Meta launches Muse Image: it looks things up and fixes its own mistakes while generating images, ranking No. 2 worldwide
Sibling model Muse Video was previewed at the same time, already supports native audio generation, and will soon open to creators.
- Meta Superintelligence Labs released its first batch of self-built media generation models since the lab was founded: Muse Image, now live, and Muse Video, released as a preview.
- When generating an image, Muse Image calls search and code tools to verify content and check details, turning image generation into a process of researching while drawing.
- The model checks and revises its own drafts mid-generation, a self-correcting habit that emerged spontaneously during reinforcement learning training — the team never designed it on purpose.
- On the Arena human-preference leaderboard (as of July 5, 2026), Muse Image ranks No. 2 in all three of text-to-image, single-image editing, and multi-image editing; Muse Video ranks No. 3 in text-to-video.
- Muse Image is now live in the Meta AI App, meta.ai, US Instagram Stories, and WhatsApp in some countries, with Facebook support coming soon; Muse Video will soon open to creators.
Meta steps in directly with two self-built image and video generation models
On July 7, 2026, Meta Superintelligence Labs released Muse Image and Muse Video — the lab's first self-built media generation models since it was founded.
Muse Image is now officially live, while Muse Video is out as an early preview. The former generates images, the latter generates video, and both share the same pretrained foundation.
Worth noting: when Muse Image generates an image, it works step by step like an assistant that looks things up, writes code, and revises its own drafts — and on the Arena human-preference leaderboard, it ranks No. 2 in text-to-image, single-image editing, and multi-image editing alike.
Some of Muse Image's official showcase results. Source: Meta AI Blog.
Turns out AI image generation can search the web, write code, and draw — all at once
Older image models mapped your one sentence directly into a picture, with no verification step in between. Muse Image turns generation into an agentic process (one where it calls tools on its own and works through the task step by step): it calls search and code tools first, checking and calculating along the way, before assembling the final image.
Writing code. During reinforcement learning, Muse Image learned to write and run code to generate accurate charts and QR codes, then correct the image based on the rendered result. To draw a scannable QR code, it actually computes the QR code with code, opens the image to confirm it scans, then places it into the scene.
Web search. It also learned to search, aligning generated images with real, live information. With search turned on, prompts that depend on current events and real-world facts come out more factually accurate.
Muse Image can also work alongside Muse Spark — the two models share tools and plan together, combining code and media generation to make animations, image-rich web pages, and playable visual mini-games.
An example: drawing a poster with a QR code that actually scans
The scenario in the original post: a Korean-comic-style young woman standing in front of an ICML 2025 poster, scanning a QR code that links to meta.ai. Here's how the model worked through it step by step:
Korean comic style · ICML 2025 poster
linking to meta.ai
confirm it actually scans
look down at their phone
Left: initial version, the person is staring at the QR code; right: after the model fine-tuned it, the person now looks down at their phone screen, while the art style, poster, and QR code details all stay unchanged. Source: Meta AI Blog.
Search on vs. off — what changes
For prompts that rely on current events or real-world facts, the model can only draw on what it already remembers, which makes mistakes more likely.
It looks things up first and draws afterward, giving higher factual accuracy (in Meta's internal ablation study, the search-on version had a higher win rate).
More agentic examples (click to expand)
- Fractal poster: first computes a Julia set and a Sierpinski triangle in Python, composites a clean base image, then applies a mid-century Swiss grid layout.
- Fighting-game flipbook animation: generates punch and dodge frames one by one, keeping lighting and character design consistent throughout.
- Pet-raising mini-game: generates six images of a cat at kitten, teen, and senior stages, converts them to base64 and embeds them directly in HTML, delivering a webpage that's playable the moment it opens with no external files needed.
- 2026 summer outfit: first searches fashion trends and product catalogs, then produces an outfit image you can shop directly.
- Moon-formation infographic: first searches scientific diagrams and facts about the giant-impact hypothesis, then draws a six-panel vertical infographic.
- Redecorating a room with secondhand furniture: based on your uploaded room photo and city, searches Facebook Marketplace for suitable used furniture, then produces a rendering.
