A reliable formula

Image prompting has its own grammar, and a simple formula gets you most of the way. Describe four things, roughly in this order: subject, style, composition, and lighting.

"A red fox curled asleep, watercolor style, centered composition, soft morning light"

Each word is a lever. "Subject" is what; "style" sets the medium and mood (photograph, oil painting, 3D render, anime); "composition" controls framing (close-up, wide shot, centered, rule of thirds); "lighting" sets atmosphere (golden hour, neon, studio, dramatic shadows). Stack them and the model has a clear target.

The dials beyond the words

Most tools expose a few settings that matter as much as the prompt:

• Steps — how many denoising passes. More can mean more detail, but 30–50 is plenty; beyond that you mostly waste time.
• Guidance scale — how literally the model obeys the prompt. Low is loose and creative; high is faithful but can look forced. The middle is usually the sweet spot.
• Seed — the starting noise. Fix it to reproduce or tweak an image; change it to explore fresh variations.
• Negative prompts — list what to avoid ("blurry, extra fingers, text, watermark"). Often the fastest fix for a recurring flaw.
• Aspect ratio — set it deliberately to match where the image will be used.

How to actually work

Don't expect the first image to be perfect. Generate a small batch, pick the closest, then refine — adjust one part of the prompt or one dial at a time so you learn what each change does. Image prompting rewards iteration far more than it rewards the "perfect" first prompt.

Treat the prompt as the description and the settings as the controls. The words say what to make; steps, guidance, seed, and negatives decide how it comes out.

Try this: Take a basic prompt ("a coffee shop") and add the four formula parts one at a time — style, then composition, then lighting — regenerating after each. Watching the image sharpen with every added clause teaches the grammar faster than any reference list.