Runs on Muse Spark 1.3 · free tier, with limits

Three tools wearing one name.

Muse Image is Meta’s image generation and editing tool, and it quietly covers three different jobs: making something from a description, changing an image you already have, and steering new output with a picture you supply. Nearly everyone who ends up fighting it is doing one job using the mode built for another which is a much easier problem to fix than a bad prompt.

The three modes Where it falls short

3 modes 4 places approximation bites Free tier, lightly metered No separate install

Same model underneath Muse Spark 1.3 but three different ways of asking.

How it works

The three modes, and the symptom of choosing wrong

Each mode is good at something the others handle badly. The useful diagnostic is not “is my prompt good enough” but “am I in the right mode at all”.

Generate from a description, with nothing to start from

Blank page

The mode everyone knows. You describe something and get an image of it. It is at its best when the subject is common, the specifics are yours to choose, and you are happy to be surprised by the parts you did not specify.

You are in the wrong mode if: you keep adding clauses to pin down a look. Description is a poor tool for style, and a prompt that has grown to four lines is usually a reference image in disguise. Prompts that hold up.

Edit change one thing about an image you already have

Keep what works

The most underused of the three. Editing preserves everything that was already right, which is the one thing regenerating cannot promise: a fresh attempt will differ in ways you never asked for, including the parts you liked.

You are in the wrong mode if: you are regenerating an image that was 90% there, hoping the next roll keeps the good parts. It will not. Editing instead.

Direct hand it a picture and say “like this”

Show, don’t tell

Style, palette, composition and mood are things a picture communicates instantly and a sentence communicates badly. Supplying a reference collapses several rounds of describing into one step, and it is the fastest route to a consistent look across a set of images.

You are in the wrong mode if: you are trying to describe a look you can already see somewhere. Show it instead. How references steer output.

Why the wrong mode feels like a prompting problem

All three modes take text, so choosing wrongly never announces itself. You get a result, it is not right, and the obvious move is to reword. Rewording cannot fix a mode mismatch, so the loop continues until the allowance or the patience runs out.

The tell is repetition: if your third attempt is a longer version of your first, stop writing and change modes instead.

Honest limits

It gives you the neighbourhood, not the address

The failure mode of image generation is not ugliness. It is approximation output that is near what you asked for, in a way that is fine for some jobs and useless for others.

Ask for a red bicycle against a brick wall and you will get a red bicycle against a brick wall. Ask for your red bicycle, or the exact brick of a particular building, and you will get something adjacent plausible, well made, and not the thing.

That gap is invisible when the subject is generic and glaring the moment a specific detail has to be correct. Most disappointment with image generators lives exactly there.

It also explains why the tool feels excellent for some work and hopeless for other work that looks superficially similar. A background, a mood piece, an illustration of a concept all of these are happy in the neighbourhood. A product your customers will recognise, a diagram whose labels must be right, a likeness someone will check: these need the address.

None of which is a reason to avoid it. It is a reason to know, before you start, whether your job tolerates approximation because if it does not, no amount of prompting will change the answer.

A reference image is how you move the target closer, rather than hoping a dot lands on it.

Approximation bites

Words inside the picture

Text rendered into generated images is a known weak point across image generators generally. Treat any words in the output as something to check, and for anything where the wording matters, add the text yourself afterwards.

Approximation bites

A specific real thing

A particular product, building or person comes back as a convincing relative rather than the thing itself. If your audience knows what it should look like, they will notice which is why product work usually starts from a photo.

Approximation bites

Counts and arrangements

Exactly five of something, arranged in a stated order, is the kind of instruction that gets approximately followed. Check anything countable rather than assuming it came out as asked.

Approximation bites

Consistency across a set

The same character or style across several images will drift unless you anchor it. This is the clearest case for working from a reference rather than re-describing each time.

Backgrounds and texturesGeneric by design, so the neighbourhood is the destination.

Mood and atmosphereNothing here needs to be exact to be right.

Concept illustrationsStanding in for an idea rather than depicting a thing.

Variations on a themeWhere being surprised is the point, not the problem.

Whether you may use generated images commercially is a terms question rather than a technical one, and it belongs to Meta’s own current terms rather than to a guide. Check before building anything commercial on them.

Where you meet it

On a Mac, it is already installed

There is no separate Muse Image app to download. It comes with the assistant you already have.

Inside the Meta AI Mac app

Image generation sits alongside everything else the app does, so if you have installed it you already have this. Apple silicon and macOS 15 or later, same as the rest of the app.

Free, with limits you will rarely notice

Image generation is one of the things that meters on the free tier, but lightly enough that most people never meet the ceiling. Meta One raises it, and it is usually video rather than image work that makes a subscription worthwhile.

If it stops behaving

Generation failing outright is a different problem from generation disappointing you, and it has its own short checklist Muse Image not working.

Go deeper:

FAQ

Muse Image questions

What is Muse Image?

Meta’s image generation and editing tool, part of the family powered by Muse Spark 1.3. It covers three different jobs: making an image from a description, editing an image you already have, and steering new output using a reference you supply. The three modes.

Why does my prompt keep not working?

Often because the job is not a prompting job. If you are chasing a specific look, a reference image gets you there faster than a longer description; if the image is nearly right, editing beats regenerating. Long prompts are what people reach for when they have picked the wrong mode.

Can Muse Image edit photos I already have?

Yes, and it is the mode people underuse. Editing keeps everything that was already right, which regenerating does not a fresh attempt changes things you never asked it to change.

What are reference images for?

Showing rather than describing. Style, palette, composition and mood are all things a picture communicates in one step and a sentence struggles with over several how that works.

Why is the result close but not quite right?

Because generators produce something in the neighbourhood of the request rather than an exact rendering of it. That is fine when the subject is generic and a problem when a specific detail has to be correct. Where that bites.

Can it get text inside an image right?

Written text inside generated images is a known weak point across image generators, so treat any words in the output as something to check rather than trust. For anything where the wording matters, adding the text yourself afterwards is the reliable route.

Does Muse Image cost anything?

Image generation is one of the things that meters on the free tier, but it is light enough that most people never think about it. A Meta One subscription raises the allowance, and it is usually video rather than image work that makes that worthwhile.

Where do I use it on a Mac?

Through the Meta AI Mac app, which generates images alongside everything else it does. It needs no separate install and works on the free tier.

Can I use the results commercially?

That is a terms question rather than a technical one, and the answer belongs to Meta’s own current terms rather than to a guide. Check before building anything commercial on generated images.

Should I use Muse Image or Muse Video?

If your message is a state how something looks use images. If it is a change over time, you need video. Images are also far cheaper to iterate on, so it is worth settling a look in stills even when a clip is the final deliverable the full comparison.

Mode first · prompt second

If the third try is just a longer second try, change mode.

Generating, editing and directing are three different jobs. Most of what feels like a prompting problem is one of them being asked to do another’s work.

The three modes Prompts that hold up

Already in the Mac app
Free tier, lightly metered
References beat long prompts
Check anything countable

Independent guide not affiliated with Meta. Usage limits and terms for generated images are Meta’s to set; check current terms before relying on them.