What should a ChatGPT Images 2.5 prompt contain?

Six things, in roughly this order: the intent (what kind of image and what it is for), the subject, the composition and aspect ratio, any exact text in quotes with placement, the style and medium, and constraints that say what to avoid or keep. A generation prompt can skip the reference clause; an edit prompt cannot.

Are shorter prompts better in Images 2.5?

OpenAI has not published guidance on prompt length. In our experience the length that matters is the number of decisions you leave to the model, not the word count. A 120-word prompt with no contradictions beats a 40-word prompt that leaves the text, ratio, and reference use unstated.

How do I stop Images 2.5 from adding text I did not ask for?

Name every piece of text in quotes, say where each goes and which is largest, and end with a clause such as 'no other text anywhere in the image.' Empty space with no instruction is where invented copy appears.

Can I use the same prompt structure in the API?

Yes. The structure is model-facing, not app-facing. In the API you would put the aspect ratio in the size parameter rather than in the prompt text, and use the edit endpoint with a source image for edits.

Back to blog

Images 2.5September 9, 20266 min

How to Prompt ChatGPT Images 2.5

A working structure for Images 2.5 prompts: intent, subject, layout, exact text, style, and the constraints that stop the model from guessing. With before-and-after rewrites.

By Depikt Team

What changed, in one sentence

ChatGPT Images 2.5 is better at doing precisely what you say and, importantly, at leaving alone what you do not mention. OpenAI describes it as following editing instructions more reliably, keeping reference subjects recognizable, and handling more complex layouts. (The launch details are in what's new in Images 2.5.)

For prompting, that moves the effort. In earlier models you spent words fighting drift. Now the words that pay are the ones that remove a decision from the model: what the image is for, what the text says, what the reference is, what must not change.

This guide is the structure we use in the Prompt workspace. It works typed by hand too.

The six-part structure

Not every prompt needs all six. Every prompt that fails is missing at least one.

1. Intent

One line: what kind of image this is and what it is for. "Editorial poster for a print run." "App store screenshot." "Product photo for a white-background listing." Intent sets the defaults for everything below, and the model uses it. A "poster" gets typographic hierarchy; a "photo" gets a lens.

2. Subject

Who or what, with the two or three details that make it this subject and not a generic one. "A ceramic pour-over dripper, matte sand glaze, on a walnut board" beats "a coffee product." If a reference image supplies the subject, say so here (see part 6).

3. Composition and ratio

Aspect ratio, framing, camera position, where the subject sits. State the ratio even in ChatGPT; the model does not read your mind about vertical versus square. Our aspect ratio guide was written for GPT Image 2 and the recommendations still hold: 2:3 for posters, 4:5 for feed posts, 16:9 for slides and thumbnails, 1:1 for marks and icons.

4. Exact text

If text should appear, quote it verbatim, give each piece a position and relative size, and close the door on extra text. Images 2.5 handles defined layouts better than 2.0; it does not read your intent from a blank corner.

Prompt

Text, exactly as written: "RAIN CHECK" (headline, top, largest); "A short film by Ana Ruiz" (below, small caps). No other text, logos, or watermarks.

5. Style and medium

Medium first, then the two or three properties that define the look: "risograph print, two inks, visible grain" or "35mm photo, Portra 400 look, soft window light." Avoid stacking six adjectives; pick a medium and let it carry the style. The vocabulary cheat sheet is model-agnostic and still the best place to steal precise words.

6. Constraints and preservation

What to avoid, and, for edits, what to keep. This is where Images 2.5 earns its update. Because the model is better at preserving unmentioned details, an explicit PRESERVE list is more likely to be honored, which means it is worth writing one. The CHANGE / PRESERVE / MATCH framework covers this in depth; the short form is: name the change, list what stays, and say the new elements must match the old ones in light and grain.

Five rewrites

Rough idea on top, Builder-style prompt underneath. These are our examples, not OpenAI's.

Poster

Rough: "a poster for my pottery open studio"

Prompt

Editorial print poster, 2:3 portrait, for a pottery open-studio day. Subject: one hand-thrown stoneware bowl, speckled oatmeal glaze, photographed straight-on against warm off-white paper. Composition: bowl centered in the lower two-thirds, generous margin. Exact text: "OPEN STUDIO" (top, largest, condensed grotesk, near-black); "Sat 14 Sept · 10:00-18:00" (directly under headline, one-third the size); "Kiln Lane Pottery, Unit 4" (bottom-left, small). No other text. Style: risograph two-ink print, ochre and near-black, slight misregistration, paper grain. Avoid gradients, drop shadows, and any decorative border.

