# What is DALL·E 3, and what changes when ChatGPT writes your image prompts?

Date: 2023-09-21 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-dall-e-3/

**Answer:** DALL·E 3 is OpenAI's new image generator, announced on September 20, 2023, which follows the details of a written description far more faithfully than earlier systems and can even place readable text in images. It is built into ChatGPT, so you can describe an idea in ordinary words, let ChatGPT write the detailed prompt, and then ask for changes in conversation; it reaches ChatGPT Plus and Enterprise subscribers in October.

Yesterday OpenAI announced **DALL·E 3**, the third generation of its text-to-image system. Like its predecessors, it turns a written description, called a **prompt**, into a new picture. The difference OpenAI emphasizes is accuracy: DALL·E 3 is designed to draw what you actually asked for, including the small details that other tools tend to skip.

The second big change is where it lives. DALL·E 3 is built directly into ChatGPT. Instead of learning special prompt tricks, you chat about what you want, and ChatGPT writes the detailed instructions for the image model.

## What OpenAI announced, and when you can use it

DALL·E 3 is in what OpenAI calls a **research preview**, meaning a limited early stage while testing continues. The schedule:

| Who | When |
|---|---|
| ChatGPT Plus and Enterprise subscribers | Early October 2023, inside ChatGPT |
| Developers, through the API | Later this fall |
| Labs (OpenAI's existing image web app) | Later this fall |

The pictures belong to the person who makes them, just as they did with DALL·E 2. OpenAI says you do not need its permission to reprint, sell or put them on merchandise.

OpenAI has not published technical details about how DALL·E 3 is built or trained. Its announcement page is mostly side-by-side examples, which Ars Technica notes were chosen to show the system at its best.

## What is new: following the prompt

Anyone who has used an image generator knows the frustration. You ask for a red bicycle leaning on a blue door, with a cat in the basket, and you get a red door, no cat, or two bicycles. OpenAI says today's systems tend to ignore words or descriptions, which pushes users to learn **prompt engineering**, the craft of phrasing a request in odd, repetitive ways to get the result. DALL·E 3 is meant to make most of that unnecessary.

Two improvements stand out in OpenAI's samples:

- **Detail and composition.** Long prompts with several objects, colors and relationships come out with the parts in the right places. OpenAI says fine details such as hands are also better than with DALL·E 2.
- **Text in images.** Earlier models produced garbled letters on signs and labels. DALL·E 3 can render short phrases legibly. One example shows a cartoon avocado in a therapist's chair, with a speech bubble saying it feels empty inside, and a hole where its pit should be.

OpenAI's own comparison uses the same prompt for both versions: “An expressive oil painting of a basketball player dunking, depicted as an explosion of a nebula.” The DALL·E 3 image is visibly closer to that description.

## How the ChatGPT connection works

Here is the workflow OpenAI describes, step by step:

1. **You describe an idea** in ChatGPT, as briefly or as fully as you like: "A poster for my daughter's lemonade stand, cheerful, summer colors, with the price 50 cents."
2. **ChatGPT writes detailed prompts** for DALL·E 3, filling in style, layout and specifics you may not have thought to mention.
3. **DALL·E 3 generates images** from those prompts.
4. **You ask for changes** in plain words: "Make the lemons bigger and put the price in a yellow circle." ChatGPT adjusts the prompt and DALL·E 3 draws again.

The idea is that ChatGPT acts as a brainstorming partner and translator between everyday language and the precise descriptions an image model needs. It also means the image can draw on the conversation so far, such as a story you were writing together. Ars Technica points out that Microsoft's Bing Chat, which also uses OpenAI technology, has offered images inside a conversation since March, so the idea is not entirely new; the tighter fit with ChatGPT is.

## Safety and artists

OpenAI lists several limits built into DALL·E 3:

- Like earlier versions, it restricts violent, adult and hateful images.
- It is designed to decline requests for a public figure by name.
- It is designed to decline images in the style of a **living artist**.
- Artists and photographers can now ask for their images to be excluded from the training of OpenAI's future image models.

OpenAI says outside experts, known as red teamers, tested the model for risks such as propaganda and misinformation. It is also experimenting with a **provenance classifier**, an internal tool that tries to tell whether an image was made by DALL·E 3, and promises more details soon.

## Why it matters

Image generators have become very good at producing attractive pictures; the weak point has been control. A tool that reliably draws what you describe, including readable words, is far more useful for practical work such as posters, classroom materials, product mock-ups and illustrations for stories.

Putting the model inside ChatGPT also changes who can use it well. People who never learned the special phrasing that other tools reward can simply explain what they want and refine it in conversation. Whether DALL·E 3 matches its showcase examples in everyday use will become clear once subscribers get access in October.

> **Try it:** Write a one-sentence description of an image you would like, then rewrite it as a detailed paragraph covering subject, setting, style, colors, mood and any text. Compare the two. The second version is roughly what ChatGPT will write for DALL·E 3, and it is good practice for any image tool you use today.

## How do you describe an image so DALL·E 3 gets it right?

Be specific about what matters and silent about what doesn't. Name the subject, where it is, the style (photo, watercolor, pixel art, flat illustration), the main colors and the mood. If you want words in the picture, put them in quotation marks and keep them short.

Describe relationships clearly: "a small brown dog sitting to the left of a large orange cat" leaves less room for error than "a dog and a cat." Say what the image is for, such as a poster, a book cover or a slide, because that hints at layout and level of detail.

Then use the conversation. Instead of starting over, ask for one change at a time: brighter lighting, a different angle, fewer objects. Small, clear adjustments are how you steer an image toward what you pictured.

## Sources

- DALL·E 3: https://openai.com/dall-e-3
- OpenAI announces DALL-E 3, a next-gen AI image generator based on ChatGPT: https://arstechnica.com/information-technology/2023/09/openai-announces-dall-e-3-a-next-gen-ai-image-generator-based-on-chatgpt/