# How do you make images with ChatGPT Images 2.0?

Date: 2026-05-01 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/make-images-chatgpt-images-2-0/

**Answer:** In ChatGPT, describe the image you want in plain language, including the subject, the setting, the style, the format and any exact text, then refine it with short follow-up requests. For a set of related images, write a short style guide first and reuse it word for word, or turn on thinking mode on a paid plan so ChatGPT can plan several consistent images from one prompt.

ChatGPT Images 2.0, released on April 21, runs on OpenAI's new image model, called **gpt-image-2** in the API. Everyone using ChatGPT gets the improved quality, including sharper small text, icons and interface elements, and better handling of non-Latin scripts. Paid plans (Plus, Pro and Business) can also turn on **thinking mode**, in which the model plans the image before drawing it, can look things up on the web, and can produce up to eight images that belong together from a single request.

You don't need special syntax. The skill is the same as with any good prompt: say clearly what you want, then improve it one change at a time.

## Writing a strong image prompt

A useful image prompt covers six things. You won't always need all of them, but checking the list prevents most disappointing first results.

| Element | Question to answer | Example |
|---|---|---|
| Subject | What is in the picture? | A loaf of sourdough on a wooden board |
| Setting | Where, and when? | A small bakery counter, early morning |
| Style | Photo, illustration, watercolor, flat icons? | Soft watercolor with ink outlines |
| Composition | What's in front, what's left empty? | Bread in the lower third, space at the top |
| Format | Shape and use | Vertical, for a phone story |
| Text | Exact words, in quotes | "Fresh every Saturday" |

Put any words that must appear in the image inside quotation marks, and say where they go. Text rendering is much improved, but OpenAI's own guide still lists precise text placement as a limitation, so always read the result letter by letter before you publish it.

## A worked example: four images for a bakery

A small bakery wants four matching images for social media: bread, croissants, a cake and its opening hours. Start with one image and get it right.

```prompt
A soft watercolor illustration with thin ink outlines of a round
sourdough loaf on a wooden board, on a small bakery counter in early
morning light. Warm palette: cream, wheat, terracotta, muted green.
Bread in the lower third, plain cream space above it. Vertical format
for a phone story. At the top, in a simple hand-lettered style, the
words "Fresh every Saturday".
```

Then refine with short, specific changes: "Make the light warmer," "Smaller lettering," "Remove the knife." Each follow-up should change one thing, so you can tell what worked.

If you build with the API, the same model is available through OpenAI's Image API. This Python sketch uses the official `openai` library; organization verification may be required before GPT Image models work on your account.

```python
import base64
from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY from the environment

STYLE = ("Soft watercolor with thin ink outlines. Warm palette: cream, "
         "wheat, terracotta, muted green. Subject in the lower third, "
         "plain cream space above. No other text than requested.")

def make(subject, filename):
    result = client.images.generate(
        model="gpt-image-2",
        prompt=f"{STYLE}\nSubject: {subject}",
        size="1024x1536",      # portrait
        quality="medium",      # low for drafts, high for final files
    )
    with open(filename, "wb") as f:
        f.write(base64.b64decode(result.data[0].b64_json))

make("Three croissants on a linen cloth", "croissants.png")
```

The guide lists `low`, `medium`, `high` and `auto` for quality, and suggests low quality for quick drafts. Note that gpt-image-2 does not yet support transparent backgrounds, and very large outputs, beyond roughly 2560 by 1440 pixels, are labeled experimental.

> **Try it:** Pick something you post about regularly. Write one image prompt using all six elements from the table, generate it, and make three follow-up changes, one element at a time. Save the final prompt; you will reuse it below.

## Things to check before you use an image

Look for extra fingers, warped objects, misspelled words and logos that look real but aren't. Make sure no real person is shown in a way they haven't agreed to. Image generation can take up to a couple of minutes for complex requests, so drafts at low quality save time. And if you plan to sell or advertise with the images, read OpenAI's current usage terms first.

## How do you keep a consistent style across a set?

Write a **style guide prompt**: a short paragraph that fixes the medium, palette, lighting, composition and lettering style, like the `STYLE` text above. Paste it unchanged at the start of every request in the set, and change only the subject line. Small wording changes in the style part cause visible drift.

Use your best image as a reference. In ChatGPT, upload it and ask for the next image "in exactly the same style as this one." In the API, the edits endpoint accepts one or more reference images with a new prompt, and gpt-image-2 always reads input images at high fidelity, which helps keep details stable.

On a paid plan, try thinking mode for the whole set at once: "Create four matching story images for a bakery: sourdough, croissants, a birthday cake, and our hours (Wed to Sun, 7 a.m. to 3 p.m.), all in this style." Because the model plans the set together, characters, objects and colors are more likely to match.

Even so, OpenAI notes that recurring characters and brand elements can still drift between images. Lay the finished set side by side, regenerate any image that stands out, and keep the prompts that worked in a document so the next set starts from the same place.

## Sources

- Image generation (API guide): https://web.archive.org/web/20260429060637/https://developers.openai.com/api/docs/guides/image-generation
- OpenAI's ChatGPT Images 2.0 thinks before it generates, adding reasoning and web search to image creation: https://the-decoder.com/openais-chatgpt-images-2-0-thinks-before-it-generates-adding-reasoning-and-web-search-to-image-creation/