What 9,599 Real AI Image Prompts Actually Contain

What we measured

Most prompt advice is one person's habit generalised into a rule. This is a count.

We took 9,599 AI image prompts from the LocalBanana gallery — every published item with a prompt longer than 20 characters — and measured what they actually contain: how long they are, how they're structured, and which technical vocabulary shows up. Each prompt is stored with the model that ran it and the image it produced.

How long should an AI image prompt be?

The median prompt is 83 words. The mean is 128, which tells you the distribution has a long right tail — a substantial group of people write far more than the middle.

Median engagement rises about tenfold from the shortest bucket to the longest. The jump is not gradual — it happens between 79 and 80 words, where median views go from 5 to 19.

That threshold is the practical takeaway. Prompts under about 80 words tend to describe a subject. Prompts over it have room to also specify light, lens, composition and mood — which is when the model stops guessing.

Does prompt structure matter more than length?

Yes — and this was the most surprising result in the corpus.

13% of prompts are written as structured JSON rather than prose: an object with keys like scene, subject, lighting, camera, style. The rest are sentences.

An 11× gap in median engagement. JSON prompts are also longer on average (278 words vs 105), so the obvious objection is that we are just re-measuring length.

We controlled for it. Within every length band, structured prompts still win:

Same length, different structure, consistently different outcome. The effect is strongest in the 80–149 band, where structuring a medium-length prompt is associated with roughly 8× the median engagement of writing the same amount as prose.

Why it plausibly works: a JSON object forces you to name your slots. You cannot leave lighting empty without noticing. Prose lets you write a beautiful sentence that never mentions where the light comes from.

{
  "subject": "young woman, mid-20s, shoulder-length dark hair",
  "wardrobe": "oversized cream knit sweater",
  "setting": "sunlit café window seat, mid-morning",
  "lighting": "soft window light from camera left, warm bounce",
  "camera": "85mm, f/1.8, eye level, shallow depth of field",
  "mood": "quiet, unposed, candid",
  "style": "editorial photography, fine film grain"
}

CCD Flash Hotel Lazy Heel-Off Moment

Nano Banana Pro

Prompt▾

{"portrait_prompt":{"subject":{"description":"Based on <User Portrait>, a young woman lying on white sofa, seen from directly above, relaxed lazy pose, looking up at camera","features":{"hair":"Messy tousled hair spread on sofa cushion","expression":"Relaxed, slightly tired, intimate late-night vibe"},"pose":"Lying on back on sofa, one hand reaching down to touch her heel still on her foot, fingers hooked on the back of her high heel about to slip it off, other hand resting near her head or in her hair, legs slightly bent, the moment of taking off shoes"},"attire":{"type":"Based on user-uploaded clothing reference","details":["Black blazer dress","Sheer black stockings","Heels removed and held in hand"],"color":"As per clothing reference"},"composition":{"shot_type":"Top-down overhead shot, full body visible","focal_length":"28-35mm wide angle, compact camera","camera_angle":"Directly overhead, bird's eye view, looking straight down","camera_height":"Standing directly over subject","framing":"Subject fills frame diagonally, some sofa and floor visible around edges, casual snapshot framing","aspect_ratio":"9:16"},"lighting":{"type":"Direct on-camera flash, harsh and unflattering","direction":"From camera position straight down","mood":"Hard flash creating harsh highlights on skin and fabric, distinct shadow cast on sofa beneath subject, flash reflection in eyes, overexposed hot spots on forehead and cheekbones"},"color_palette":{"film_simulation":"CCD compact camera with flash","style":"Early 2000s point-and-shoot digital camera, direct flash, high ISO noise, warm incandescent ambient mixed with cool flash, slightly muddy colors","tones":["Cool flash on skin","Warm ambient in shadows","Blown highlights"],"grade_notes":["highlight: BLOWN OUT, clipped whites on skin and sofa","shadow: warm ambient light visible in corners","contrast: harsh from flash, deep blacks where flash doesn't reach","noise: VISIBLE high ISO digital noise, color noise in shadows","sharpness: slightly soft, compact camera lens quality","flash falloff: bright center, darker edges"]},"environment":{"setting":"Hotel room interior, nighttime","background":"White sofa, marble floor partially visible, window with city lights or curtains in background, indoor ambient warm lighting mixed with flash","atmosphere":"Late night, private moment, after-party vibe"},"mood":"Intimate snapshot, late night hotel room, caught on compact camera with flash, raw and unpolished","technical_tags":["top-down overhead shot","direct on-camera flash","CCD compact camera","high ISO digital noise","visible noise grain","blown highlights","harsh flash lighting","early 2000s digital photo","point-and-shoot aesthetic","NOT professional","candid snapshot"],"character_reference":"Based on user-uploaded character reference","clothing_reference":"Based on user-uploaded clothing reference","negative_prompt":"professional studio lighting, softbox, natural light only, side lighting, rim light, no flash, perfectly exposed, no noise, clean image, magazine editorial, too polished, AI look, plastic skin"}}

