GPT Image 2.5 Prompt Generator
and the settings to run it at

You described the picture and got back almost it. Type the same line here: this GPT Image 2.5 prompt generator hands it back in the order the model actually reads — and tells you the quality, size and frame to run it at, so the next attempt is the one you keep.

What have you got to start with?

Nothing to attach. Anything you leave out, the model fills in for you — which is where most disappointing results come from.

Say it however you'd say it out loud

Or try one

Free, no sign-up, no daily cap. Checking a prompt happens inside this tab and never reaches us. Only a finished image costs credits, and those numbers are on the pricing page.

Before and after

What the prompt generator changed, line by line

The same request, twice. On the left is what people type; on the right is what the model needs before it starts filling in the blanks for you. Nothing here was written for the screenshot — the right-hand column is produced by the engine on this page, from the sentence on the left, every time the site builds.

What you type

an infographic for our Q3 numbers, three bars, clean, for a slide

Nine words. Plenty for a person. Everything it leaves out — the ground, the light, the type, what to leave off — the model decides for you, and that is why the result was almost right.

What you get back

Scene

A flat presentation slide on an off-white ground, generous margins, nothing behind the chart, Even, shadowless studio-flat lighting with no vignette, Sober and corporate.

Subject

A three-bar vertical column chart comparing quarterly revenue, the tallest bar on the right, Flat vector infographic, solid fills, 1px rules, no gradients.

Details

Straight-on, the whole slide in frame, Eye-level, no perspective on the chart, Title across the top third, chart centred, legend along the bottom edge. One accent blue against three neutral greys, crisp edges, true horizontal baselines. The text reads "Q3 Revenue", Top left, above the chart, Bold grotesque, sentence case, roughly 1/12 of the slide height. The text reads "Q1 Q2 Q3", Under each bar, centred on the bar, Regular weight, small, evenly tracked.

Constraints

Exclude: no extra labels or callouts beyond the ones specified, no watermark, no drop shadows.

Look at the order. The room comes before the thing standing in it. Nobody describes a picture that way out loud, and it is exactly what OpenAI's guide asks for: "organize the prompt as scene, subject, details, and constraints, using labeled sections." Getting that order right is most of the distance between the two columns — and it is the part that is tedious to do by hand every single time.

What guessing costs you

The settings it picked, and why

A sentence is the easy half. What decides whether it works is what you run it at, and that is the number nobody hands you — pick the wrong quality tier and you either get soft labels you have to re-run, or a bill several times larger than the picture needed. Here is what this request came back with.

an infographic for our Q3 numbers, three bars, clean, for a slide

Settings computed for the infographic request above
Quality	Size	Frame	Format	How many
Ultra	2K	3:2	PNG	1

That run costs 84 credits. The cheapest setting on this model is 6; the dearest is 288 — a 48× spread. Guess low on a chart and the labels come back soft, so you run it again. Guess high on a plain product shot and you paid many times over for a picture you were getting anyway. This is the guess we take off you.

Small or dense type. The prompting guide names exactly this case: compare medium or high before shipping.
Infographics carry labels, arrows and relationships that only survive at the top of the ladder.

What your kind of picture should run at

Eight jobs, and what the tool sets for a typical one of each. Every row is produced by the same function a live request goes through, and the credit figure comes from the one that bills it — so this is not advice about settings, it is the settings.

Recommended GPT Image 2.5 quality, size, frame, format and credit cost for eight kinds of job
The job	Quality	Size	Frame	Format	Credits each
Product photo for a shop page	Standard	1K	1:1	WebP	6
Poster or something you'll print	HD	4K	2:3	PNG	72
Infographic, chart or diagram	Ultra	2K	3:2	PNG	84
App or website mockup	HD	2K	3:2	PNG	48
Character, several views	Standard	1K	3:2	PNG	6
Logo or icon	Standard	1K	1:1	PNG	6
Illustration or artwork	Standard	1K	2:3	PNG	6
Edit of a photo you already have	HD	1K	2:3	WebP	24

A purpose alone does not decide this — the details do, which is why the tool reads your actual sentence rather than asking you to pick a category. The rows assume a typical job of each kind: a poster you will print, a diagram with real labels on it, an edit that has to keep someone recognisable. Notice how few of them need the top of the ladder. Five of the eight sit at Standard, and the one that costs 84 credits earns it on labels that would be illegible at 6.

