How to use GPT Image 2.5 in ChatGPT
From first prompt to a finished edit chain that has not drifted. Every step below is something you do in the chat window, not in code.
Everything here happens inside ChatGPT, on a free account or any paid tier. GPT Image 2.5 rolled out to all ChatGPT tiers on release day, along with ChatGPT Work and Codex, across desktop, mobile and web. If you can open a conversation, you can run every step of the GPT Image 2.5 workflow below.
Two of the ten steps use features that exist only in the chat app. Sketch and templates have no API equivalent, so if you are planning to move this workflow into a product later, note which parts you would have to rebuild yourself.
Ten steps, in order
- 1
Open the image surface
In ChatGPT, describe the image inline, or select More and then Images. Generation runs in the background — you can keep chatting while it works.
- 2
Describe a picture, not keywords
Name the subject, the setting and the light in ordinary sentences. Adjective piles like 'beautiful, 4k, masterpiece' tell the model nothing about what to put in the frame.
- 3
Quote any text exactly
Put words you want rendered inside quotation marks: the sign reads "HARBOR LIGHT" in a bold serif. Unquoted text is a suggestion; quoted text is an instruction.
- 4
Set the aspect ratio up front
Use the aspect ratio picker, or state the ratio in the prompt. Re-cropping later costs a generation; asking for it now costs nothing.
- 5
Start from a template if you are stuck
Poster and merch templates give you a layout skeleton to overwrite, which is faster than negotiating a composition from a blank prompt.
- 6
Draw it with @Sketch
For spatial arrangements — who stands where, what sits on top of what — a thirty-second sketch beats three paragraphs of description. Type @Sketch and draw.
- 7
Edit with the selection tool
Open the editor, select the region, then describe the change. Highlights are approximate, so include the area in the prompt as well when precision matters.
- 8
Change one thing per turn
Scoped edits are the headline improvement, but they work best when the instruction is singular. 'Make the label matte black, keep everything else exactly' is a good turn.
- 9
Chain edits instead of restarting
Because earlier edits persist in the thread, refining across ten turns now beats rewriting one enormous prompt. Restart only when the concept itself changed.
- 10
Share the prompt with the image
When something lands, attach the prompt so a teammate can rerun it against their own reference photo instead of reverse-engineering yours.
Using Sketch properly
Sketch is the genuinely new interaction in GPT Image 2.5. Type @Sketch
in the composer, a drawing surface opens, and whatever you draw becomes a visual reference for the
image that follows. The obvious use is the one OpenAI leads with: some ideas are easier to draw
than describe.
The non-obvious use is spatial arrangement. Describing where four objects sit relative to each other in words is genuinely hard — "the mug behind and slightly left of the book, with the lamp cropped at the top edge" is three attempts waiting to happen. Thirty seconds of rough boxes conveys it exactly. Draw the layout, then let the prompt carry the style, the light and the materials.
Do not draw well when sketching for GPT Image 2.5. A tidy sketch invites the model to treat your linework as the intended style, which is rarely what you want. Rough boxes and stick figures read clearly as structure rather than as art direction.
Editing without losing what you had
The editor opens when you select any image in a conversation or in your images library. Inside it you can highlight a region with Select, change the aspect ratio, undo and redo your selections, and save when you are done. OpenAI is candid that highlights are approximate and edits may extend beyond the selected area, which is why the prompt should name the region as well as the selection marking it.
You can also skip the selection tool entirely and just describe the edit in the conversation. In this generation that path works better than it used to, because the model changes only what you name. Selection is worth the extra step when two similar objects are in frame and the words alone would be ambiguous.
Aspect ratio is available after the fact, and it regenerates at the new ratio rather than cropping. That is useful for repurposing an asset, but it is not free — the composition gets reconsidered. Ask for the ratio you need up front whenever you know it.
The multi-turn discipline
The behaviour that makes GPT Image 2.5 worth learning properly is that edits persist. Earlier changes carry through the thread and each new instruction builds on the last, instead of the image quietly degrading with every round as it did before. That turns the conversation into a version history, and it rewards a specific discipline.
Change one thing per turn. Name what stays fixed. When something lands wrong, ask for the previous state rather than starting over. And when a result is good, share the prompt with it — a teammate can then rerun the same idea against their own reference photo instead of reconstructing yours by hand.
Common mistakes and their fixes
✗ Generating drafts at high quality
✓ Fix: Iterate at low, then re-run the winning prompt at high. The tier spread is roughly 35x per image, and you are discarding the draft anyway.
✗ Stacking four edits into one turn
✓ Fix: Scoped editing is per-instruction. Split them: one change, then the next, and the thread will carry both.
✗ Leaving text unquoted
✓ Fix: Quote every string you want rendered. Unquoted words get paraphrased into something that looks like text but is not.
✗ Cropping to a ratio after the fact
✓ Fix: Ask for the ratio in the prompt or the picker. A post-hoc crop throws away composition the model spent its budget on.
✗ Reaching for Sunburst by default
✓ Fix: Flare is the default for a reason. Sunburst buys edit precision at the cost of generation time; pay that only when the asset ships.
✗ Assuming a per-image price
✓ Fix: Billing is token-based. Size and quality both move the number, so estimate before you queue ten thousand jobs.
✗ Describing a mood instead of a frame
✓ Fix: 'Cinematic and moody' is not an instruction. Name the light source, its direction, and what it falls on.
✗ Restarting the thread after a bad edit
✓ Fix: Ask for the previous state back instead. The conversation is the edit history, and losing it costs you every earlier refinement.