Image prompting has a small number of high-leverage words and a very large number of words people add out of habit that do nothing. Learning which is which is most of the skill.
The five that matter
1. The medium. The highest-leverage word in any image prompt.
photograph · oil painting · flat vector illustration · pencil sketch · 3D render · watercolour · charcoal
Naming it rules out entire visual territories at a stroke. Without it, the model picks, and it usually picks a glossy digital-art default nobody asked for.
2. The lighting. Second highest, and the one most often omitted.
soft window light from the left · harsh midday sun · single candle, deep shadows · overcast, flat · backlit at golden hour · studio softbox
Lighting drives perceived quality more than any other descriptor. It is the difference between a snapshot and a photograph.
3. The composition. Where things are and what the camera is doing.
close-up · wide establishing shot · shallow depth of field · subject centred · shot from below · background falls out of focus
4. The subject, concretely. Nouns, not adjectives. "A chipped enamel mug on a scratched oak table" beats "a beautiful rustic mug."
5. The exclusions. What you do not want: no text, no watermark, uncluttered background, no people.
What does nothing
- "Stunning, beautiful, amazing, breathtaking." Pure noise.
- "8k, 4k, ultra HD, high resolution." Does not change output resolution.
- "Masterpiece, award-winning, trending on artstation." Vestigial from older models and largely inert now.
- Long comma-separated tag soup. Modern models read sentences. Twenty keywords dilute rather than sharpen.
- Naming living artists. Often blocked, legally risky, and less controllable than describing the style — see AI art generators.
Building one up
Watch what each addition does:
a coffee cup
Generic. A mug, centred, ambiguous lighting, café-stock aesthetic.
photograph of a white ceramic coffee cup
Medium set. Illustration and painterly output are now excluded.
photograph of a white ceramic coffee cup on a light oak table, soft window light from the left, shallow depth of field
Medium, subject, surface, lighting, lens. This is where most of the quality gain happens — and note that none of it is adjectives.
…, no text, no logo, background softly out of focus
Exclusions last. A usable product-style image, without a single superlative.
Describing style without naming an artist
Break the style into its attributes:
- Instead of a painter → thick impasto, muted earth palette, visible canvas texture
- Instead of a photographer → available light, shallow depth of field, slight motion blur, grainy film stock
- Instead of an illustrator → bold outlines, limited flat palette, no gradients
More controllable as well as safer: you can adjust one attribute at a time instead of hoping a name carries what you meant.
Prompts for common jobs
Product shot
photograph of [product] on a plain light grey seamless backdrop, soft even studio lighting, slight reflection beneath, centred, no text, no props
Blog header
wide photograph of [scene], natural light, generous empty space on the left for text overlay, shallow depth of field, muted colours
Icon or spot illustration
flat vector illustration of [subject], three-colour palette, bold outlines, no gradients, white background, centred
Atmospheric background
[scene], out of focus, low contrast, muted, no subject in sharp focus
Habits worth more than the wording
Generate four, pick one. These models are not deterministic. Selection beats prompt-perfection.
Iterate, do not restart. If an image is 80% right, edit it. Regenerating from scratch loses the 80%.
Change one thing at a time. Rewriting the whole prompt makes it impossible to know what helped.
Keep prompts that worked. With a note on which model produced them — the same prompt behaves differently across models.
Common questions
What makes a good AI image prompt? Medium, lighting, composition, concrete subject, and explicit exclusions. Adjectives and quality words add almost nothing.
Does saying "8k" or "4k" help? No. It does not change the output resolution.
Should I use long keyword lists? No. Modern models read sentences; long tag lists dilute the prompt.
How do I get a specific art style? Describe its attributes — palette, texture, line, lighting — rather than naming an artist.
Why do I get different images from the same prompt? They are not deterministic by design. Generate several and choose.
How do I get the same character twice? You cannot, from text alone. Generate once, then edit that image.
Do prompts work across different image models? The principles do. Specific phrasings often do not — keep a note of which model produced a result you liked.