Four levers
Most weak generations describe only the subject. Strong prompts specify four things:
- Subject — what is in frame, and what it is doing.
- Composition — shot type, angle, framing, what occupies the background.
- Lighting — direction, quality, time of day, colour temperature.
- Style — medium, era, render engine or reference aesthetic, plus mood.
A prompt skeleton
[subject, doing something specific], [shot type and angle], [lighting], [background], [style and mood], [aspect ratio or use case]
Example: "A ceramic pour-over coffee set on a linen cloth, three-quarter product shot at eye level, soft north-facing window light from the left with gentle falloff, uncluttered warm-grey background, editorial food photography, shallow depth of field, calm and premium."
Iterating visually
Change one lever per run. Same subject, harder light. Same light, tighter crop. That is how you build a usable image instead of collecting twelve unrelated ones. If your tool exposes a seed, fix it so the change you made is the only thing that differs.
Editing beats regenerating
When 80% of an image is right, do not reroll it. Use AI Image Editor to change the region you dislike, Background Remover to isolate a subject, or Image Upscaler to add resolution for print. Regeneration throws away the composition you already liked.
Known limits
Text inside images, hands, reflections, precise brand logos and fine repeating patterns remain unreliable across every current model. Plan around them: add text in a design tool afterwards, crop hands out, and never rely on a generated logo. Always inspect at full size before commercial use — artefacts hide at thumbnail scale.