GPT Image 2.5 prompt — style transfer by assigning the reference a role
Say what the reference is *for*. Otherwise the model copies its content too.
| Mode | Model | Quality | Size | Ratio | Credits |
|---|---|---|---|---|---|
| Image to Image | Think | Standard | 1K | 2:3 | 10 cr |
Published by OpenAI as an example for this model — not generated on this site. Model: gpt-image-2.5-sunburst. Source: OpenAI The published example ran at 1024x1536, a legal size we do not currently offer; the frame here is the same shape at our nearest size.
A pixel-art image goes in; a man on a motorcycle comes out, in the same pixel-art style. The whole technique is in the first six words — "Use the same style from the input image" — which tell the model that this reference is a style sheet rather than a subject. Attach a reference without saying what role it plays and you will get some of its content in your output.
The prompt, nothing cut
Use the same style from the input image and generate a man riding a motorcycle on a white background.
This exact prompt, at these exact settings
The mode and all five settings are locked to this template, so there is nothing to line up by hand. Press Run this prompt above and the panel opens filled in — prompt, mode, quality, size and frame — and your first image needs no account. If your plan sits below what this ran at, the line under the panel says so.
This example was published at Think · Standard · 1K. A free run is Fast · Standard · 1K, so expect less fine detail and softer small text. The composition is the same; the finish is not.
What is safe to change
Most libraries publish this list and stop, which is why so many copied prompts come back worse than the original: the swappable nouns are the safe half. The three sections under it are the other half — the clauses that are doing the work, the ones that break it if you touch them, and why it is written in this order.
The role assignment
"Use the same style from the input image" is the entire technique. Narrow it when you want less — "use the same palette", "use the same line weight", "use the same medium" — and you get a more surgical borrow.
The new subject
Described completely separately from the reference, with no shared words. Keeping the two apart is what stops the model blending the reference's subject into yours.
The background
"On a white background" removes one variable. Style transfer is easier to judge, and easier for the model to execute, when the new image is not also inventing an environment.
Three ways to break it
Each one is a real failure with a reason attached. A rule without a reason is not usable.
01Attaching the reference without naming its role
OpenAI's fundamentals are explicit: identify each input by number and purpose. An unlabelled reference is ambiguous between style, subject, background and composition, and the model resolves that ambiguity by taking a bit of each.
02Using a photograph as the style reference
It works, but weakly. Style transfer is strongest when the reference has an obvious visual medium — pixel art, watercolour, woodcut, line drawing. A photograph's "style" is mostly its lighting and grade, which transfers as a mood rather than a look.
03Asking for a complex scene in a strong style
A distinctive medium constrains detail. Pixel art has a resolution budget; woodcut has a line budget. Ask for a crowd scene in either and the style will break down or the scene will.
Inside the Borrow one image's style and draw something else with it prompt
This is the shortest prompt in the library and it teaches the most transferable single habit: tell the model what each reference is for.
OpenAI's prompting fundamentals put it as assigning roles to references — identify each input by number and purpose, subject or style or clothing or background, and explain how they should combine. It sounds procedural. It is actually the difference between style transfer working and not working, because an image attached to a request without a stated role is genuinely ambiguous. Should the output contain what is in the reference? Look like it? Be composed like it? All three are common intentions, and a model with no instruction picks a blend.
The blend is what people are seeing when they complain that style transfer "leaks". You attach a pixel-art castle, ask for a motorcycle, and get a motorcycle next to a suspiciously castle-shaped structure. Nothing malfunctioned — the reference was offered as evidence and the model used all of it.
The second thing this prompt does right is keep the two halves lexically separate. The reference clause talks only about style; the subject clause talks only about the man and the motorcycle and the background. No word appears in both. That sounds trivial but it is what makes the boundary legible: any descriptive word that could attach to either half is a word the model has to disambiguate.
Where the technique is weakest is worth knowing. It is strong when the reference has a *medium* — pixel art, watercolour, woodcut, halftone, blueprint. Those are visual grammars with rules the model can apply to a new subject. It is much weaker when the reference is a photograph, because a photograph's style is largely its light and its grade, and those transfer as atmosphere rather than as a look you would recognise.
Borrow one image's style and draw something else with it — common questions
- Why does my style transfer bring across the original subject?
- Because the reference had no stated role. Say "use the same style from the input image" — an unlabelled reference is ambiguous and the model uses all of it.
- Can I transfer just the colours?
- Yes, and it is often better. Narrow the role: "use the same palette" or "use the same line weight" gives you a much more surgical borrow than "style".
- Does it work from a photograph?
- Weakly. A photograph's style is mostly lighting and grade, which transfers as mood. References with a distinct medium work far better.
- How complex can the new subject be?
- Less complex than you would like. A strong medium has a detail budget; ask for a crowd in pixel art and either the crowd or the pixel art gives way.
Other prompts that start the same way
- Think · Standard · 1K · 2:3
Input 
Result Change what someone is wearing without changing their face
Nine preservation clauses and nine prohibitions, because each one closes a known drift.
Bring an image10 cr - Think · Standard · 1K · 2:3
Input 
Result Put something from one photo into another
Numbers both inputs, names the element and its destination, then forbids everything else.
Bring an image10 cr - Think · Standard · 1K · 2:3
Input 
Result Cut a product out onto real transparency
Isolates the product on real alpha, and names the painted checkerboard as forbidden.
Bring an image10 cr
Written and maintained by Andy SwiftPublished Last updated