How to Write Better AI Image Prompts
AI image generators can only work with what you tell them. A vague prompt like "a cat" gives you a random cat. A good prompt describes the scene the way a photographer or illustrator would plan it. Here's a simple framework that works in every major tool.
The six building blocks
1. Subject
Start with what the image is about, and be specific. "A ginger cat" is better than "a cat". "A sleepy ginger kitten curled up on a knitted blanket" is better still. Name the action, the expression and anything the subject is holding or wearing.
2. Setting
Where is it? A rainy Tokyo street at night, a sunlit kitchen, a foggy pine forest. The background sets the tone for everything else.
3. Style or medium
Say what kind of image you want: a 35mm film photo, a watercolor illustration, a 3D render, an oil painting, a flat vector icon, an anime still. This single phrase changes the result more than almost anything else.
4. Lighting
Lighting creates mood. Try golden hour, soft window light, harsh midday sun, neon glow, studio softbox, or dramatic rim light.
5. Composition and camera
Close-up, wide shot, bird's-eye view, low angle, centered and symmetrical. For photo styles, lens details help: an 85mm portrait lens gives a blurred background, a wide-angle lens exaggerates space.
6. Color and mood
Muted pastel tones, high-contrast black and white, warm autumn palette, cold blue shadows. Add a mood word too: calm, eerie, joyful, nostalgic.
Putting it together
Here's a prompt built from all six blocks:
A sleepy ginger kitten curled up on a chunky knitted blanket beside a rainy window, soft natural window light, shallow depth of field, 85mm lens, warm cozy color palette with muted creams and oranges, calm peaceful mood, 35mm film photo
Tips for each generator
Midjourney
Midjourney responds well to short, vivid descriptions followed by style keywords. Use parameters at the end, such as --ar 16:9 for widescreen images. Avoid long lists of negatives in the main prompt; describe what you want instead.
Stable Diffusion
Stable Diffusion models usually prefer comma-separated tags, with the most important ones first. Quality tags like "highly detailed" help on many models, and a separate negative prompt (for example "blurry, low quality, extra fingers") removes common problems.
Flux
Flux understands full sentences very well. Write two to four clear sentences, and put any text you want in the image inside quotes, like a sign that reads "OPEN".
DALL·E and ChatGPT
Natural language works best. Describe the scene as if explaining it to an artist, and ask for changes conversationally afterwards.
Common mistakes
- Too many styles at once. "Watercolor, 3D render, photorealistic" confuses the model. Pick one.
- Contradictions. "Bright sunny night" forces the model to guess.
- Burying the subject. Put the most important thing first.
- Expecting an exact copy. Prompts guide the model; they don't duplicate an image. Iterate and adjust.
Learn faster from real images
One of the quickest ways to improve is to study how an image you like translates into words. Upload it to our image to prompt generator, read the prompt it writes, then edit it and see how each change affects the result.