Muse Image by Meta: AI Image Generator & Editor

Meta Superintelligence Labs

Muse Image, tested against Meta's own claims

Muse Image is the first image model from Meta Superintelligence Labs, released in July 2026 alongside the Muse Spark language model and a preview of Muse Video. Meta says it follows instructions faithfully, edits with precision, composes from many references and renders fine detail such as text, plots and QR codes. We ran each claim on Voor. Pick a test to see the brief, the output and our verdict.

Muse Image how-to card titled REPOT A SUCCULENT IN 4 STEPS with four illustrated, captioned panels — Text to image · 4:5 · 1408×1760 · 4 credits

Meta's claimFollows instructions faithfully; text renders cleanly

Claim holds

Every instruction met. Title and all four captions are letter-perfect.

Asked for	Got
Title “REPOT A SUCCULENT IN 4 STEPS”	Exact
Four captions, numbered 1–4	All four exact, in order
2 by 2 grid, one illustration each	Yes, and each picture matches its step
Terracotta, sage and charcoal on cream	Yes
Nothing else asked for	Added a “7 DAYS” calendar to step 4

Read the prompt

A clean illustrated how-to card on cream paper. Title at the top: “REPOT A SUCCULENT IN 4 STEPS”. Below it, four numbered panels in a 2 by 2 grid, each with a small flat illustration and one caption… Terracotta, sage green and charcoal palette.

Meta credits the plot and QR skills partly to code execution learned in training. Whatever the model does inside Meta's own apps, the chart and codes above are what a plain prompt returned here. Use Muse Image for the layout and the words, and create anything that must be measured or scanned with a dedicated tool.

Multi-reference edits

Compose a Muse Image scene from separate photos

The edit route takes up to ten reference images in one request. Number them in the prompt and give each a job: which image is the scene, which supplies an object, which supplies a material. We made the three references with Muse Image itself, then asked for a single composite. Pick a reference to see what carried over.

Muse Image composite: the speckled mug steaming on the bistro table and the mustard scarf draped over the chair — Image 1 · sceneRailing, basil and rosemary pots, table, chair, rooftops and morning haze all stay in place; the camera does not move.

“Use image 1 as the scene. Place the mug from image 2 on the bistro table with steaming coffee inside, and drape the mustard scarf from image 3 over the back of the chair. Keep the balcony, plants, morning light and camera angle from image 1.”Edit · three references · 3:2 · 4 credits

Single-photo edits

Restyle a room with Muse Image, keep the walls

Meta shows Muse Image redesigning rooms from a phone photo. Our test asked for five furniture changes in one pass, then listed what had to survive. Drag across the photo and check both lists.

Changed

Sofa reupholstered in olive-green velvet
Low walnut sideboard added on the right
Woven rug, framed abstract print, fiddle-leaf fig

Held still

Window, trees and neighbouring houses
White walls, oak floor, skirting boards
Camera height and angle, daylight direction
The floor lamp, which the prompt never mentioned

For one careful change at a time, the image to image AI guide shows how to keep an edit narrow.

After the Muse Image edit: olive velvet sofa, walnut sideboard, woven rug, abstract print and fiddle-leaf fig

Before: a plain white living room with a grey sofa and a floor lamp

Canvas

Every Muse Image shape, one pixel budget

Muse Image renders each ratio at roughly 2.5 megapixels. Voor offers eleven presets from 21:9 to 9:21, plus auto. Wide and tall frames trade height for width instead of losing detail — compare the outlines below, drawn at equal area from the files we received.

Muse Image 21:9 panorama of a fishing village with lit wooden houses and snowy mountains at blue hour — 21:9 · 2352×1008 · “Panoramic landscape photograph of a small fishing village along a curved harbour at blue hour…”

21:92352×1008 · 2.37 MP
3:21920×1280 · 2.46 MP
4:31792×1344 · 2.41 MP
1:11600×1600 · 2.56 MP
4:51408×1760 · 2.48 MP

Writing for Muse Image

Make every Muse Image instruction countable

01
Quote every word
Titles and captions go in quotation marks, in reading order. All five strings on the repotting card came back exact.
Title at the top: “REPOT A SUCCULENT IN 4 STEPS”.
02
Give structure a number
State the panel count, the grid and the axis range. Muse Image honours structure; check any data it draws against your source.
Four numbered panels in a 2 by 2 grid.
03
Give each reference a job
Number the images and say what each one supplies. Name the details of an object that must not change.
Use image 1 as the scene. Place the mug from image 2 on the bistro table.
04
List what must survive
End an edit with the things that stay. Muse Image kept every item we named and changed only what we asked for.
Keep the window, walls, floor, camera angle and daylight exactly as they are.

4 credits: per image
4 credits: per edit, with 1 to 10 references
≈2.5 MP: WebP output at every ratio
15–35 s: per image in our runs

Muse Video, Meta's companion model, is still listed as coming soon; the Muse Video page tracks what is known. For packaging lettering with a premium finishing route, compare MAI Image 2.5 from Microsoft, or browse every model in text to image.

Muse Image FAQ

What is Muse Image?

Muse Image is Meta's image generation and editing model, the first image model from Meta Superintelligence Labs, released in July 2026. It creates pictures from detailed prompts and edits or combines up to ten reference images. Meta highlights instruction following, clean text and multi-reference composition.

Is Muse Image on Voor the same as the Meta AI app?

It is the same model family, used directly. Meta runs Muse Image inside the Meta AI app, meta.ai, WhatsApp and Instagram Stories with presets, markup tools and a chat assistant. On Voor you work with the model itself: a prompt, an aspect ratio and optional reference images.

How many credits does Muse Image use?

4 credits per generated image and 4 credits per edit, whether the edit uses one reference or ten. The aspect ratio does not change the price. New Voor accounts receive free welcome credits, enough for several Muse Image tests.

How many reference images can Muse Image combine?

Between one and ten per edit. Number them in your prompt and say what each contributes, for example the scene, an object or a fabric. In our three-reference test Muse Image kept the mug's glaze and the scarf's cable pattern while placing both into the balcony photo.

Can Muse Image make charts and infographics?

It is very good at the layout and the words: titles, captions, axis labels and numbered panels came back exact in our tests. The data geometry is not reliable. Our bar chart printed the right values but drew every bar too tall, so build real charts in a charting tool.

Does Muse Image generate QR codes that scan?

Not in our tests. Meta lists QR codes among the model's strengths, but none of the three codes we generated on Voor decoded with a standard reader, including a flat code on white. Generate the QR code with a dedicated tool and place it into the design.

What resolution and aspect ratios does Muse Image support?

Every output is roughly 2.5 megapixels: 1600×1600 for a square, 1920×1280 at 3:2 and 2352×1008 at 21:9. Voor offers eleven presets between 21:9 and 9:21, plus auto, which lets the model choose the shape for you.

What file format does Muse Image return?

Results are saved as WebP images. Current browsers and most design tools open WebP directly; convert to PNG or JPEG if an upload form or older app in your workflow needs it.

How do I write a good Muse Image prompt?

Make each instruction countable. Quote every word you want printed, state numbers such as panel counts or grid size, give each reference image a job, and end an edit with a list of what must stay the same. Long, specific prompts suit this model better than short style tags.