A ChatGPT account is the only thing between you and a first image. No separate app, no key, no setup.
Every step below comes from OpenAI’s own help centre. Where a number does not exist publicly, this page says so rather than filling the gap.
GPT Image 2.5 quick start: three doors
There are three ways in:
- A conversation — the default door. Describe the picture and ChatGPT makes it.
- The Images library — a standalone gallery with its own create button.
- Codex — the same generation and editing, inside a coding session.
OpenAI states that Images 2.5 is available to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile and web. For what the two API models are for, see the GPT Image 2.5 overview.
Where GPT Image 2.5 actually lives
| Door | How you open it | What you get |
|---|---|---|
| Conversation | Ask for the image, or choose More then Images | A generated picture inline, ready to keep editing |
| Images library | Open Images on web or mobile → Create Image | A saved gallery you can return to |
| Templates | Sidebar → Images → Templates | A preset layout instead of a blank prompt |
| Sketch | Type @ in the mobile composer → Sketch | You draw the outline first, then describe it |
| Codex | Generate or edit from a Codex session | The same models without leaving your code editor |
Two details are easy to miss: the gallery has its own address on web and mobile, and Sketch is a mobile entry point.
Your first image in five steps
The sequence is short; one step decides the result.
- Sign in to ChatGPT and open a fresh conversation.
- Choose Images if you want the picture parked in the gallery rather than buried in the thread.
- Write the prompt. State the subject, the framing, the light and what must not appear.
- Set the aspect ratio with the picker, or put the ratio in the prompt itself.
- Send it. The image appears in the conversation, or in your library.
The prompt that works first time
A first prompt usually fails for boring reasons: too many subjects, no stated framing, and a list of things to avoid that reads like a second subject list.
Subject: one matte ceramic mug with a handle, centred, slightly turned.
Background: seamless pale studio sweep, soft even light, faint contact shadow.
Framing: square, the mug fills about two thirds of the frame, even margin.
Exclusions: no text, no logo, no watermark, no props, no hands.Every line in that block is answerable from the picture itself; “beautiful” is not, so leave it out.
OpenAI says an image may take a few minutes depending on complexity, and you can keep working while it renders.
Editing the image you just made
You do not start over to fix one detail, and you need not draw first.
Select, or just describe it
OpenAI documents two editing routes and does not rank them:
- Select part of the image with the selection tool, then describe the change in the chat panel.
- Skip the selection tool and describe the edit in the conversation.
The same page is frank about the first route’s limit: highlights are not always precise, and an edit may extend past the area you selected. Change one thing at a time and a stray edit is easy to spot.
ChatGPT Images can add text, add detail inside an image, or make a background transparent, and you can upload an existing picture and describe what should change.
Templates and Sketch, in practice
Two entries in the table above are built for starting faster rather than prompting better.
Templates open as a preset layout — a poster, a merch mock-up — and you fill the slots. One official caveat: templates are not yet available in Work mode.
Sketch hands the outline to you: draw on the mobile canvas, confirm, then describe what it should become.
The Sketch guide and the templates guide go through each.
Saving, sharing and deleting
Nothing you make is lost by default, and one deletion is less obvious than the rest.
- Saved automatically. Every image lands under Images, so you can browse and reuse it.
- Out of the app. Copy it to the clipboard, save it to the device, or share it into another app.
- Prompt sharing. On mobile, the share sheet has a prompt template you can copy as a link.
- Deleting. An image goes when the conversation that made it goes — hover the thread title on web, or long-press it on mobile.
How many images a plan allows is not documented: OpenAI has not published per-tier quotas, so treat any number you read elsewhere as somebody’s observation. The free access limits page collects what is and is not published.
Two examples: first plate and first edit
The two examples below are the same job at two moments: making a picture, then fixing its background.
Example one: a first product plate
A maker needs one clean frame of a mug for an online listing.
Subject: one matte ceramic mug with a handle, lit from the upper left.
Background: seamless pale warm grey sweep, no horizon line, no props.
Framing: square, generous margin so the listing grid can crop it later.
Detail: visible clay texture, soft contact shadow, no blown highlights.
Exclusions: no text, no logo, no watermark, no hands, no second object.Send it once and read the result before sending again. If the light is wrong, change only that line — a prompt that also rewrites the framing teaches you nothing about which change did the work.
Example two: swapping a background
The mug is right; the desk behind it is not.
Keep: the mug, its shape, its proportions and its contact shadow.
Change: replace the whole background with a seamless pale studio sweep.
Keep: the direction of the light falling on the mug, unchanged.
Do not: add props, text, reflections or a second object.The edit is described, not selected, which is faster when the change covers the whole frame. If the result spills into the mug, delete that turn and re-send — a fresh attempt is cheaper than a chain of repairs.
GPT Image 2.5 quick start FAQ
- Do I need a paid plan? OpenAI says ChatGPT Images is available on all tiers, so a free account can generate. What each tier allows in volume is not published.
- Which model am I using inside ChatGPT? The interface does not name one. The two documented model IDs belong to the API — that split is covered in what changed in GPT Image 2.5.
- Why did an edit change something I did not mention? Selection is not always precise, per OpenAI’s own help page. Select more tightly, or describe the edit instead of drawing it.
- Can I edit a photo I already have? Yes. Upload the image and describe the changes you want.