How to use ChatGPT Image 2.5, and where to open it
There are three honest doors. Pick the one that matches the next picture you need, not the one with the longest feature list.
Most people who search “how to use GPT Image 2.5” are not looking for a manifesto. They want to know which tab to open, what to type, and why the first try sometimes comes back empty.
Here is the map we wish someone had handed us.
Three doors
ChatGPT. This is the official one. Sign in, ask for a picture the way you would ask a colleague, attach a photo if you have one, or start from a sketch. If you already live in that chat, stay there. You get the real Images 2.5 model, and official files carry OpenAI’s provenance marks.
This site. A single page. Prompt box, optional reference photos, a ratio, a size. No install. The first 1K image does not need an account. After that we ask you to sign in, because unbounded anonymous generation is how these tools get wrecked.
The API. If you are wiring this into an app, OpenAI publishes Flare and Sunburst. That is a developer path, and we kept it off this page on purpose. The comparison note is next door.
If you are using the generator here
Scroll to the box on the homepage. You do not need a new app.
- Write the scene in ordinary language. Subject, place, light. One job per sentence is plenty.
- If the picture has to stay loyal to a real object or a real face, drop a reference photo in. You can add up to five. One good photo beats three muddy ones.
- Pick a ratio if you already know the shape — 16:9 for a wide still, 9:16 for a phone story, 1:1 if it has to sit in a grid. Auto is fine when you do not care yet.
- Start at 1K. It is faster, and it is enough to see whether the idea works. 2K and 4K can wait until the frame is right.
- Press Generate. If the result is close, do not start over. Say what to change: the cloth, the sky, the type. Leave the rest alone on purpose.
The first image without an account is a 1K preview on a lighter model, so you can see the flow. After you sign in, generations run on GPT Image 2. If you specifically need official Images 2.5 — the one OpenAI announced — open ChatGPT.
If you are using ChatGPT
You already know how to talk to it. A few habits still help.
Describe the room before you describe the mood. “North light, pale stone, one bottle, label facing us” will beat “make it premium and cinematic.” If you upload a reference, say what must not move: the face, the logo, the furniture layout. If you start from a sketch, say which lines are structure and which are just scribbles.
When it misses, name the miss. “The bottle grew a second label” is more useful than “try again, better.”
Three first prompts that usually work
These are starting points, not magic words. Change the object to whatever is on your desk.
A product, no drama.
“Catalog shot of a matte sage bottle on limestone, soft north window light, label sharp and centered, no extra logos, no props except a folded linen napkin.”
A poster with real type.
“Poster for a sparkling yuzu drink on pale lemon paper. Headline in one line: FRESH AND CLEAN. Lots of empty margin. No fake barcodes.”
A person you already have a photo of.
“Same person as the reference, same coat, walking a wet city street at dusk. Do not change the face or the rust color of the coat. Only the street and the weather.”
If you are coming from a pencil drawing, say so in the first sentence: “Finish this room sketch as a loft living room, warm oak, morning light. Keep my furniture layout.”
When it feels stuck
Nothing happens after the first picture. On this site, the free try is one generation. Sign in if you want another. New accounts get a small credit gift; a 1K image still costs 10 credits, so that gift is a head start, not a second free print. Plans are on the pricing note.
2K or 4K is greyed out. The unsigned preview is 1K only. Size opens up after you have an account.
The reference seems ignored. Say what must stay. “Keep this label” is a stronger instruction than uploading the photo and hoping. Also check that the file actually landed in the drop zone — we see people type the prompt and forget they never attached the picture.
The model “improved” the thing you liked. That is the old habit of image models. Tell it what is already finished. “Do not restyle the bottle. Only replace the background with a winter table.”
Square 4K comes back smaller. Some combinations are limited upstream. If you asked for 1:1 at 4K and got 2K, that is a constraint, not you doing it wrong.
If you came here because you wanted a free box that just makes a picture, the next note is written for that: a generator in the browser.