Work with me
image OpenAI

GPT Image 2.5

OpenAI's newest image model comes in a fast variant and a quality variant. Both do 4K and both edit an existing image without wrecking the rest of it.

At a glance

Resolution
1K, 2K or 4K
Aspect ratios
15 presets, from 1:3 to 3:1
Inputs
text, plus reference images when you edit

Good for

  • Editing one part of a picture and leaving everything else alone
  • Keeping a face or a product the same when you move it to a new scene
  • Large print-size output straight out of the model
  • Fast first drafts, then a slower pass on the one you keep

Weak at

  • Very wide or very tall crops, because the shape limit stops at 3:1
  • Long paragraphs of small text inside the image
  • Speed at the top quality setting, where each image takes real time
  • A published reference-image limit, which the providers do not agree on

What it is for

GPT Image 2.5 is a still image model that is unusually good at doing what you asked and nothing else. You give it a picture and one change, and it makes that change. The rest of the frame stays where it was.

That makes it the model to reach for when an image is already close. Swap a background, fix a label, put the same person in a different room. It is less exciting for wild invented art, and much better for work that has to match something you already have.

It ships as two variants. Flare is the fast one and it is the right default. Sunburst spends longer on each image and holds fine detail better, so use it for the version you are going to keep. Both take the same settings, so you can draft on Flare and re-run the winner on Sunburst without rewriting anything.

The hero image on this site was made with Sunburst, through OpenArt.

What it takes in

Text, and reference images when you are editing rather than generating from scratch.

Output runs at 1K, 2K or 4K. There are fifteen shape presets and they go from 1:3 to 3:1, so a normal social crop is fine but an extreme banner is not. You can also ask for an exact pixel size instead of a preset, as long as both edges divide by 16 and the long edge stays inside the 3:1 rule.

How to prompt it

Say what stays before you say what changes. This model reads that as a lock, and the parts you named stop drifting.

Keep the instruction in plain sentences. It does not need a shot list or a pile of style words, and a long stack of adjectives makes the edit vaguer, not sharper.

An edit that changes one thing 
Keep the person, their face, their hair and their jacket exactly as they are in
the reference image. Keep the camera position and the framing.
Replace the background with a quiet city street at dusk, wet pavement, warm shop
lights out of focus behind her. Match the light on her face to the new
background: soft, coming from her left, slightly warm.
Do not change her pose. Do not add text, logos or extra people.

Works on either variant. Draft on Flare, finish on Sunburst.

For a fresh image rather than an edit, describe the subject, then the light, then the camera, then the format. Ask for 4K only when you actually need it. At that size everything takes longer and most of the detail is lost once you post it.

Where it falls down

Shapes wider than 3:1 are out. If you need a thin banner, generate at 3:1 and crop, or build the wide frame somewhere else.

Small text inside the image is still a weak spot. It handles a short sign or a label. A paragraph on a poster will come back mangled. For that job use Nano Banana Pro instead.

Top quality is slow. The five quality levels go from low to max, and the top two are noticeably heavier. If you are iterating, stay low and only move up once the composition is right.

Finally, it is conservative. Ask it for something strange and it will quietly give you the sensible version. That is exactly why it is good at edits, and exactly why it is not the model for a wild idea.

Sources