None of those five numbers was typed in by hand — they come out of the same tables our generator consults before it sends your request, which is why they can be this specific. Our quality names and the API's disagree on four out of five. Every size-and-frame pair has to land on edges divisible by 16 and keep its long side at or under 3840px. Nothing wider than 3:1 gets through at all. A prompt tool with no generator behind it can tell you to "use high quality", and that is the end of what it knows.

Where are you starting?

Start from words, a photo, or a sketch

Pick where you actually are. These are the same three modes the generator runs — which is what separates this prompt generator from one that hands you text and wishes you luck. One press carries your prompt straight across with the settings already in place — no copying, no re-picking, no second guess.

Text to image
Nothing but an idea. The prompt carries the whole picture, and the frame is yours to choose.
No attachments
Open this mode
Image to image
You have the shot and want one thing different. The keep-list is what stops everything else moving.
References up to 16
Open this mode
Sketch to image
You have the layout and want it rendered. The drawing holds the geometry; the words bring the materials.
References up to 4
Open this mode

Read from the generator's own list of modes, so nothing can appear here that the tool cannot actually run. That is the difference between "copy this into your image tool" and one press that arrives set up.

The edit that keeps wrecking everything else

Change one thing, keep the rest

You asked for a different jacket colour and got a different face. Ask again and you get a different room. This is the hardest thing to get out of an editing model, and most prompt tools treat "don't change her face" as one more adjective competing with everything else in the paragraph.

A cream labradoodle stands on sunlit patio paving on a lead, a person’s trainer and jeans at the left edge, black woven patio chairs and a pool behind — One instruction: dress the dog. **Everything the instruction did not mention held** — the pose, the lead in the same hand, the trainer at the edge of frame, the chairs, the pool, the paving joints, the direction of the sun. That is what a keep-list buys you, and it is the difference between an edit and a re-roll. Pair published by OpenAI with the GPT Image 2.5 launch; ours to show, not ours to claim.

The same labradoodle in the same pose and the same spot, now wearing a white star-spangled stuntman suit and helmet with STUNTMAN on the chest — lead, trainer, chairs, pool, paving and shadow all unchanged — One instruction: dress the dog. **Everything the instruction did not mention held** — the pose, the lead in the same hand, the trainer at the edge of frame, the chairs, the pool, the paving joints, the direction of the sun. That is what a keep-list buys you, and it is the difference between an edit and a re-roll. Pair published by OpenAI with the GPT Image 2.5 launch; ours to show, not ours to claim.

She stays the same person
the face and identity of every person
Same camera, same crop
the camera angle, framing and composition
Every word stays as written
all text, labels and typography exactly as written
Same light, same shadows
the lighting direction, colour temperature and shadows
Product shape untouched
the exact shape, proportions and geometry of the product

What your ticks turn into

Change only what is described above. Keep unchanged: the face and identity of every person; the camera angle, framing and composition; the lighting direction, colour temperature and shadows. Leave every other part of the source image exactly as it is.

Tick them; don't write them. OpenAI's guide asks an edit to say change only X and then list what has to survive, and the reason it works is that the list arrives as a separate instruction rather than as more description. Your ticks travel with the prompt and the sentence gets assembled once, at the moment it is sent.

Already have a prompt?

Check a prompt you wrote somewhere else

Whichever prompt maker or chatbot you got it from, paste it into the Check tab and get an answer in a second — no account, no credit, and nothing sent to us. Every complaint quotes the line it objected to and tells you what to write instead.

