GPT Image 2.5 prompting guide: prompts, edits and quality steps
Write GPT Image 2.5 prompts that hold up across revisions, describe reference edits and masks clearly, and use the six quality steps without wasted renders.
Read as MarkdownA good GPT Image 2.5 prompt names the subject, the composition, the light and the finish, and then says what must stay fixed. Write it once in plain sentences, draft it on Flare at a low quality step, and change one thing per revision. For edits, describe the change and the parts to preserve in the same prompt, and use a mask when only one region may move.
Most wasted renders come from changing several things at once, or from asking the model to guess what "better" means. The sections below turn that into a routine you can repeat.
What belongs in a GPT Image 2.5 prompt?
Four parts cover most briefs. Put them in this order so the most important information comes first:
- Subject. What the image is of, with the details that identify it: "a matte ceramic pour-over kettle with a walnut handle."
- Composition. Framing and placement: "three-quarter view, centered, generous empty space on the left."
- Light and setting. "Soft window light from the right, on a linen surface."
- Finish. The look of the whole image: "editorial product photograph, fine grain."
Then add the constraints that decide whether the image is usable: "no text, no logos," or "the handle must be fully visible." A constraint written as a plain sentence is easier to check than one buried in a list of adjectives.
Keep the prompt as long as the brief needs and no longer. The limit is 32,000 characters, but a prompt that repeats itself gives the model conflicting emphasis rather than more precision.
How do you ask for exact text in an image?
Quote the text and say where it goes: a poster with the headline "Morning Pour" in bold serif type across the top third, and a small caption "Single origin, slow brewed" beneath it. Spell out capitalization and line breaks if they matter.
GPT Image 2.5 handles layouts with type better than earlier GPT Image models, but rendered words can still come back misspelled. Treat every image with text as needing a proofread before it ships. If a word keeps breaking, shorten it, or leave the space empty in the image and set the text in your own layout tool.
How should you describe a reference edit?
Send the image in images and write the prompt as an instruction, not a fresh description. State the change first, then what to keep:
Change the bottle to deep blue glass. Keep the label, the composition, the background and the lighting exactly as they are.
The second sentence carries as much weight as the first. Without it, the model is free to improve things you did not ask it to touch. You can pass up to 16 reference images, for example a product shot and a style reference. Say in the prompt what each one is for, such as "use the first image for the product and the second for the color palette."
When is a mask worth using?
Use a mask when the edit has a boundary: replace the sky, change one object, remove a label. The mask is a PNG with the same dimensions as the first reference image, and its transparent area marks the region to change.
A mask guides the model rather than cutting pixels exactly, so describe the change in the prompt as well. "Replace the sky with a clear dusk gradient" gives the masked area a target; an empty prompt leaves it to chance. Leave the mask out when the change is global, such as a new lighting mood across the whole frame.
Which quality step and model should you draft with?
GPT Image 2.5 has six quality steps: auto, low, medium, high, xhigh and max. Higher steps take longer to render. Match the step to the stage of the work:
| Stage | Model | Quality | Why |
|---|---|---|---|
| Exploring composition | Flare | low | Fast feedback while the prompt is still changing |
| Refining details | Flare | medium or high | Enough detail to judge texture and type |
| Final asset | Flare or Sunburst | high to max | Only once the prompt is settled |
| Precise edit | Sunburst | high | Tighter control over what stays unchanged |
auto lets the model choose the step. That is convenient, but it makes comparisons between runs harder, so set an explicit step while you are iterating. The same goes for size: auto leaves the canvas to the model, while an explicit value such as 1536x1024 keeps every revision comparable. See Flare vs Sunburst for the model choice in more depth.
How do you revise without losing a good result?
Change one variable per attempt and write down what you changed. A short log is enough:
| Attempt | Change | Result |
|---|---|---|
| 1 | Base prompt, Flare, low | Kettle too small in frame |
| 2 | "fills two thirds of the frame" | Good framing, light too flat |
| 3 | "hard morning light, long shadows" | Keep; move to high |
When an attempt works, keep its exact prompt and settings with the image. Asking for n: 4 on a settled prompt is a cheap way to get variations to choose from without rewriting anything.
Frequently asked questions
Can the prompt ask for a transparent background?
Not reliably by wording alone. Set background to transparent and output_format to png in the request, then describe the subject as an isolated object in the prompt.
Should I write prompts differently for Flare and Sunburst?
No. Both read the same prompt and the same fields. Settle the prompt on Flare, then run the same text on Sunburst if the job needs tighter control.
Will the same prompt return the same image twice?
No. Each render varies. Keep the prompt and settings fixed when comparing, and generate a few images with n if you need options.
Make every revision answer one question
Write the brief in plain sentences, draft fast on Flare, and change one thing at a time until the image is right. Then spend the higher quality steps on the version you will keep. Try a prompt in the GPT Image 2.5 Playground, and see OpenAI's image prompting guide for the vendor's own advice.
