Use an image style reference without copying its composition

Example result.

1. Decide what you want from the reference

Look at the style reference and write down the one or two properties you actually want: light direction, shadow softness, surface texture, colour family or how busy the background is. Then write down what is in the reference that you don’t want: its subject, props, surface, clothes, wording and camera angle.

“Make it like this” hands the reference authority over everything, including the breakfast you didn’t order. Use only references you have the right to use, and don’t ask for a copy of someone else’s distinctive work.

Why this works: a named property (“low warm light from the left”) is something the model can transfer. “The vibe” invites it to transfer the whole picture.

What you should see: a short list of wanted properties and a longer list of things the reference must not lend.

2. Attach the images and give each one a role

Upload the images in a fixed order and refer to them by that order in the prompt. ChatGPT’s own image guidance recommends exactly this: give each image a role and identify it by its position.

Create [output and use] from the attached images.
Image 1 is the subject. Keep its shape, colour, surface and every visible detail exactly.
Image 2 supplies only [the properties you want]. Do not take its subject, props, surface, wording, background objects or camera angle.
Image 3 controls only the layout: [where the subject sits and what space stays clear].
Return one image. Then list any part of this request the images could not support.

[output and use] is what the image is for, such as “a 3:1 web banner for the spring sale”. [the properties you want] is your list from step 1. The layout line describes the sketch in words, so the model doesn’t have to guess.

Why this works: each attachment gets one job and an explicit list of things it can’t contribute. A text instruction states intent; it doesn’t force the tool’s behaviour, which is why step 3 exists. Firefly’s own controls give you a strength setting as well as the words.

What you should see: one candidate image and a short note on anything the model couldn’t do.

3. Check the candidate one role at a time

Compare in a fixed order. First the subject against Image 1: same shape, same colour, same details. Then the treatment against Image 2: did you get the light you asked for? Last, the layout against Image 3, plus any text.

Why this works: a beautiful light effect makes it easy to miss that a handle has moved or a prop has wandered in from the reference. Checking one role at a time stops the best part of the image hiding the worst.

What you should see: a verdict for each role: kept, borrowed correctly or leaked.

Check the result

• The subject matches Image 1 in shape, colour, surface and small details.
• The light or texture matches what you named in step 1, not the whole look of Image 2.
• Nothing from Image 2’s scene (objects, surface, wording, angle) appears in the result.
• The layout follows Image 3, with the clear area still clear.

The usual failure: the style reference leaks its scene. You get the light you wanted, plus the reference’s surface, props or camera angle. Spot it by covering the subject and asking: does the background look like Image 2’s background? If so, it has been copied, not borrowed.

The light from Image 2 is right, but the result also copied [the leaked items] from Image 2.
Remove them. Keep the subject exactly as Image 1 and the layout as Image 3.
Keep only [the property you want] from Image 2.
Return one image and say what you changed.

[the leaked items] are the objects, surface or angle you spotted; [the property you want] is your step 1 list again.

If generation keeps redrawing the subject itself, stop prompting and use a pixel editor or design tool to place the original subject on the new background. A longer instruction can’t guarantee an exact copy of your product.

Fictional example. Fernhill Ceramics needs a spring-sale web banner for its teal glazed jug.

Image 1  jug-source.jpg     Teal glazed jug, curved handle on the right, small chip-free spout, plain grey backdrop
Image 2  bakery-window.jpg  Low morning sun from the left, long soft shadows; marble counter, basket of croissants, shot from slightly above
Image 3  banner-sketch.png  3:1 frame; jug in the right two-thirds; left third empty for the sale text
Wanted from Image 2: low warm light from the left, long soft shadows

First candidate. The light is lovely, and the jug is sitting on a marble counter next to a basket of croissants, photographed from slightly above. The handle has also drifted to the left.

Role check. Subject: leaked (handle moved). Treatment: borrowed correctly. Layout: fine, but the counter and the basket came from Image 2. The jug did not ask for breakfast.

Repair. The repair prompt above, with [the leaked items] as “the marble counter, the croissant basket and the overhead camera angle” and [the property you want] as “low warm light from the left and long soft shadows”, plus one added line: “The handle is on the right in Image 1.”

Second candidate. The jug is back on a plain surface, handle on the right, shot at eye level, with long shadows to the right. The left third is clear for the text.

• The first prompt named the light but not the angle, so the model took the reference’s camera position too. The step 2 prompt now lists camera angle under “do not take”.
• Covering the jug made the counter obvious; with the jug visible, the eye went straight to the nice light.
• The handle drift was only caught because the subject check came first.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Sources

• OpenAI: Image generation
• Gemini Apps Help: Generate and edit images with Gemini Apps
• Adobe Firefly: Style reference images
• Adobe Firefly: Match image composition to a reference image