Meta Muse Image: Ten References, One Clean Edit

Everybody has a folder. A product shot from the studio, a mood photo from a trip, a jacket a client loves, a font sample from a poster, a background someone pinned in Slack six weeks ago. The picture you actually need is all of those at once, and no single prompt has ever been able to hold that much.

Meta Muse Image is built for exactly that folder. It's the first image model from Meta Superintelligence Labs, it composes one image from many references, and when you ask for an edit it changes the thing you named and leaves the rest of the frame alone. It's live on each::labs now, as a text to image model and as an edit model that takes up to ten reference images in a single call.

That last number is the reason to pay attention. Most of the pain in image work isn't making something pretty. It's keeping six things true at the same time.

What Meta Muse Image Actually Is

Meta introduced Muse Image alongside a preview of Muse Video in a July 2026 post on its AI blog, calling it its most advanced image generation model. The pitch has three parts: it follows instructions, it edits with precision, and it composes from multiple references. Meta also ranked it second on Arena for text to image, single-image editing and multi-image editing, based on human-preference votes as of July 5, 2026. That's a snapshot, not a crown, but it says the model plays in the top tier on all three jobs rather than being a generator with an edit mode bolted on.

Inside Meta's own apps the model behaves more like an agent than a renderer. Meta describes it calling search to ground images in real-world facts, writing and running code for accurate charts and QR codes, and refining its own output with local edits or full regenerations. Meta says that self-correcting habit emerged during training simply because it produced better images. Those tool loops are a product feature of the Meta AI app. What you get on each::labs is the model behind them, exposed as two clean endpoints you can wire into your own pipeline.

Two Endpoints, One Model

On each::labs, Muse Image comes as Meta Muse Image Text to Image and Meta Muse Image Edit. They share the same controls, which keeps switching between them painless.

Text to image takes a prompt and, optionally, an aspect ratio. You get nine to pick from, running from an ultra-wide 21:9 through 16:9, 4:3, 3:2, square, 2:3, 3:4 and 9:16 down to a tall 9:21. Leave it empty and the model picks dimensions from the prompt itself, which is handy when you're exploring and don't know yet whether the idea is a banner or a story. A single request can return anywhere from one to ten images, in WebP, PNG or JPEG.

Edit adds an image list. Give it one picture and it's an editor. Give it several and it becomes a compositor that pulls a person from one, a jacket from another, a room from a third. The ceiling is ten references per request, which is more than most real briefs ever need.

Compose From References, Don't Describe Them

Here's the habit to break. When you have a reference, stop describing it in words. A paragraph about "a mid-century walnut armchair with tapered legs and mustard boucle upholstery" is a worse version of the photo you already have. Meta built Muse Image to combine people, objects, clothing, styles and environments from many reference images, and it supports text and images interleaved in the prompt. Use that. Put the chair in the image list and spend your words on what the model can't see: where the chair goes, how the light falls, what the picture is for.

A good multi-reference prompt reads like stage directions, not an inventory:

Place the armchair from the first image in the living room from the second image, angled toward the window. The woman from the third image sits in it, wearing the coat from the fourth image, reading. Late afternoon light from the left, soft shadows, editorial interior photography, 3:2.

Refer to each image by its position. "The first image" and "the coat from the fourth image" are unambiguous in a way that "the chair" isn't, especially once three of your references contain furniture.

Edit One Thing and Name What Stays

Meta's own claim for editing is simple: Muse Image changes only what you request and stays coherent across editing turns. That second half matters more than it sounds. Plenty of models can recolor a sofa once. Fewer can recolor it, then swap the rug, then relight the room, and still hand you the same room at the end.

You'll get the most out of that behaviour by being explicit in both directions. Say what changes. Then say what doesn't.

Change the sofa fabric to deep green velvet. Keep the room, the lighting, the camera angle and every other object exactly as they are.

It feels redundant. It isn't. The "keep" clause takes one line to write, and it turns a vague request into a contract the model can check its work against. For a chain of edits, make one change per turn. If you stack four requests into one prompt and the result goes sideways, you've no idea which one caused it.