Checks it runs 14

This gets refused, or does something you did not ask for

Weighting syntax
That is Stable Diffusion weighting syntax. GPT Image 2.5 has no weights — it reads the brackets and the number as literal text.
A brass pocket watch on a walnut desk, (masterpiece:1.3), (best quality:1.2)
TrySay it in words instead. "(dramatic lighting:1.4)" becomes "strongly dramatic lighting, the key light hard and low".
Fundamental 2 — no special syntax
A negative prompt
There is no negative prompt field on this model. Whatever follows this goes into the description as ordinary text.
Negative prompt: blurry, extra fingers, watermark
TryWrite exclusions as sentences in the prompt body: "an empty desk, no laptop, no watermark".
Fundamental 6 — state exclusions explicitly
A LoRA or checkpoint trigger
That is a trigger for a local checkpoint or LoRA. There is nothing on this side to load it.
A brass pocket watch on a walnut desk <lora:productStyle_v2:0.8>
TryDescribe the look you were loading it for, in words.
Fundamental 2 — no special syntax
A citation pointing past the uploads
This prompt cites Image 3, and only 1 reference is attached.
Apply the style of Image 3 to the product in Image 1
TryCite only what you uploaded, in upload order — the model has nothing to bind a citation that points past the end of the list.
Fundamental 7 — identify each input by number and purpose
A frame the endpoint refuses
4:1 is wider than the endpoint accepts. Measured, it answers: "maximum supported aspect ratio is 3:1".
TryUse 3:1 or 1:3 and crop the rest. There is nothing between those and the limit to ask for.
measured — endpoint refusal

This will run — you probably won't want the result

Words in the picture, described rather than quoted
You have asked for words in the picture but not marked off which words.
A shop window at dusk where the sign reads the name of the cafe
TryPut the literal string in quotes — the headline reads "Signal over Noise" — and say where it sits and how it is set.
Fundamental 5 — put required wording in quotes
Type on a tier that will not hold it
Type is the first thing to fall apart below the higher tiers, and this prompt has type in it.
TryCompare HD before you ship this one. The guide names small text and dense information panels as cases worth the extra latency.
Fundamental 5 — compare medium or high for small text, dense information, or multiple fonts
An edit with nothing marked to keep
This is an edit and nothing is marked to keep. Everything you did not mention is fair game.
TryTick what has to survive — identity, framing, text, lighting, product geometry — and we turn the ticks into the guide’s own "change only X" sentence.
Fundamental 6 — say "change only X" and list what to preserve
Camera numbers sold as a simulation
Camera numbers here are a look, not a simulation. Asking for them exactly will not make the optics exact.
A portrait at exactly f/1.4, 85mm, precisely simulated optics
TryKeep the spec if you like the look it suggests, but describe the result you want — "shallow depth of field, the background fully soft".
Fundamental 3 — camera specifications are cues, not a guarantee
A transparent background asked for in words alone
Asking for transparency in the prompt does not set it. Turn the Background chip to Transparent as well, and keep the format on PNG or WebP — JPEG has no alpha channel.
A brand logo isolated on a transparent background, no checkerboard pattern
TryKeep the line — it changes how the subject is composed — and set the chip too. It is a Starter unlock; on the free tier, ask for a flat evenly lit backdrop you can key out.
Worked example "Create a transparent product cutout" — the parameter and the sentence are two requests
A transparency request that invites the checkerboard
Transparency requests come back opaque with a painted-on checkerboard often enough that the guide says to rule it out by name.
TryAdd the exclusion explicitly: "isolated on a fully transparent background, no scenery, no solid backdrop, no checkerboard pattern, no drop shadow".
Worked example "Create a transparent product cutout" — "A drawn checkerboard is not transparency"
Several changes in one edit
There are 3 separate changes in this one edit. Each extra one is another chance for the rest of the picture to drift.
Change the shirt to green, remove the mug, add a plant on the shelf
TryRun the first change, then feed that result back and make the second. Repeat the keep list each time.
Fundamental 8 — iterate deliberately, one change at a time
A frame the model takes and this panel does not
21:9 is inside what the model takes, but it is not one of the 7 frames this panel offers.
TryNearest available is 3:1. Pick that and crop, or set a custom size — every pair we send has both edges divisible by 16 and a long edge at or under 3840.
measured — our panel, not the model
Two media in one request
This asks for a photograph and for an illustration in the same breath. You will get whichever one it read last.
A photorealistic product shot of a kettle in flat vector style
TryPick the medium, then use the other word for the subject rather than the render — "a photograph of a flat-vector poster on a wall" is a coherent request.
Fundamental 3 — name the visual medium