This "self-correcting" habit is something the AI figured out on its own during training
Muse Image reviews and improves its own work within its own chain of thoughtThe step-by-step reasoning the model writes out before giving its final result — like scratch notes on paper.. This self-correction takes three forms:
If some small detail is wrong, it makes a local edit to the current draft.
If a large area is wrong, it just starts over and generates a fresh image.
Switches to tools instead — searching or writing code — to get the facts right.
The team didn't design this behavior. It emerged on its own during reinforcement learning training: the model discovered that revising its draft produced better images, which earned a higher reward, so it learned this whole sequence of actions by itself.
An example: it caught its own broken formula in a magazine layout
The model was assembling an elegant magazine page — after laying out a mathematical proof, a headline, and a portrait, it looked back, noticed the summation formula was missing a division sign, and corrected it to:
Left: first draft, the formula under the headline reads "S = n(n + 1) 2" — missing the division sign and unreadable; right: the model noticed the problem on its own review and changed it to "S = n(n + 1) / 2," while also re-checking the layout details along the way. Source: Meta AI Blog.
Self-correction on vs. off — what changes
The model turns in whatever it generates on the first pass — small flaws in the draft stay in the final image.
It reviews and revises its own draft during generation, producing higher-quality images (in Meta's internal ablation study, the self-correction-on version had a higher win rate).
The longer the AI thinks, the more accurate its image — but the gains eventually level off
Like language models, Muse Image gets better the more it thinks before generating an image. Given more test-time computeThe extra computation the model spends before producing its actual result: more reasoning, more tool calls, more rounds of draft revision., it reasons more, calls tools more, and self-corrects more.
Test-time compute is like drafting a few more times and checking your exam paper a few more times before turning it in. Thinking longer usually makes the answer more accurate, but check too many times and the extra benefit from each additional pass keeps shrinking.
Meta observed that as reasoning intensity is turned up, the human-preference Elo score rises along with it, in a roughly log-linear relationship. Interestingly, this compute spans two very different kinds of work — reasoning uses text tokens, generation uses image tokens — but final quality depends on the combined total compute across both.
How you spend that compute matters too. Best-of-N (generating many images at once and picking the best one) improves quickly at first but plateaus fast. Spend the same compute on deliberate reasoning instead, and quality keeps climbing; add tools on top of reasoning, and the effect compounds further, because tools let the model reach things it wouldn't otherwise know — like looking up a missing reference, or writing code to nail down a detail.
Editing images is now a matter of pointing at exactly what you want changed
Muse Image only touches the part you name in your edit, leaving everything else untouched. The original post gives a set of instructions, each one very specific:
Even the text on a sign can be swapped precisely
One instruction was to change the sign to read "$3.00 ALL DAY," change "no free parking" to "FREE PARKING ON WEEKENDS," and change the phone number to 555-5555. Only the text changes — the rest of the layout stays put:
555-1234
555-5555
Illustration recreated from the sign-text edit instructions in the original post, showing "pointing at exactly what you want changed" as a precise replacement.
Multiple rounds of edits still stay consistent
Muse Image supports editing round after round without drifting off track. The original post includes a conversation that goes from start to finish, with each step building on the previous image:
The same cat, dog, and café elements stay consistent across five rounds of edits.
The original post has another set of similar multi-turn examples: "turn the living room into Japandi style → but bring back the lamp and cabinet from the first image → finally make a before-and-after comparison image," with each step building on the output of the last:
Cram several reference images into one prompt, and the AI can still piece it together correctly
Muse Image can take elements from several reference images — a person, an object, an outfit, a style, a scene — and piece them together into one new image. The prompt can also interleave text and images. For example: put [this person] on [this bicycle], have them wear [this outfit], passing by [a park bench], all rendered in the style of [some image].
