ChatGPT Images 2.5 Official Guide, Reorganized

Slot one: purpose

Say what the image is and what it is for. The guide asks you to define the result first, naming the subject and the use, for example a product photo, an ad, a diagram. When the content gets complex, break the prompt into sections, scene, subject, detail and constraints, with a heading on each.

Slot two: composition and constraints

Next, say how the frame is arranged: composition, aspect ratio, and what has to sit where. This is the slot I drop most often, and the result is the subject cropped away, empty space in the wrong place, and text squeezed against the edge.

Slot three: visible details

The guide calls this one Describe visible details: state the material, the lighting, the colour and the visual medium. For a real photo, say "photorealistic" or "real photograph" outright and describe the framing and the texture. Camera specs (a 50mm lens, say) are only a hint about the look, not a promise that the lens is physically simulated. For wide-angle, cinematic, low-light, rain or neon scenes, the guide suggests writing out scale, atmosphere and colour, because mood adjectives alone are not enough.

Slot four: people and action

If there are people in the frame, say how much of the body is in shot, how big the person is next to the object, where the eyes are looking, and how the hands touch things. The guide's own examples are sentences like "full body in frame, including feet", "looking down at an open book", "both hands resting naturally on the handle".

Slot five: text in quotes, verbatim

For words that appear in the image, the guide asks you to put them in quotes verbatim and describe placement and font separately. For brand names or unusual words, spell them out letter by letter when needed. End with a line saying no other text, then check the spelling and legibility yourself. For small type, dense information, or several fonts at once, the guide suggests running quality at medium and at high and comparing the two.

Official example (the Render exact text section):

Slot six: exclusions

The last slot is what you do not want. The guide uses the same pattern across its examples: list the unwanted elements as a row of short phrases at the end. Common exclusions are watermarks, logos, extra text, checkerboard backgrounds, shadows, over-retouching, a stock-photo feel, gradients and ornament.

Swap one thing, keep everything else

Official example (the Change furniture in a room section):

Give each reference image a role

When you drop in several reference images, the guide asks you to number each one and say what it is for: subject, style, clothing, background. Then say how the inputs combine and which element moves where. Without that step the model has to guess what each image is for.

Official example (the Combine references section):

Change the clothes, keep the person

Portrait editing lives or dies on what you pin down. The guide's clothing example lists it in detail: face, features, skin tone, body shape, pose, identity, expression, hairstyle, proportions. Only the clothing changes, and the lighting, shadows and colour temperature have to match the original so it does not look pasted on.

Sketch to finished image

Turning a sketch you already have into a finished image comes down to two things: preserve the original layout, proportions and perspective, then fill in plausible materials, lighting and surroundings. The guide specifically asks you to add "do not add new elements or text" so the model does not improvise. With no Sketch canvas, draw on paper, photograph it, upload it, and this section still works.

Official example (the Turn a drawing into a realistic image section):

Transparent cutouts

A cutout needs two things at once: the prompt has to ask for the subject in isolation, and on the API side you set background to transparent and output PNG or WebP. The guide has one very practical warning: a drawn checkerboard is not transparency. Real transparency means the file carries an alpha channel, so open the file you get and check hair, glass, shadows and object edges. On later edits, repeat "keep the transparent background". The ChatGPT web app has no parameter to set, so write the requirement into the prompt in full and check the edges yourself once you have the image.

Official example (the Create a transparent product cutout section):

Remove an object

Removal is the shortest pattern: name the thing to remove, then tell it to leave everything else alone. The guide's version is two sentences.

Official example (the Remove an object section):

Put a person into a new scene

For compositing a person into a new scene, the guide asks for four things to be spelled out: natural lighting, believable detail, how much of the body is in frame and where they look, and how the person interacts with the scene. Then name the features and proportions that must not change. If you want it to look like a casual snapshot, say no cinematic lighting, no dramatic grading, no stylised composition.

Step one: get the starting image right

The starting image has to nail the text and the subject in one go. The guide's example wraps the billboard copy in "EXACT, verbatim, no extra characters" and spells out font, contrast, centring and letter spacing, then asks for the text to appear once.

Step two: change one condition at a time

Send the image from the last step back in, then name only the one thing that changes. The guide's demo for this step is a single sentence.

Official example (the Change one condition section):

Step three: restate what must be kept, every round

The editing section of the guide has an important warning: repeated edits can still wipe out the detail you meant to keep, so restate those constraints and check every round. If an area must not move by a single pixel, the more reliable route is compositing the approved edit back onto the original, not relying on the prompt alone.

Character consistency: build a character sheet first

For a picture book, an illustration series or serialised content, the guide's approach is to generate a character reference first, fixing the look, proportions, clothing and temperament, then repeat those defining details on every new scene and change only the setting and the story.

One: product photo

For products, OpenAI focuses on material, packaging and print legibility, and asks for original designs that do not infringe.