Product photo

Rough: "product shot of our water bottle"

Prompt

E-commerce product photo, 1:1. Subject: a 750 ml matte steel water bottle, forest green, brushed steel cap, no label. Composition: bottle upright, centered, filling 70% of the frame height, on a seamless pure white background. Lighting: large soft box from upper-left, gentle gradient reflection down the bottle's left side, soft contact shadow only. Style: clean catalog photography, 85mm lens, no props. Avoid: reflections of a room, colored casts, extra objects.

Infographic

Rough: "infographic on how a heat pump works"

Prompt

Explainer infographic, 4:5 portrait, for a landing page. Layout: three horizontal panels top to bottom, numbered 1-3, each panel a simple line diagram on the left and a two-line caption on the right. Exact text: title "How a heat pump works" (top, largest); panel captions verbatim: 1 "Outdoor air passes over the evaporator coil"; 2 "The compressor raises the refrigerant's temperature"; 3 "The condenser releases heat indoors". Footer: "Source: US Department of Energy" (bottom, smallest). No other text or labels. Style: flat vector, two-color (deep blue, warm gray) on white, 2px consistent line weight. Do not draw a house cross-section or add decorative icons.

Note the footer. If the image carries real-world information, give the model the exact source line rather than letting it invent one.

Reference-led portrait

Rough: "put my friend in a 1970s magazine cover" (with a photo attached)

Prompt

Magazine cover, 2:3. Reference image: use as SUBJECT IDENTITY only. Keep the face, hairline, and build recognizable; do not copy the reference's clothing, background, or lighting. Subject wears a 1970s wide-collar shirt and is lit by a single warm tungsten key from camera-right. Composition: head-and-shoulders, subject looking slightly past camera, masthead across the top. Exact text: "TEMPO" (masthead, largest); "The interview" (small, lower-left). No other text. Style: 1976 fashion editorial, film grain, slightly faded color. Avoid modern clothing, logos, and glasses unless present in the reference.

Edit

Rough: "change the sky to sunset" (with a photo attached)

Prompt

Edit the attached photo. CHANGE: the sky only, to a late-sunset sky with warm orange near the horizon fading to violet above, thin high clouds. PRESERVE: everything below the horizon line, the buildings, the people, the road, the crop and framing, and the original grain. MATCH: warm the light on the sunlit faces of the buildings to agree with the new sky; keep shadow directions unchanged. Do not add or remove any object.

Mistakes that still waste generations

Describing the end state only. "A photo of Sam in Tokyo" tells the model nothing about which parts of the attached photo of Sam matter. Say what the reference is for.
Adjective stacking. "Stunning, cinematic, ultra-detailed, 8K" adds nothing a lens, a light, and a medium would not add better. The 2.0-era list of prompt mistakes still applies.
Unquoted text. If the words are not in quotes, the model treats them as a description and may paraphrase.
Contradictions. "Minimal" and "richly detailed" in the same prompt. The Critique mode flags these first because they cap the score of everything else.
No ratio. You get a square. You wanted a poster.
Forgetting PRESERVE on turn four. Multi-turn consistency improved; it did not become memory. Restate what stays on every edit.

A note on prompt length

You may see advice that Images 2.5 prefers short prompts. OpenAI has not said that. Our experience is that the number of decisions you leave open is what matters, not the word count. A 120-word prompt with no contradictions and every text string quoted will beat a 30-word prompt every time. What you should cut is repetition and vibes, not specification.

Use the tools

Paste a rough idea into the Prompt workspace and it will produce a prompt in this structure, with the intent stage shown so you can see what it inferred.
Paste an existing prompt into the Critique mode for a score across intent fidelity, clarity, contradictions, composition, reference handling, edit preservation, text and layout, style coherence, factual integrity, and efficiency.
Browse the Library for 500 worked examples from the GPT Image 2 era, each with a note on why it works, and remix any of them for 2.5.

Sources

OpenAI, "Introducing ChatGPT Images 2.5," September 8, 2026: openai.com/index/introducing-chatgpt-images-2-5

Build a prompt that follows this guide.

Paste a rough idea. Get back a structured prompt.

Open Prompt

ChatGPT Images 2.5: What's New and What It Means for Prompting

OpenAI released ChatGPT Images 2.5 on September 8, 2026: better reference fidelity, edits that leave the rest of the image alone, more consistent multi-turn editing, more accurate layouts, and up to 50% lower latency. Here is the launch, sorted by what it means for the prompt you write.

Images 2.5

Precise Edits and Reference Images in ChatGPT Images 2.5

OpenAI's two biggest claims for Images 2.5 are better reference fidelity and edits that change only what you asked for. Both depend on the prompt saying what the reference is and what must stay. Here is how to write those prompts, with examples for identity, style, product, composition, and sketch references, and a routine for multi-turn edits.