Muse Image Prompt Guide: Text, Charts, QR Codes and Edits

All guides · Muse Image in Fuser

Quick answer: the best way to prompt Muse Image in Fuser is as a layout brief: name the medium, say where everything sits, put every word in quotes, and give each connected image one job. For anything that must be exact beyond spelling, such as a chart's scale or a QR code, draw or generate it first and connect it as an input for Muse to copy.

Muse Image is the first image model from Meta Superintelligence Labs, released on 7 July 2026. Meta says it "thinks through your prompt first", planning its layout and blending several visual references at once, and that it renders text cleanly (Meta). In Fuser, one Muse Image node does both jobs: with no image connected it generates from text, and with one to ten images connected it edits or combines them. Every example below is a real Muse Image run from October 2026, shown with its exact prompt, including the parts that came out wrong.

The Muse Image prompt formula

Write the prompt in this order:

1. Medium and light. Name the medium, film or print process and the light: "35mm colour negative, hard morning sun" or "a two-colour risograph on cream paper".

2. Layout. Say where each element sits: "title at the top left", "three small lines at the bottom left", "a line drawing in the top right corner".

3. Exact text. Put every word in quotation marks and ask for each line once.

4. Roles. With images connected, cite them in order as @Image1, @Image2 and say what each one supplies.

5. What stays. For edits, name the one change, then list what must stay exactly as it is.

Text: quote it, place it, ask for it once

Muse spelled every word we gave it in our tests; its slips were in counting and numbering, covered below. On this seed packet it got all 282 requested characters right, including the en dashes and the middle dots, and it used both inks we named:

Small print (text-to-image, 2:3): "The back of a paper seed packet, flat print design on kraft paper in black and dark green ink, shown flat and square-on, filling the frame. Exact text, each line once, as written: at the top, "SWEET PEA" in bold capitals, and under it "Lathyrus odoratus · Painted Lady" in italic. A heading "SOWING" followed by three small lines: "Sow indoors Feb–Mar, or outdoors Apr–May.", "Soak seeds overnight. Sow 2 cm deep, 8 cm apart.", "Pinch out the tips at 10 cm tall." A heading "GROWING" followed by two small lines: "Give climbing support from day one.", "Pick flowers often to keep them coming." A row of three small facts: "Height 180 cm", "Flowers Jun–Sep", "Fragrant". At the bottom, tiny: "Approx. 30 seeds · Lot 0412 · Packed for 2027". A small line drawing of a sweet pea flower in the top right corner. No other text."

Numerals need the same care as words. Our first run of the desk fan on the right printed its speed scale as 0 1 2 3 3; the version shown spells out how many numerals there are and that each appears once:

Counted numerals (text-to-image, 3:4): "Three-quarter studio view of an invented late-1960s tabletop desk fan. Its rounded shell is warm grey plastic with a perforated circular front grille; on its side, a large transparent acrylic speed wheel turns over a black scale printed with exactly four crisp white numerals, evenly spaced: "0", "1", "2", "3", each appearing once. A small engraved aluminium nameplate on the base reads "ARIA 3". The fan sits on a bent tubular chrome steel stand, its cable trailing, against a seamless neutral grey cyclorama in soft studio light."

Check sizes as well as spelling. Muse set the packet's "small" sowing notes at a comfortable reading size, larger than we meant.

Charts: the labels read, the scale doesn't

Muse gets chart text right. In every chart we made, the title, months, values, axis numbers and notes were spelled correctly. Bar heights were another matter. For a six-bar chart of swift counts (12, 48, 131, 96, 27 and 3), Muse drew an axis whose step from 100 to 150 was twice the height of the others, and against that axis its bars read roughly 17, 74, 143, 110, 37 and 5. Spelling the scale out ("a linear scale … each bar's top sits exactly at its value") did not fix it: the re-run labelled two gridlines 150. A chalkboard chart from the same week drifted too: we asked for stacks of 2, 3, 3, 4 and 6 lemons and got 2, 3, 3, 5 and 7.

What worked was drawing the chart first and letting Muse style it. We drew the same data as a plain black-and-white chart in a few lines of code, connected it as @Image1 and asked for the look:

Restyle an exact chart (edit, @Image1 = the drawn chart, 3:2): "Restyle the chart in @Image1 as a two-colour risograph print on cream paper in ink blue and fluorescent pink, photographed flat and square-on. The JUN bar is pink; everything else is blue. Keep the chart's data exactly as it is: the same axis labels and gridline positions, the same height for every bar, the same numbers, month labels, title and source note, in the same places. Change only the colours, the paper and the print texture."

Every bar kept its height, within about two units of its value, and the risograph texture and colours came through. In Fuser you can draw the chart in a Creative Code node and save the frame as a PNG Photo, which arrives as an image node you can wire into Muse.

QR codes: supply the code, don't ask for one

Meta's launch post says you can "write a custom prompt to build a functional QR code" in Meta AI (Meta). Through the Muse Image node in Fuser we could not get one. We asked twice for a code that encodes a web address, once chalked on a board and once printed on a card. Both came out looking like QR codes, and neither scanned, inverted or not.