Copy

A structured prompt in the wild — every slot filled, nothing left to the model's discretion.Open in LocalBanana

Which lighting keywords do people actually use?

Lighting is the vocabulary people reach for most after the subject itself. Share of all 9,599 prompts containing each term:

Two things stand out. rim light at 3.9% is far more common than its reputation suggests — it has quietly become standard vocabulary rather than an advanced technique. And flash at 3.5% reflects how thoroughly the direct-flash look took over: people are asking for the aesthetic of a cheap camera on purpose.

Notably rare: top light (0.05%), candlelight (0.2%), blue hour (0.2%). Not because they don't work — because few people think to ask.

Which camera settings show up most?

One in seven prompts mentions depth of field — the single most common technical instruction in the entire corpus. Among focal lengths, 85mm leads at 4.9%, roughly 1.4× the rate of 35mm and 2× 50mm, which matches the portrait skew of the library.

Apertures are much rarer than focal lengths: f/1.8 appears in 2.0% of prompts, f/2.8 in 1.3%. Most people name the lens but not the opening — which is odd, since the aperture is what actually produces the shallow depth of field 11% of them asked for by name.

What do different models get asked for?

The same corpus, split by which model ran the prompt:

People route photographic work to Nano Banana, text and layout work to GPT Image, and write the least technical prompts of all for Midjourney. We unpack that split in Nano Banana vs DALL·E.

Five things to change in your next prompt

1. Get past 80 words. That is where median engagement steps up, and it is roughly the length at which a prompt can carry subject, light, lens and mood at once.
2. Structure it. Even a plain list of labelled slots — subject, setting, lighting, camera, mood, style — reproduces most of the JSON effect. The point is being unable to leave a slot empty.
3. Name the light. Six per cent of the corpus says studio lighting and most of the rest says nothing at all about light. Saying anything specific puts you ahead of the median.
4. Give the aperture, not just the lens. Eleven per cent ask for shallow depth of field; only 2% specify f/1.8. Name the opening and stop hoping.
5. Steal a rarely-used term. blue hour, candlelight, side lighting and top light are each under 0.5% of the corpus. They work fine — they are simply under-asked.

FAQ

How long should an AI image prompt be?

The median in a 9,599-prompt corpus is 83 words, and median engagement steps up sharply above 80 words — from 5 to 19 median views. Aim for 80–150 words for most work: enough to specify subject, lighting, camera and mood without padding.

Do longer prompts produce better AI images?

Longer prompts correlate with higher engagement in our data, but length is not the mechanism — specification density is. A 200-word prompt of adjectives will not beat a 90-word prompt that names the light, the lens and the composition.

Should I write AI prompts in JSON?

Structured prompts show markedly higher median engagement than prose at the same length — 141 vs 17 median views in the 80–149 word band. Strict JSON syntax is not the point; labelled slots are, because they make an unfilled slot visible.

What is the most common technical term in AI image prompts?

Depth of field, in 14.5% of prompts, followed by close-up (9.3%) and bokeh (5.4%). Among lighting terms, studio lighting leads at 6.5%.

Which focal length is used most in AI prompts?

85mm, in 4.9% of prompts — ahead of 35mm (3.5%) and 50mm (2.4%). That ordering reflects a portrait-heavy corpus; a landscape-heavy library would likely invert it.