# How do you get better images from the new ChatGPT Images?

Date: 2025-12-24 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/get-better-images-new-chatgpt-images/

**Answer:** Write image prompts in a fixed order (scene, subject, key details, constraints) and say what the picture is for, use camera and lighting words for realistic photos, and put any text you want in the image in quotation marks. Then improve the result with small, single-change follow-ups instead of rewriting the whole prompt.

Since December 16, image creation in ChatGPT has run on a new OpenAI model, available to developers as **GPT Image 1.5**. OpenAI says it follows instructions more closely, renders denser and smaller text, and produces images up to four times faster. The same day, OpenAI published a prompting guide for the model in its Cookbook, its public collection of developer examples. The advice in that guide works just as well when you type into ChatGPT, and it boils down to one idea: describe the image the way a photographer or designer would brief a colleague.

If you want to skip writing altogether, the new **Images** space in the ChatGPT sidebar offers preset styles and trending ideas you can apply to your own photos. For anything specific, a good prompt still makes the difference.

## A structure that works

OpenAI's guide recommends writing every prompt in the same order:

1. **Scene or background:** where we are, the time of day, the setting.
2. **Subject:** the main person, object or idea.
3. **Key details:** materials, colors, textures, the visual medium (photo, watercolor, 3D render).
4. **Constraints:** what must or must not appear, such as "no logos" or "no extra text".

Add the **intended use** at the start, because "a poster for a library event" and "a product photo for a website" call for very different levels of polish. For long requests, short labeled lines are easier for the model to follow than one dense paragraph.

Three more habits help:

- **For realism, talk like a photographer.** Lens, framing and light ("50mm lens, eye level, soft window light from the left") steer a photo more reliably than "ultra-detailed 8K".
- **For text in the image, quote it.** Put the exact words in quotation marks or capitals, and state the font style, size, color and position. For an unusual name, spell it out letter by letter.
- **For several images, number them.** "Image 1 is my product photo; Image 2 is the style I like. Apply the style of Image 2 to Image 1."

## A worked example: a poster for a community event

Here is a first prompt for a winter event at a small-town library:

```prompt
Use: a printable poster (portrait, letter size) for a public library.
Scene: a cozy small-town library at dusk in winter, snow outside the
tall windows, warm lamps inside.
Subject: an older woman and a young boy reading a picture book together
in armchairs, seen from the side.
Details: soft illustrated style, like a modern children's book; warm
golds and deep blues; plenty of empty space at the top.
Text, exactly as written, in a clean rounded sans-serif:
Top, large: "STORY NIGHT"
Bottom, smaller: "Thursday 7 pm · Free · All ages welcome"
Constraints: no other text, no logos, no watermark.
```

Suppose the first result is close but the bottom line of text is cramped and the boy looks older than intended. Fix one thing per message:

| Round | Follow-up message | Why |
|---|---|---|
| 2 | "Make the boy about six years old. Keep everything else the same." | One change, with a protect instruction |
| 3 | "Make the bottom line of text larger, with more space around it. Same words, same font." | Text fixes need the exact words repeated |
| 4 | "Same image, but in landscape format for a website banner." | Reuse the approved design |

The details of each result will vary; this table is illustrative. Check the final text letter by letter before printing, because text is much better than before but not flawless. OpenAI also lists limits worth knowing: some art styles, images with many faces, and text in non-Latin scripts can still go wrong.

> **Try it:** Write a prompt for something you need this month, such as an invitation or a header image, using the four-part structure plus a "Use:" line. Then improve it in three single-change follow-ups and compare the first and last versions.

## Why small steps beat long rewrites

TechCrunch notes that many image tools redraw the whole picture when asked for a small change, which is why people learned to start over. The new model is designed to keep lighting, composition and faces consistent across edits, so a short follow-up now tends to change only what you mention. Rewriting the entire prompt throws that advantage away and invites a completely new image.

## How do you edit just one part of an image?

Name the part precisely, say that only that part may change, and list what must stay the same. Location words help: "the mug on the left", "the sky above the roofline", "the second line of text". OpenAI's guide suggests the pattern "change only X, keep everything else the same", and adds an important detail: repeat your list of things to preserve in every follow-up, because small changes can drift over several rounds.

```prompt
In this photo, replace only the two white kitchen chairs with chairs
made of light wood. Keep the camera angle, the room lighting, the floor
shadows, the table and everything else exactly the same.
```

Developers can do the same through the API. The guide's examples call `client.images.edit` with `model="gpt-image-1.5"`, your image file and a prompt, and add `input_fidelity="high"` when faces or fine details must survive the edit. After each edit, compare it with the original side by side. If something you did not mention has changed, name it in the next message ("restore the original table"), or go back to the original file and try again with a clearer preserve list.

## Sources

- Gpt-image-1.5 Prompting Guide: https://github.com/openai/openai-cookbook/blob/fc51ab3afa16/examples/multimodal/image-gen-1.5-prompting_guide.ipynb
- The new ChatGPT Images is here: https://web.archive.org/web/20251217221724/https://openai.com/index/new-chatgpt-images-is-here/
- OpenAI continues on its 'code red' warpath with new image generation model: https://techcrunch.com/2025/12/16/openai-continues-on-its-code-red-warpath-with-new-image-generation-model/