How to write image generation prompts

Image models such as ChatGPT (GPT Image), Midjourney, Stable Diffusion or Gemini turn a description into a picture. They follow concrete visual words far better than abstract ones: "a lighthouse at dusk, warm window light" works; "a hopeful image" leaves almost everything to chance.

The parts of an image prompt

1. Subject: what is in the picture, with the details that matter: age, material, colour, count.
2. Action and setting: what is happening and where.
3. Shot and composition: close-up, wide shot, from above, centred, rule of thirds.
4. Style: photograph, watercolour, 3D render, flat illustration, a period or medium.
5. Lighting and colour: golden hour, studio light, overcast; a palette.
6. Format: aspect ratio, and whether you need empty space for text.
7. What to avoid: text, watermarks, extra limbs, clutter. Some tools have a separate negative prompt field for this.

A template you can reuse

A {{shot}} of {{subject}}, {{action}}, {{setting}}. {{style}} style, {{lighting}} lighting, {{palette}} colour palette. {{aspect}} aspect ratio. Avoid: {{avoid}}.

Filled in, it might read:

A close-up of an old ceramic teapot, steam rising, on a wooden kitchen table by a window. Soft watercolour style, early-morning lighting, muted green and cream colour palette. 3:2 aspect ratio. Avoid: text, logos, extra cups.

Fill the variables with the template filler, then paste the result into your image tool.

Iterate one change at a time

• Generate a few images, keep the best, and change one detail, such as lighting, before generating again.
• Put the most important detail early. Many models weight the start of a prompt more heavily.
• Avoid copying living artists' names. Describe the qualities you want instead: loose brushwork, high contrast, muted colours.
• Keep the versions that worked. The prompt diff shows what you changed between them.

For the general structure of text prompts, read what makes a good prompt.