Supplying a real code worked. Make the QR code with any generator, connect the image you want to change as @Image1 and the code as @Image2, and ask Muse to copy it:

Copy a supplied QR code (edit, @Image1 = the chalkboard, @Image2 = a real QR code, 3:4): "In @Image1, replace the QR code printed on the small white card pinned in the lower right corner with the exact QR code from @Image2: copy every black module precisely, keep its white margin, sharp and square-on to the camera. Change nothing else: the lemons, tape, chalk title, axis numbers, crate, window and sunlight stay exactly as they are."

The copied code scanned at full size, and still scanned with the image scaled down to 480 px wide. Everything else on the board, from the lemons to the chalk labels and the light, stayed as it was.

Several references: give each image one job

Connect images to the node's Images input and cite them in the order they are connected: @Image1, @Image2 and so on, up to ten. Say what each one contributes and what you keep from it:

Two references (edit, @Image1 = a marble bust, @Image2 = a lifeguard chair photo, 3:4): "Replace the red-haired man sitting on the wooden lifeguard chair in @Image2 with the weathered marble bust from @Image1, set on the seat at the same height, and rest his woven straw hat on the bust's carved curls. Nail a hand-painted white board to the chair's front rail reading "ON DUTY 10–18" in red brush letters. Keep the chair, the turquoise sky, the sun flaring through the hat brim, the low angle, the direct flash and the 35mm film grain exactly as they are."

Muse took the bust from @Image1 and everything else from @Image2: the chair, the low angle, the flare through the hat brim and the grain. The sign is spelled right, but we typed an en dash in "10–18" and Muse painted a hyphen.

When a borrowed object must take on another image's style, say how it should be rendered. Asked to float pianos over an impressionist riverbank, Muse first pasted them in as photographs; asking for them "fully repainted in the same thick, broken impasto brushwork as the willows" brought them much closer, though they still read smoother than the trees. In our three-reference test against GPT Image 2.5, Muse placed a chair, a vase and foggy dunes in one frame, but set the chair's feet in crusty clumps of sand; see Muse Image vs GPT Image 2.5.

Edits that change one thing

Connect one image, describe a single change and list what must stay. There is no mask input, so name the region in words:

One change (edit, @Image1 = a choir portrait made with Muse, 3:2): "Change only one thing: the woman in the middle of the front row who is shielding her eyes now wears a deep emerald-green choir robe instead of her white robe, same cut, same folds, same pose and same hand over her eyes. Change nothing else: every other singer, every face, every robe, the steps, the light and the framing stay exactly as they are."

Only her robe changed. Muse redraws the whole frame rather than pasting into it, though: in our edit test against GPT Image 2.5, 12% of the pixels outside the changed area moved by more than a 10% tolerance, against 8% for GPT Image. That is invisible at a glance but matters for approved artwork. For a chain of changes, wire one Muse node into the next and make one change per step, restating what stays each time.

Sizes, re-runs and no seed

• Aspect Ratio: Auto, 21:9, 16:9, 4:3, 3:2, 1:1, 2:3, 3:4, 9:16 or 9:21. Auto lets Muse choose; set a ratio when the image has to fit a layout.

• One resolution: Muse renders at about 2.5 megapixels whatever the ratio. Our 3:2 runs came back at 1920 × 1280, 2:3 at 1280 × 1920 and 3:4 at 1344 × 1792. There is no quality or resolution setting, so a re-run is never sharper, only different. For 4:5, render 3:4 and crop.

• Output Format: PNG (the default in Fuser), JPEG or WebP. Keep PNG for text you will edit again.

• No seed: every run re-plans the layout, so you cannot re-roll one detail. Our first climber kept its altitude labels right but spaced them unevenly; the re-run, asked for even spacing, labelled two ticks "2000 m". Keep outputs you like and change them with an edit instead of starting over.

Fix what goes wrong

• A repeated or missing number: say how many there are and that each appears once, as in the fan prompt.

• Chart bars that don't match the values: draw the chart exactly first, then ask Muse to restyle it as @Image1.

• A QR code that won't scan: don't ask Muse to invent one. Connect a real code and ask for an exact copy.

• A borrowed object that looks pasted in: describe how it should be painted or lit to match the scene.

• Dashes and sizes: proofread punctuation and type sizes as well as spelling. Our en dash became a hyphen on a painted sign.

• A detail that changes on every re-run: there is no seed. Fix the keeper with an edit.

Muse Image settings in Fuser

The Muse Image node has four inputs:

• Prompt: the description, or the edit instruction.

• Images: none for text-to-image; one to ten switch the node to edit mode.

• Aspect Ratio: Auto or one of nine ratios from 21:9 to 9:21.

• Output Format: PNG, JPEG or WebP.

Each run returns one image at a flat cost whatever the mode or ratio, and Fuser shows that cost on the node before you run it. In our tests a Muse image cost about a quarter of a GPT Image 2.5 image at High quality.

Build it as a workflow

Muse Image is most reliable when the exact parts arrive as inputs. In Fuser, draw charts in a Creative Code node, bring in a QR code or a product shot as an image, and wire them into Muse with a prompt in its own Text node so variations share it. Set long or legally exact copy as a text layer in the Compositor rather than asking any model to paint it. To see how Muse compares on the same prompts, read Muse Image vs GPT Image 2.5; for OpenAI's model on its own, see the GPT Image prompt guide, and for more models on lettering, the best AI models for text in images.