AI image prompts: a practical guide
Structure, style references, lighting, aspect ratios and negative prompts — how to get the picture you actually wanted.
6 сентября 2026 г.
Image models do not read minds. They read structure. Here is the structure that works across Midjourney, Leonardo, Ideogram and the rest.
The formula
Subject + action + environment + lighting + style + framing + technical parameters.
"A ceramic coffee cup on a linen tablecloth, morning window light, soft shadows, editorial food photography, shallow depth of field, 3:2".
Be specific about light
Lighting changes the mood more than any style keyword. Golden hour, overcast, single hard key light, neon rim light — pick one and say it.
Style references beat style words
"Cinematic" means almost nothing. A named genre, a decade, a medium or an uploaded reference image gives the model something concrete.
Negative prompts
Say what you do not want: extra fingers, text, watermark, cluttered background. Most tools support this natively and it removes the majority of obvious failures.
Aspect ratio first
Decide the ratio before you iterate — changing it late re-rolls the whole composition.
Iterate in small steps
Change one variable at a time. Changing subject, lighting and style at once tells you nothing about what actually helped.
Licensing
Free tiers usually restrict commercial use. If the image goes into an ad or a product, confirm the rights on your current plan before you ship it.