Two: flyers and text posters

For text-heavy images, the guide's approach is to write the prompt like a spec: state the message, the design elements and the style, put the words in quotes verbatim, and say outright that the layout has to be readable.

Three: logo

The guide's logo example stresses three things: simple shapes, legible at every size, delivered on a transparent background. It also puts "checkerboard" straight into the exclusions, because the guide makes a point of it: a drawn checkerboard is not real transparency.

Official example (the Design a reusable logo section):

Four: process diagram

For flowcharts and infographics the guide adds one more warning: beyond the look, verify the labels and the factual relationships yourself. A wrong name or a reversed arrow makes a beautiful diagram unusable.

Five: four-panel comic

Turning a story into a comic comes down to splitting it into one visual beat per panel, described concretely and driven by action, so the model can lay out panels that read with rhythm.

Six: slide graphic

The guide suggests a landscape size for slides, and writing the numbers and labels straight into the prompt instead of letting the model invent them. When the image has small type, a legend, axes or a footnote, set quality to high. The guide adds its own reminder: the market figures in its example are fictional design inputs, and you should replace them with verified data before real use.

The guide adds one more line: when you change prompts or change models, run representative inputs and compare quality, response time and cost, all three together.

How the two models differ

How to choose: on a new pipeline, try Flare first when speed matters and Sunburst first when quality is strict. Once Sunburst passes, test Flare with the same prompts and inputs, and switch if it also passes and is faster. Keep Sunburst only when the quality gap genuinely calls for it. One reminder from the guide: measure speed on your own workload, because a figure from someone else's workload does not transfer.

Parameters

Custom resolutions have four limits: no side above 3,840 pixels, both sides multiples of 16, the long side no more than 3:1 against the short side, and total pixels between 655,360 and 8,294,400. Output above 3,686,400 total pixels (the equivalent of 2560x1440) is experimental.

The way to use quality is to pick the model first and tune quality after. On a first comparison, hold the quality setting that both models support fixed, and hold the prompt, the reference images and the output size fixed too. The same quality label on a different model does not mean the same picture quality or the same response time. Step quality up when it falls short, then step back down once it passes, to see whether you can get faster while quality stays acceptable. Use xhigh and max only when they genuinely fix a quality requirement that is failing and still fit your time budget.

For transparent assets, set background to transparent explicitly, use PNG or WebP, and check the decoded alpha channel once you have the file, especially hair, glass, shadows and object edges. output_compression applies to JPEG and WebP only, not PNG.

Three things that matter when migrating an old pipeline

The guide's migration flow has six steps. Here they are boiled down to three points.

1. Save a baseline first: collect representative production prompts and reference images, including hard edits, exact text, faces, product geometry and transparent assets, and record the model, the request settings and the results you use today.
2. Change only the model in the first comparison: hold the prompt, reference images, size and output format fixed. Compare instruction following, preservation of identity and product, text accuracy, whether anything got changed that should not have been, and transparency. Repeat each request several times to see how stable it is, and test editing pipelines as a whole chain, not one step at a time.
3. Change one setting at a time: compare quality levels before touching the prompt. Measure typical response time, slow responses, failures and retries, plus the cost per usable image. The guide specifically warns against assuming a faster model is cheaper, so check the current pricing. When you go live, send a small slice of traffic first, watch the same metrics, then scale up, and keep the old model as a fallback while it is still supported.

I can't find Sketch in my ChatGPT

The launch post says typing @Sketch in ChatGPT opens the canvas, but it looks like a batched rollout, and my own account still didn't have it on 2026-09-10. If you don't see it, take the other route: draw the sketch on paper or in any drawing app, photograph or screenshot it, upload it as a reference image, then paste the "sketch to finished image" prompt above. It still keeps your original layout and perspective.

The text in my image keeps coming out wrong

Do the four moves from slot five: put the words in quotes verbatim, say how many times they appear and where, describe font and placement separately, and end with a line saying no other text. For small type, dense information or several fonts, move quality from medium to high and compare again. For brand names or unusual words, spell them out letter by letter when needed. Always check spelling and legibility yourself once the image is out.

The ChatGPT web app does not have those parameters. What now?

quality, size and background are API parameters, and I did not find matching settings in the web app. What you can do there is write the requirement into the prompt: for a transparent background, spell out "fully transparent background, no solid backdrop, no checkerboard", and for landscape just say landscape composition, then check the image yourself. The guide is written for the API. The prompt patterns work fine for me in the web app, and the parameter layer is the part the web app does not have.

How much faster is it really

The launch post says generation latency drops by up to 50% compared with Images 2.0. The API docs say GPT-Image-2.5 Flare has 50% lower latency than GPT-Image-2. The guide also warns that the real figure has to be measured on your own prompts, reference images, output sizes and quality settings, because numbers from another workload do not transfer.