ChatGPT Images 2.5: how to choose between Flare and Sunburst

What this page answers

ChatGPT Images 2.5 (also written GPT Image 2.5) turned the model choice from one into two, and added two quality settings above the previous ceiling. How you write a prompt has not changed, but "which tier and which setting for this round" is now a decision you make every time.

The method and the example images here come from a third-party guide, morphic.com's ChatGPT Images 2.5 walkthrough, cross-checked against OpenAI's own image prompting guide. Wherever a third-party claim and the official position differ, the text says so, and every example image is labelled with its source.

The official guide insists on keeping parameters out of the prompt: "transparent background" and "4K" belong in the request, while the prompt only describes the picture. The same split applies to selection — picking the model and picking the quality setting are two separate decisions, and the model comes first.

• This page assumes you already know how to write a prompt. The full method lives in the official prompting handbooks; this page is only about choosing.
• The third-party guide dates the release to 8 September 2026. Every "faster" or "more stable" claim below is that guide's usage observation; the official material gives no fixed improvement ratio.
• The example images come from the third-party guide. Copyright and licensing rest with the source, and they are used here only to show the difference in output.

Flare or Sunburst: the whole decision in one table

The two models do not differ in what they can do, but in what they optimise for. The official position: Flare is the small model, optimised for speed, with image quality comparable to GPT Image 2; Sunburst is the base model, optimised for quality, with higher image quality than GPT Image 2. Both support image generation, editing and transparent backgrounds.

The real fork is whether this round of work has a hard quality floor. The table puts the official position and the third-party recommendation side by side.

• Test quality before speed. The other order leaves you with a fast default you cannot ship.
• Do not migrate by swapping the model name: the guide recommends saving a baseline of representative production prompts and reference images covering difficult edits, exact text, faces, product geometry and transparent assets.
• Test the complete sequence of edits, not just individual steps — stability in editing workflows only shows up across successive rounds.

The official advice points one way, and it is enough to simply follow it: if your work already meets the quality bar, start with Flare and see how much time comes back; if it does not, establish the result with Sunburst first, then re-test whether Flare reaches the same bar.

There is a trap here, and both the official and the third-party guidance warn about the same thing: a speed-up on one workload may not hold on another. Do not trust a blanket number — run your own work through both.

Before you choose: what both models share

Before switching models, work out which layer your problem is on. According to the third-party guide, this generation improved four things on both tiers, which means these gains have nothing to do with choosing Flare or Sunburst:

• Sharper detail, with more natural lighting and texture.
• People and objects from reference images are preserved more completely — which decides whether clothing swaps and product edits are usable at all.
• Results stay steadier across a run of edits, with less drift over multiple turns (the official material still lists drift as a known limitation).
• Two quality settings were added above the previous ceiling (xhigh and max), for work that needs fine print detail.

If your problem is mushy detail, or a face that stops looking like itself after two rounds, a different model will not fix it — that is a prompt and process problem, and the two handbooks are the place to start. Model choice only trades quality against speed.

Quality settings: the first thing after picking a model

Both models share the same setting labels, but a higher setting does not guarantee a better result for every prompt. The third-party workflow is worth copying: iterate the wording at low, then run the final at the setting the job actually needs. Step up until requirements are met, then step back down once to see whether the lower setting still passes.

• The official order is: fix the model from the table above first, then touch quality.
• For a first two-tier comparison, hold the explicitly selected quality, the prompt, the reference images and the output dimensions all constant. The same quality label does not mean the same image quality or response time across models.
• Use xhigh and max only when they fix an unmet requirement inside your latency budget.

Sizes and cost also shape the choice

Output size is the most easily missed part of selection: the same prompt at 1:1 and at 16:9 do not cost the same, and 1024 does not cost the same as 4K. The time Flare saves can be eaten by one unnecessary square hero image.

• Up to 4K: each edge no more than 3,840 pixels, both edges multiples of 16, and a long-to-short edge ratio between 1:3 and 3:1.
• Outputs above 2560x1440 are marked experimental, which is worth knowing before a workflow depends on them.
• At the same quality setting a square is the most expensive shape, so a 16:9 or 21:9 output costs less than 1:1 at the same nominal resolution.
• Let the destination decide the shape: landscape for 16:9, vertical for 9:16. Do not generate square and crop.

Choosing by job: seven typical cases

Put the rule onto concrete work and selection becomes a lookup. The left column is the job in front of you, the right column is the tier to start with — note "start with": once the first image is out, the rule still applies, so re-test Flare wherever quality already passes.

Once chosen: the five prompting habits

With the model settled, output quality goes back to the prompt. The third-party guide attributes most problems to five habits; everything below is an expansion of one of them.

• The fourth habit is the easiest to miss: once a request carries seven or eight conditions, stop writing prose and split it into labelled blocks, adding an Include ONLY this text (verbatim): block when copy is involved. It changes nothing about the model's behaviour — it just stops one requirement from drowning in a paragraph.
• The official guide says the same thing in its own words: for complex requests, write the prompt as scene, subject, details and constraints, organised into labelled sections. Both roads lead to the same place.

Your first prompt: describe the picture, then add two things

Write the subject, the setting and the lighting in ordinary sentences. Words like "beautiful" and "4k" barely do anything, because they never say what should be in the frame. The third-party contrast is blunt:

Generate this version, look at what comes back, then change one thing at a time. Two additions are worth making early.

First: if you want photographic realism, ask for a photorealistic image and then describe framing and texture — 35mm film, medium close-up at eye level, shallow depth of field, subtle grain. Camera and film language is a cue for style; the model does not reproduce it as a physical simulation. That is the official position too: treat camera specifications as cues for appearance, not a guarantee of physical accuracy.

