# How AI image generation works

A plain-language look at how image models turn your words into pictures, and why results vary.

You don't need to know how image models work to use them. But a little background helps you understand why prompts matter and why the same prompt gives different pictures.

## Two main approaches

**Diffusion models** start from random noise, like TV static. Step by step, they remove the noise and shape it into a picture that matches your prompt. Think of a sculptor slowly revealing a figure from a block of stone.

**Autoregressive models** build a picture piece by piece, a bit like how a chat model writes a sentence word by word. Because they share their roots with language models, they tend to be good at following detailed instructions and at writing readable text in images.

Many modern models mix ideas from both. For you, the practical difference shows up in the results: some models are better at realism, others at layout, text or edits. See [Choosing a model](/guide/choosing-a-model).

## Why prompts matter

The model learned from huge numbers of images paired with descriptions. Your prompt points it toward the kind of images it has seen. That's why:

- **Specific words work better than vague ones.** "Golden hour side light" tells the model much more than "nice light".
- **Style words change a lot.** "Watercolor", "35mm film photo" and "3D render" pull the picture in very different directions.
- **Missing details get filled in.** If you don't mention the background, the model picks one for you.

## Why results vary

Each generation starts from a different random point. Same prompt, same model, different picture. This is useful: it's how you get options.

- Create 2 to 4 images at once to compare variants.
- If one variant is almost right, open it and use **Reuse prompt and settings** to try again, or **Edit this image** to fix a detail.

![A floating fantasy castle with dragons](/showcase/g06.webp)

## What models find hard

- Exact counts of many small objects
- Hands and fingers in complex poses (much better in recent models)
- Long passages of text
- Keeping the same face across separate generations without a reference

For the last one, see [Keeping a character consistent](/guide/character-consistency).

## Related

- [Complete prompting guide](/guide/prompting-guide)
- [Aspect ratio and number of images](/guide/aspect-ratio-and-variants)