Everything Muse Image can do

Muse Image is Meta's image model. It plans each picture before it draws it: it can look things up on the web, run code to lay out charts and tables, and read up to ten pictures at once. Every picture comes out at about 2.5 megapixels, in eight shapes from 21:9 to 9:16.

Every picture below was made by the step under it. Change a placeholder and press Gen to make your own.

Name the job first

Open the prompt with what the picture is for, here a website hero, then write what a photographer would settle before the shoot: who, doing what, where each thing sits, where the light comes from. The model plans the layout from that, so ask for the empty space too: each of these keeps the left third clear for a headline.

Give counts as numbers and give every object a place. Thinking is the planning before the drawing: high, the default, holds counts and even spacing like these; low is quicker and enough for one subject on a plain backdrop. Both cost the same.

Words as written

Put every word you want set in quotes, give each line a role (the maker, the headline, the small print) and list them top to bottom. Words outside quotes are a description, and the model writes its own wording for them.

Keep each price with its item, in one quoted line, so the two can't drift apart on the board. Lettering can still drop a letter from one run to the next, and there is no seed to repeat a good run: raise Files per gen and keep the cleanest. Each file costs the same as one picture.

Charts that add up

With the Code tool on, as it is by default, the model writes and runs code to lay a chart out before drawing it, so the bars stand at the numbers you give. Put every value in the prompt, next to its label, and name the kind of chart and the axis range.

The same tool builds a QR code that scans. Write the link out in full and keep it short, and keep the code plain: black on white, with a clear margin and nothing drawn over it. Scan it with your phone before you print it; every run builds a new one.

Looked up, not made up

With Search on, the default, the model looks the subject up before drawing it. Name the exact place and where you stand: a named bridge seen from a named bank comes back as that bridge, not one like it.

Search finds facts too. Ask for the real values and the model looks them up instead of inventing numbers that fit the layout. Search costs nothing extra; turn it off when there's nothing real to find, such as your own numbers, an invented character or a drawing style of your own.

Point with words

An edit has no brush: the sentence does the pointing. Name the thing by what it looks like and where it is (the dirty plates at the left end of the counter), say what fills the gap it leaves, and list what stays. An edit keeps its picture's shape on its own.

To change one thing about one object, name the part, its new finish, and the highlights and shadows it keeps. The color changes and the light stays. Each edit here starts from the cleared counter.

For an old photo, name each fault: the color cast, the grain, the crease. Then pin the people and the place, so the model mends the photo instead of redrawing the day. Drop in your own old photo in place of this one, and check the faces against it afterwards.

Several pictures, one frame

A step can read up to ten pictures. To keep a character the same across a set, give every step the same sheet and the same words: these steps share one placeholder for the character, so its description never changes from one picture to the next.

This poster reads three pictures. Give each one a job and call it by what's in it: the courier from the sheet, the object she carries, the place from the travel photo. The model redraws all three in one style and one light. Where she stands and what she holds come only from the prompt.

To build on any step, press Mix to copy its recipe into your own page.