Most of the red ones are here because the phrasing that fills prompt-sharing sites was written for a different family of models. Weights, negative prompts and LoRA triggers are not stripped and ignored by GPT Image 2.5 — they are read as literal words sitting in your description, which is why a prompt that worked beautifully in Stable Diffusion can come back looking like it ignored you. The amber ones are judgement calls, and each one says whose judgement it is.

Before you spend a credit

What it refuses to promise

Three things this generator gets asked for every day and will not simply write down for you. Writing them down is free for us and expensive for you: the prompt runs, the credit goes, and the thing you asked for is not in the picture. Each one below says what is actually true and what to do instead.

Asked for daily
A transparent background from the sentence alone
GPT Image 2.5 produces real alpha, and the panel now asks for it — but on the chip, not in the prose. A prompt that says "transparent background" while Background is set to Auto returns an opaque picture of a subject that was composed to float. So the generator writes the line and the checker tells you to set the chip, rather than letting you find out from the file.
Set Background to Transparent with PNG or WebP — JPEG has no alpha channel, and that pair is refused rather than flattened. It is a Starter unlock; on the free tier ask for a flat evenly lit backdrop you can key out. Either way, read why a drawn checkerboard is not transparency before you trust a result that looks right.
Habit from other tools
A negative prompt
There is no negative prompt field on this model. A negative prompt: line is not stripped out — it is read as part of your description, so the things you were trying to exclude arrive as things you asked for.
Exclusions become ordinary sentences: "an empty desk, no laptop, no watermark". They go in the constraints block, and that does work.
Habit from other tools
Weights and trigger words
(dramatic lighting:1.4), <lora:…> and double-bracket emphasis are Stable Diffusion vocabulary. Here the brackets and the number are just more characters in your sentence.
Say what the multiplier was for: "strongly dramatic lighting, the key light hard and low" gets you there, and it survives being edited later.

The first one is worth being fussy about, because it has been wrong in both directions here. We once wrote that the model could not do transparency; it can. We then wrote that our route could not ask for it; it can now, by the direct API. What is left is the part that was true the whole time and is the only part you can act on — the parameter and the sentence are two separate requests, and one without the other gets you an opaque file. The one thing still missing is on us: nothing here decodes the returned image to confirm the alpha is real.

Then go make it

Take it straight to the generator

There are two buttons under every result and they deliberately hand over different text. Which one you want depends on where the prompt is going — our panel builds part of it for you, and anywhere else will not. Thirty seconds of reading here saves the “why are these different” moment later.

Open this in the generator
Carries the prompt, the mode, your Keep-Intact ticks and all five settings into our own panel, already filled in. Press Generate and you are done — your first image is on us.
Prompt · mode · ticks · five settings
Copy
Puts the full assembled version on your clipboard, keep-list sentence and all — because wherever you are pasting it has no tick boxes to build one from. It is the same text our panel would have sent.
The whole thing, as one block of text

Goes toText to image Image to image Sketch to image

Before you paste anything

GPT Image 2.5 prompt generator questions

What it costs, what it sends, and how it differs from asking ChatGPT

Is this GPT Image 2.5 prompt generator free?

Yes, and there is no account, no email and no daily cap. Rewriting a line runs a small text model that costs us a fraction of a cent — nowhere near what rendering a picture costs — so there is nothing to meter and no card at the end of it. Only a generated image spends credits, and the rates are on the pricing page.

Does my prompt get sent to your server? Runs in your browser

Only if you press Write. The Check tab is code running in your own browser — paste a client brief or an unreleased product description into it and nothing about it reaches us, because no request is made. Writing does go to our server, because something has to run the model that rewrites it.

