy, the pricing, and the eventual API surface are now Meta’s to set, not a licensing partner’s. It also marks a clear escalation of Meta’s in-house generation push — the company is no longer renting the models behind its most visible consumer AI features, it is building them, as the next section explains.

Most coverage led with “Muse Image is #2 on Arena” and stopped there. Here is the nuance that got buried: Meta states the #2 rank across text-to-image, single-image editing and multi-image editing — measured by human-preference Elo, as of early July per Meta — but its own blog post never names the #1 model that Muse Image sits behind. If you see Muse Image’s #1 rival named as a Meta-confirmed fact, that is a misread of the source.

Separately, Arena’s own account disclosed the specific Elo numbers Meta withheld. Those scores — and the gap they imply between Muse Image and whatever sits at #1 — come from Arena, not from Meta, and Elo is a live, community-vote metric that drifts as more votes land, so read the table as a point-in-time snapshot rather than a fixed standing. We deliberately do not name the #1 model, because Meta does not, and its identity only surfaces in unofficial third-party leaderboard posts. The “gap vs #1” column is our own arithmetic on Arena’s published Elo figures.

Read the gaps carefully. In images, Muse Image trails the leader by roughly a hundred Elo points on Arena’s snapshot — a real but not enormous margin in a metric where a few dozen points separate near-peers. In video, the three top models sit within about 68 Elo of each other, which is close enough that ordering can flip as vote counts grow. The honest framing for a brand deck is “competitive top-tier, not the outright leader,” with the caveat that these standings are Arena’s and will move.

If you are sizing up how this fits the wider field, our data-led reads on the broader AI image-generation market and AI video generation adoption data put Muse’s arrival in context. For a hands-on comparison of the tools it competes with, see the current AI video generator landscape.

The most distinctive thing about Muse Image is that it behaves less like a single-pass diffusion model and more like an agent that plans before it draws. Meta describes it as agentic: the model reasons through a prompt, searches the web, and plans before generating. In practice that means it can invoke tools and refine its own output rather than committing to a first guess.

The editing surface is where this pays off for everyday use. Muse Image supports multi-reference composition — blending several input photos into one image — and targeted edits to existing shots, such as removing a photobomber, changing the camera angle, or erasing fog. It is the difference between “generate something new” and “fix the photo I already have,” and the second is often what a social team actually needs.

The single most useful thing a brand can know on launch day is what actually works right now versus what Meta has only announced. The two are easy to conflate in the coverage. Muse Image is genuinely live and free across several consumer surfaces; Muse Video and the advertiser integration are not. Here is the clean split.

The practical takeaway for a social team: you can start testing Muse Image inside the Meta AI app and Instagram Stories today, but do not build a campaign timeline that assumes Muse Video or Advantage+ access exists yet. Both are real commitments from Meta, but they are dates to watch, not tools to deploy. If your team is standing up an AI-assisted content pipeline, our social media management work builds exactly these workflows around what is actually shippable.

Muse Image is free for consumers, but the commercial logic points straight at advertising. Meta says Muse Image is coming to Advantage+ creative in the weeks ahead, where it will let advertisers generate lifestyle visuals, new product backgrounds, and static creatives pulled from video — all while aiming to preserve brand identity. That positions an in-house image model directly inside Meta’s highest-value surface.

For scale, the commercial stakes sit with where Muse plugs in. Meta’s automated Advantage+ suite is the revenue engine these models are being pointed at, and dropping a first-party image model into it means creative generation, ad targeting, and delivery increasingly live under one roof. That is the real reason a “free” consumer image model matters to Meta’s business — it feeds the highest-value surface the company operates, with no licensing partner in the middle. Any performance lift Meta attributes to Muse-powered creative is, for now, vendor-stated and unbenchmarked, so treat it as a claim to test rather than a result to bank.

The strategic read: Meta is following the same playbook as its rivals, folding a first-party generation model into the ad stack so creative production, targeting, and delivery live under one roof. That is convenient, and it is also a concentration of control worth naming. This builds on the direction we covered in Meta’s AI creative tools for advertisers. For teams weighing how much of their creative to route through it, our paid media management practice treats platform-native AI as one input to test, not a default to adopt wholesale.

Within hours of launch, attention turned to one design choice. Muse Image includes an @-mention feature that lets a user tag any public Instagram account and pull that account’s public photos into a generated image — without notifying the tagged person. The feature defaults to opt-out: it is on unless a user goes into settings and disables it. Meta’s stated policy is blunt on notification.

Most coverage framed this as a consumer-privacy story. It is also a brand story. A public brand account is exactly the kind of source a stranger could pull into a generation, and the opt-out default means the safe posture — turning the setting off — requires a deliberate action most teams will not know to take. The same-day pushback was immediate, with users describing the consent gap in stark terms.

The concrete action for social teams is small and worth doing now: review the AI-generation and tagging settings on every brand Instagram account you manage, decide whether to opt out, and document the decision. This is a governance question, not a technical one, and it is the kind of thing our social media team folds into account setup and brand-safety reviews.

On the provenance side, Muse Image embeds Content Seal — an invisible watermark designed to survive cropping, compression, resizing, and even screenshots — and Meta offers a public detection tool at meta.ai/identification to check whether an image carries it. That is a genuinely useful control for anyone who needs to know if a visual came out of Meta’s model.

The gap worth flagging: video watermarking is planned but not yet shipped. So while Muse Image outputs can be identified, Muse Video outputs cannot yet be checked the same way — one more reason the video model’s preview status matters. If provenance and disclosure are part of your compliance posture, image is covered today and video is a wait-and-see.

Muse changes the practical checklist for social and paid teams in a few specific ways. For everything else, the picture is unchanged. Here is how we would triage it.

The through-line is discipline about the live-versus-announced line. Muse Image is a real, free, capable tool you can use today with appropriate guardrails; almost everything else Meta announced is a date to watch. Teams that treat the two the same way will either miss a genuine opportunity or over-promise on capabilities that do not exist yet.