What score did human judges give it
This ranking comes from Arena. It runs large numbers of real head-to-head human votes on "which image / video is better," and computes a leaderboard score (Elo) from them — the higher the score, the more people preferred that result, similar to how a competitive game's rank tiers work. As of July 5, 2026, here's where the Muse series stands:
| Category | Model | Arena rank |
|---|---|---|
| Text-to-image | Muse Image | #2 |
| Single-image editing | Muse Image | #2 |
| Multi-image editing | Muse Image | #2 |
| Text-to-video | Muse Video | #3 |
Meta's official full leaderboard (top 10): in all three image categories, OpenAI's GPT Image 2 takes first place (1280–1466 points), with Muse Image right behind it in second; in video, first place is Google's Gemini Omni Flash (1527 points), second is ByteDance's Seedance 2.0, and Muse Video ranks third (1459 points). Higher scores mean more approval from human judges. Source: Arena AI Leaderboard, as of July 5, 2026.
Muse Video is still a preview version. Meta says it's competitive on prompt adherence, visual fidelity, and temporal coherence, and names two areas still being improved: audio-visual sync and physical accuracy of fast motion.
Who can use it now, and how do you prove an image was made by AI
Muse Image is already live today in the Meta AI App, meta.ai, US Instagram Stories, and WhatsApp in some countries, with Facebook support coming soon. Muse Video will soon open to creators and will also join Meta AI, with native audio generation support.
How to check whether an image was AI-made: Content Seal
To let people tell whether an image was AI-generated, Muse Image has Content Seal built in — an invisible watermarking system. Any image generated with Muse Image in the Meta AI App or on meta.ai carries an invisible provenance mark that survives cropping, compression, resizing, and screenshotting. Meta has also previewed a detection tool that lets you check whether an image carries a Content Seal. Watermarking for video is planned to follow soon.
What you can do once it's connected to Meta's own products
Muse Image is tied into Meta's social ecosystem, and there are already a few kinds of use cases on the ground:
We didn't design this behavior. It emerged on its own during reinforcement learning training, simply because self-correction produced better images and earned a higher reward. Meta AI Blog · Introducing Muse Image and Muse Video
AI image generation, from "one sentence in, one image out" to "look things up, fix its own mistakes, then output"
Meta launches Muse Image, turning image generation into a process of researching and revising drafts along the way; one page with visuals to see how it pulls this off.
↓ Read this page · one animated diagram inside
An AI image tool is one where you type a sentence and it gives you an image. Previously, it drew the whole thing in one go from what it "remembered," with no checking, no looking back, in between.
✘ But anything that needs to match reality — a scannable QR code, an accurate chart, real text on a sign — often comes out wrong
Because the old approach "mapped" text straight into an image in one shot, with no verification and no looking back — if it got something wrong, it wouldn't even know.
Muse Image turns image generation into an agentic process (one where it calls tools on its own and works through the task step by step): search first, compute with code, verify, then generate.
⟶ straight to one image
(no checking in between)
⟶ search / verify with code
⟶ draft ⟶ self-check
⟶ output
So how exactly does "checking and revising along the way" work? See the diagram below.
Muse Image looks back and revises its own draft mid-generation — this is called "self-correction": a small issue gets a local fix, a large area that's wrong gets scrapped and redone, or it switches approach entirely and goes to search or writes code. Crucially, this habit wasn't designed by the team — it emerged on its own during reinforcement learning: the model discovered that revising drafts scored higher, and it taught itself the whole routine. Give it more time to think, and it searches more, revises more, and draws more accurately — though the gains eventually level off.
Arena is a leaderboard computed from real head-to-head human votes (Elo score — the higher, the more people approve of its results, an algorithm similar to a competitive game's rank tiers). As of July 5, 2026, here's where the Muse video model stands on the text-to-video leaderboard:
Can you use it now? Muse Image is already live on Meta AI App, meta.ai, US Instagram Stories, and WhatsApp in some countries, with Facebook support coming soon; Muse Video will soon open to creators. These images all carry a Content Seal — an invisible watermark that survives cropping, compression, resizing, and screenshotting, with a companion tool to help you check whether an image was made by Meta AI.
but this code…
nothing
it got moving first.
② Write code to verify the code
③ Open it to confirm it scans
it's more like an assistant doing its own work
the formula yourself?
nobody taught it that.
text-to-image · single-image editing · multi-image editing
ranks No. 2 in all three
- × Scores are Meta's own tests
- × Win rates not independently reproduced
- × Muse Video is still preview-only
and this skill grew on its own during training — the team never taught it.