Where do the recommended settings come from? The bit that saves you money

From the same tables our generator reads, through a function with no AI in it. The rewriter is only ever asked what it can see — is there small type, is this going to print, is there a face close up — and the quality, size, frame, format and count are worked out from that. Done the other way round, a model asked for a quality tier cheerfully answers "8K", and a number typed into this page by hand drifts away from the panel within a month.

Will better settings just cost me more credits?

Often the opposite. The spread between the cheapest and dearest run on this model is 48×, and most jobs do not need the top of it. You only get stepped up for the cases that genuinely fall apart lower down — small type, dense information panels, several typefaces — and left on the default the rest of the time, because headroom you cannot see in the result is headroom you paid for. The expensive mistake is rarely the cheap tier anyway; it is running the same job four times.

Why not just ask ChatGPT to write my prompt? Asked a lot

You can, and the prose will be fine. Two things it cannot do: it does not know what this site's panel will actually accept, so it will write you a transparent-background prompt without telling you that the alpha comes from a chip rather than the sentence; and it will not tell you what to set that panel to, because it has no idea what our quality tiers are called or what the endpoint refuses. The rewrite is the easy half. The settings are the half that needs a backend running the model.

What kinds of images is this prompt generator for?

It reads what you are making before it writes anything, because that is what sets how much polish the job needs. Product shots, posters and print, infographics and diagrams, UI mockups, character sheets, logos and icons, illustrations, and edits of a photo you already have. A diagram and a moody portrait want opposite things from the same model — one needs every label legible, the other needs the light right — and a prompt maker that treats them the same is why so many results land halfway.

What's the difference between this and your prompt library?

Where you are starting. The library is for "give me one that works" — you take a finished prompt and change a noun. This is for "I know what I want and cannot say it properly" — you bring your own idea and it comes back structured. If the blank box is the problem, start at the library and come back here when you have something of your own.

Why isn't "change only the jacket" in the prompt it gave me?

It is — as ticks rather than as a sentence, and it gets written for you the moment you send it to our panel. If both halves wrote it, the model would read the constraint twice and hand back a picture that is about the rule rather than about your subject. That failure is real and it is documented in our own code. Press Copy instead and you get the assembled version, sentence included, for use anywhere else.

Can it write me a prompt for a transparent background? Read this one

Yes — and it will also tell you the prompt is only half of it. Transparency is two requests: the sentence, which changes how the subject is composed, and the background parameter, which decides what lands in the file. Ask in words only and you get an opaque picture of a mark that was designed to float.So the checker flags the phrase and points at the Background chip. Set it to Transparent, keep the format on PNG or WebP — JPEG has no alpha channel and that pair is refused rather than quietly flattened — and note it is a Starter unlock. A run that asks for alpha goes by the direct API rather than our usual gateway, which lists auto and opaque only; the chip picks the route for you.

Will my Midjourney or Stable Diffusion prompts work with GPT Image 2.5?

The words will; the syntax will not. Weighted tags, negative prompts and LoRA triggers are read as literal text by this model, so a prompt full of them arrives with a lot of stray punctuation sitting in the description. Paste it into the Check tab and it will point at the exact lines and say what to write instead.

How long should a GPT Image 2.5 prompt be?

Longer than one line, shorter than you'd think. The limit that bites first is attention rather than characters: past a few hundred words your constraints start competing with each other and the model starts dropping some, which is why what comes back here is four short labelled blocks instead of a paragraph. As for the hard ceiling — nobody has measured it on this route, so you will not find a number quoted here.

Stop rewriting the same prompt.

Write one, run it at the settings it came with, and see whether the next image is the one you keep. Free to write, free to check — and your first image is on us: Standard, 1K, no card.

Written and maintained by Andy SwiftPublished 9 September 2026Last updated 10 September 2026Tool version 2026.09.1

GPT Image 2.5 Prompt Generator and the settings to run it at