AI Image Prompt Generator

Order matters more than word count

Image models read your prompt as a weighted sequence, not a checklist. Tokens near the front carry more influence over composition and subject; tokens near the back nudge finish, mood and grain. That's why "watercolor portrait of a fisherman" and "portrait of a fisherman, watercolor" produce genuinely different pictures — the first commits to a medium before it even decides who the fisherman is.

A reliable order that works across all three major models:

• Subject — who or what, plus one distinguishing trait ("an elderly fisherman with a scarred hand")
• Action or state — what they're doing, or explicitly "standing still, facing camera"
• Environment — where, and how much of it we see
• Style / medium — photograph, oil painting, 3D render, specific movement
• Lighting — the single biggest lever on mood
• Camera / lens — focal length, angle, depth of field
• Technical flags — aspect ratio, quality, version parameters

Front-load what you refuse to compromise on. If the ochre coat matters more than the harbour, the coat goes earlier. This ordering logic is the same instinct behind structuring a text prompt well — see how to write AI prompts if you want the general version of the principle.

How much detail is too much

There's a ceiling. Past roughly 40-60 meaningful words, each additional descriptor dilutes the rest rather than adding to it. You'll see it happen: you add "wearing a brass compass on a leather cord" and suddenly the lighting you'd nailed goes flat. The model is redistributing attention, not stacking it.

Two failure modes to watch for. The first is contradiction: "soft diffused light" plus "harsh shadows" plus "golden hour" gives you a muddy average of three lighting setups. The second is redundancy: "highly detailed, intricate, ultra detailed, 8k, sharp focus" is one instruction typed five times, and it eats budget that could have gone to the fisherman's hands.

Cut anything that doesn't change the picture. If you removed the word and couldn't tell, it wasn't doing work. Run a draft through the prompt analyzer when you're not sure which parts are carrying weight, and count tokens if you're pushing a model's hard limit.

The same subject, three different models

Take one idea — a fisherman on a harbour wall at dawn — and watch how the phrasing has to change.

Midjourney rewards compressed, comma-separated imagery and responds strongly to art-direction vocabulary. Parameters do the structural work.

DALL·E 3 rewrites your prompt internally, so terse tag-soup gets expanded in ways you didn't choose. Write it as a sentence and it stays closer to your intent.

Stable Diffusion takes tags happily, respects weighting syntax like (ochre coat:1.3), and — crucially — depends on a negative prompt to do half the job. Build that half in the negative prompt tool.

Model-specific phrasing is worth practising directly: Midjourney, Stable Diffusion and DALL·E each have their own generator here.

Lighting and lens language that actually moves the needle

Most descriptive words are decorative. A small set genuinely changes pixels, because they're heavily represented in training captions.

Lighting terms that work: golden hour, backlit, rim light, softbox, overcast, harsh midday sun, single candle, neon reflections, chiaroscuro. Terms that mostly don't: beautiful lighting, perfect light, dramatic atmosphere.

Lens and camera terms that work: 14mm (wide distortion, looming), 35mm (documentary, natural), 85mm (portrait compression, flattering), macro, low angle, top-down, bokeh, long exposure. Terms that mostly don't: professional camera, high quality photo.

The pattern: specific, physical, nameable beats evaluative. "Rim light from behind, subject in silhouette" tells the model where photons come from. "Stunning lighting" tells it nothing.

Aspect ratio, composition, and changing one thing at a time

Aspect ratio isn't a crop applied afterwards — it's decided before generation, and it changes what the model puts in frame. Ask for --ar 9:16 and you'll get a standing figure with headroom; ask for --ar 21:9 and you'll get environment, negative space, and a smaller subject. If your subject keeps getting cut off, the ratio is often the culprit rather than the description.

Pair ratio with an explicit framing word: close-up, medium shot, full body, wide establishing shot, overhead. Without one, the model guesses.

Then iterate properly. Change one variable per generation and keep the seed fixed if your model supports it. Swap 35mm for 85mm and nothing else. Swap overcast for golden hour and nothing else. Four disciplined rounds will teach you more than forty scattershot ones, and you'll end up with a reusable template rather than a lucky accident. Save the winners into your prompt library, and read the image prompts guide for the longer walkthrough of this loop.