Muse Image: Meta's image model that searches, writes code and refines its own work

What is Muse Image

Muse Image is Meta’s image generation model. It was built by Meta Superintelligence Labs, the lab led by Alexandr Wang, and it’s the company’s first in-house image model: before it, Meta AI drew with third-party models. Meta introduced Muse Image on July 7, 2026.

The key difference is that the model works as an agent: it thinks the task through before drawing. It can search the web, write and run code, look at the result and refine it. According to Meta, self-refinement wasn’t designed in: it emerged on its own during reinforcement learning because it produced better images. Quality rises roughly logarithmically with reasoning strength.

In July, Meta reported second place on Arena in generation, single-image editing and multi-image editing. By October 26, 2026, after GPT Image 2.5 and other models arrived, Muse Image ranks 9th of 82 in generation.

For text, Meta has a separate language model — see the Muse Wonder news.

What Muse Image can do

Search before generating. Web and image search are built into the model and on by default. That helps it depict current events and real objects more accurately.

Code for precise graphics. Muse Image builds charts and QR codes with code — so the numbers on a chart are right and the QR code scans. Together with the Muse Spark text model, it makes animated GIFs, websites with images and simple games.

Text and instructions. Meta highlights legible styled text, infographics and step-by-step guides.

Multiple references. The model assembles a scene from people, objects, clothing, styles and places taken from different photos, and images can be interleaved with text in the prompt. The API accepts 1 to 10 images per edit.

Multi-turn editing. The model stays consistent across several edits in a row. In Meta AI you can circle or sketch right on the image to show what to change — for example, to remove a photobomber from the background.

Room redesign. Snap a photo of a room, and Meta AI redesigns it with real products from the web or Facebook Marketplace.

Presets. Meta AI has one-tap prompts: restoring old photos, trying trending hairstyles, claymation or 16-bit styles. You can share a preset with friends.

Muse Image in the API

Parameter	Value
Model ID	muse-image-1.0
Endpoints	/v1/images/generations, /v1/images/edits (OpenAI-compatible)
Images per edit	1–10
Images per request	up to 10
Reasoning strength	high (default, several passes) or low (single pass)
Formats	webp, png, jpeg
Price	$0.01 per image, search included
Rate limit	150 requests per minute

You only pay for successfully generated images: failed ones and those removed by safety filtering aren’t billed.

What Muse Image is good for

• Generation at scale — ad variants, catalog photos, personalization: at $0.01 per image it stays cheap even across thousands of images.
• Infographics and instructions with text.
• QR codes and charts with accurate data.
• Interior redesign from a photo.
• Assembling a scene from several photos — a person, a pet, clothing and a place.
• Restoring old photos and trying new hairstyles — via presets.

Muse Image vs alternatives

	Muse Image	GPT Image 2.5	Grok Imagine Image 2.0	Nano Banana 2
Developer	Meta	OpenAI	xAI	Google DeepMind
Arena rank (Oct 26, 2026)	9	1 (Sunburst), 2 (Flare)	6	11
Web search	built in	via reasoning models	—	yes
Input images	up to 10	several	up to 5	up to 14
Transparent background	no	yes	yes	—
API price per image	$0.01	per token	$0.04	from $0.045
In AI2Face	yes	yes	yes	yes

Muse Image is the cheapest model in Arena’s top 10: 4 times cheaper than Grok and 4.5 times cheaper than Nano Banana 2. On quality, GPT Image 2.5 and Grok Image 2.0 are ahead.

How to write prompts for Muse Image

• Describe the task, not just the picture. The model will find missing facts itself: “an infographic with current London Underground fares.”
• Give numbers for charts — the model plots them with code.
• For a QR code, give the exact link and the style around it.
• For edits from several photos, say what to take from which: “the person from the first photo, the dog from the second, the pier from the third.”
• For speed, set reasoning_strength: low in the API — one pass, same price.

How to use Muse Image in AI2Face

Muse Image is in the AI2Face generator next to GPT Image 2.5, Grok Image 2 and Nano Banana Pro.

1. Sign up — new accounts get starter credits.
2. Open the generator and pick Muse Image.
3. Write a prompt or upload photos to edit.
4. The credit cost is shown before you start. Credits don’t expire, and you don’t need a Meta account.

Limitations

• No transparent background — the API always returns an opaque image.
• No mask — edits are prompt-only, with no selected area.
• Size sets the proportions only. The model doesn’t guarantee exact pixel dimensions, and the aspect ratio can’t exceed 16:1.
• Invisible watermark. Images from Meta AI and meta.ai carry Content Seal, which survives cropping, compression and screenshots.
• Availability. Instagram effects are US-only for now, and WhatsApp only covers some countries.
• The @-mention episode. At launch you could generate images of public Instagram accounts by @-mentioning them; on July 10, Meta removed the feature after criticism.

Bottom line

Muse Image is Meta’s first in-house image model and the cheapest in Arena’s top 10: $0.01 per image with search and code built in. Its agentic approach pays off where facts matter — infographics, charts, working QR codes — and where you need thousands of variants. If top quality matters more, use GPT Image 2.5; for editing a selected area or a transparent background, Grok Imagine Image 2.0. You can compare them on the same prompt in the AI2Face generator.