Second: write down what you do not want. Name the cliché you are trying to avoid, and that AI sheen drops off the image; saying only that you want it to look real usually achieves nothing.

Getting the text and the layout right

Put the wording that must appear in quotes, state where it goes, and state how many times it should appear. That last instruction is why a tagline does not mysteriously show up twice.

All six items and their prices were written into the prompt, which is what stops the model inventing them — that is the whole trick behind this menu board.

Billboard text (exact, verbatim): "Fresh and clean". Typography: bold sans-serif, high contrast, centered. Ensure the text appears once and is perfectly legible. No watermarks, no logos.

Four requirements, none of them vague: quote the copy, require it verbatim, specify placement, describe the typographic register, and close off every other possibility. A few smaller moves pay off too — spell brand names and unusual words letter by letter when the model keeps getting them wrong; keep the copy short, because small densely set text is the first thing to smudge; and when text is small or tight, do not assume the top setting wins, but compare high against xhigh.

Once single lines of text are stable, go one level up: name each part of the layout instead of describing a mood. That is what gives the model a hierarchy to follow.

Layout hierarchy: name each block, do not describe a mood

"A concert poster: the headline act set large across the top, four support acts stacked below in decreasing type size, venue and date in small type across the bottom" — written that way, the model has a hierarchy to follow instead of guessing at relative sizes.

Charts, slides and diagrams go one step further: write the prompt as a spec, not as an illustration request. Name the deliverable, put the real numbers and labels into it, specify the type family, and close off the clichés — no clip art, no stock photography, no decorative gradients. Landscape suits a reporting page; and once a figure carries a legend, axes or a footnote, high is the floor.

Afterwards, do not just check whether it is legible — check whether the labels are right. The official checklist puts "are diagram labels and relationships correct" first.

Change one thing, not the whole image

Most people lose a good image at the editing step. A good editing prompt has the shape of one clear instruction plus a preservation list.

Write a longer preservation list than you think you need. In the official clothing-swap example, face, facial features, skin tone, body shape, pose, expression, hairstyle, proportions, background, camera angle, framing and image quality are all pinned down before anything is allowed to change. That is not verbosity: whatever you fail to name, the model is free to reinterpret.

Later rounds should get shorter, not longer. After that first heavy prompt, the next instruction can be a single line: "Make it look like a snowy winter evening." Feed the previous result back in, ask for one change, and whenever a constraint starts to slip, restate it.

Watch for drift anyway. If something you never mentioned has moved, roll back to the last acceptable result rather than fixing forward. Where a region must stay pixel-identical, the honest move is not more prompting but compositing the approved change into the original image.

Keeping a character or product consistent

If a set of images has to look like one world, stop re-describing the character each time and hoping. Build a reference image first.

Generate the character on its own with a clean background, writing out its shape, proportions, clothing and tone. Then feed that image in for every scene and restate the key points in a short consistency block: the same green hooded tunic, the same facial features, proportions and colour palette, the same personality, and do not redesign this character.

• Products work the same way: geometry and labels come from the reference image, and the prompt issues the instruction not to restyle them.
• Brand mascots, series product shots and a whole set of detail-page assets all respond to this method.
• When the model can batch several images in one request, use it — a set generated together is steadier than six separate runs.

Transparent cutouts: the one step that needs both a setting and a prompt

Transparency needs two things at once: isolate the subject in the prompt, select a transparent background in the settings, and write PNG or WebP. JPEG cannot carry a transparent channel.

Isolate the product on a fully transparent background. Centred, crisp clean edges, no halos or fringing, clean alpha edges, generous padding on all sides. Product geometry and label text preserved exactly. Do not add a solid background, scenery, checkerboard or drop shadow, and do not restyle the product.

Then open the file and check, rather than trusting your eyes. A painted checkerboard is not transparency, and that is a very common failure — worth confirming every time on e-commerce white-background and detail-page work. And if the cutout goes on to be edited further, restate the requirement to preserve the transparent background; it does not carry over on its own.

The official guide's equivalent instruction: set background="transparent" explicitly in the request, write PNG or WebP, and check the decoded alpha channel — hair, glass, shadows and object edges.

Cutouts are also the most quality-sensitive job in the whole selection question: once a label or an edge goes soft you redo the image, so it is worth starting from Sunburst.

Beyond selection: four companion features

According to the third-party guide, this update added four things to ChatGPT alongside the models themselves. They do not change the selection decision, but they change how you work with either tier. All four descriptions below are the third party's; the official page does not state them in the same place.

• Sketch: draw directly in ChatGPT and use the drawing as the basis for generation. Add "preserve the sketch's layout, proportions and perspective, keep materials and lighting realistic, and do not add new elements or text" — otherwise the model reinterprets your sketch instead of following it.
• Templates: starting points for common formats such as posters and merchandise, filled in through follow-up questions. The third-party guide says these are not yet available in Work mode.
• Annotation: mark the image directly to point an edit at one spot; the editor's selection tool does the same by region. The third party relays an official caution: a selection is not always precise, and changes can spill outside the area you marked.
• Prompt sharing: publish a finished image together with the prompt behind it, so someone else can run the same idea with their own material.

Before you ship it: four checks

Run four checks in order and most problems stop there. The first three are about content; the fourth is about the file itself.

• Is the text accurate and legible, and does it appear the number of times you asked for?
• Are the labels in a diagram genuinely correct, or does it only look tidy?
• Did identities, product shapes and label text survive the edit?
• Did this round change only what you asked for, or did something quietly move?

If transparency was required, open the file and confirm the alpha channel is really there. If the image is going to press, look at the small text at full size rather than in a thumbnail — that is precisely where a "good enough" setting gives itself away.