Where Meta Muse Image Earns Its Place

The model's sweet spot is any job where the deliverable has to match things that already exist.

Ecommerce teams have a product shot and a dozen settings it needs to live in. Feed the packshot plus a scene reference into Edit and ask for the product placed naturally, at the right scale, with matching shadows. Then reuse the same packshot across every scene, which is where a ten-reference ceiling starts paying off.

Fashion and lookbook work is the classic multi-reference problem: this model, this jacket, these trousers, this location. Muse Image was built to combine people and clothing from separate references, so the brief maps straight onto the input list.

Social and brand teams can use the text to image side to pull ten images from one request and get a spread of directions fast, then move the winner into Edit to tighten it. Meta itself points to marketing assets for small businesses as one of the uses it had in mind, which tells you the model was tuned for images that have to do a job.

Interior and real-estate visuals benefit from the "change one thing" discipline: restyle a room shot without moving the walls, swap furniture from a catalog photo, try three palettes on the same space.

How to Prompt Meta Muse Image for Clean Results

Lead with the outcome, then the inputs. "A hero banner for a winter coat launch" tells the model what kind of image it's making before any detail lands, and every later choice bends toward that purpose.

Use the aspect ratio as a creative decision, not an afterthought. A 21:9 frame invites a landscape with negative space for copy. A 9:21 frame forces a single vertical subject. If you're not sure, leave it empty for a few runs and see what shape the model chooses for your prompt. That's useful information on its own.

For text inside the image, quote the exact words and say where they go:

A minimalist poster for a jazz night. Large serif headline reading "Late Set" across the top third, small caps line "Friday, 10 PM" beneath it, a single brass trumpet lying on a deep blue velvet cloth below. 2:3.

Batch while you explore, single while you refine. Ten images on the first pass, one at a time once you've picked a direction and you're making precise edits.

When you need a lossless file for compositing or print, switch the output to PNG. WebP is the default and it's perfect for previews and web delivery, but it isn't what you want feeding a retouching step.

Muse Image Next to the Edit Models You Already Use

Muse Image arrives in a crowded, very good field. GPT Image 2.5 splits into a quality model and a speed model. Grok Imagine 2.0 is built around precision edits and text. Nano Banana 2.1 pushes extreme aspect ratios and multi-turn consistency.

The case for Muse Image is the combination: a top-tier ranking across generation, single edits and multi-image edits at the same time, plus a generous reference count. If your job is mostly one image in, one image out, you have several excellent options and you should test two of them on your own brief. If your job is "these seven things, in one picture, believably", Muse Image belongs on the shortlist. All of them sit side by side in the text to image and image to image catalogs, so comparing them on the same prompt is a model-id change, not a new integration.

Running Meta Muse Image on each::labs

The workflow is short. Open Muse Image Edit, add your reference URLs in the order you'll refer to them, write the prompt with positions ("the first image", "the third image"), choose an aspect ratio or let the model decide, and generate. For pure generation, use Muse Image Text to Image with the same prompt habits.

Because both endpoints live behind the same each::labs API as the rest of the catalog, it's easy to chain them. Generate ten options, pick one, send it to Edit with a product reference, then hand the final frame to an image to video model for motion. If you'd rather not wire that yourself, browse each::labs flows and swap Muse Image into the image step.

Frequently Asked Questions

Who made Meta Muse Image?

Meta Superintelligence Labs. Meta introduced it in July 2026 as the lab's first media generation model, together with a preview of Muse Video, which shares the same pretraining base and adds native audio.

How many reference images can Meta Muse Image use?

Up to ten per request on each::labs, through the Edit endpoint. One image makes it an editor; two or more turn it into a compositor that pulls people, objects, clothing and settings from different sources into one picture.

Can Muse Image keep the rest of my image unchanged when I edit?

That's the behaviour Meta designed it for: change only what was requested and stay coherent across several editing turns. You'll get the most reliable results by naming what must stay the same in the prompt and making one change per turn.

Is Muse Video available too?

Not yet. Meta has shown it as a preview and says it's coming to creators and Meta AI. For video today, pair a Muse Image frame with one of the image to video models on each::labs.