What 3,698 public AI image prompts tell us about how people actually prompt in 2026

Mt13 — Sun, 04 Oct 2026 05:39:59 +0000

I run PicGens, a free gallery that pairs AI-generated images with the exact prompt the creator shared publicly. Because every entry is labelled by model and use case, the collection doubles as a small dataset. I just published a new edition of it (October 4, 2026 — 3,698 published prompts), and a few patterns are worth sharing if you write image prompts yourself.

Caveat up front: this is a curated corpus of prompts creators chose to share publicly, not a measure of how many people use each model. Treat shares as "what gets shared with results", not market share.

1. GPT Image 2 dominates what people share

GPT Image 2 (ChatGPT's image model): 56.6% of prompts (2,094 of 3,698)
Nano Banana family (Nano Banana, Nano Banana 2, Nano Banana Pro): 22.8% (843), with Nano Banana Pro alone at 12.6%
Midjourney: 3.2% · Seedream: 2.4%
12.4% carried no reliable model label

Takeaway: if you're looking for prompts that were actually run and shared, ChatGPT-style plain-language briefs are by far the most common format right now.

2. Good prompts are long — about 150 words

The median prompt is 148 words (1,062 characters); the middle half of prompts runs from roughly 560 to 1,800 characters. GPT Image 2 prompts are the longest of the big models (median 169 words); Nano Banana Pro prompts sit at 135.

Takeaway: the one-line "a cat in space, 4k" prompt is not what people who get shareable results are writing. They write a short brief: purpose, subject, setting, camera, light, colour, and what must not change.

3. Text in the image is now normal

39.0% of prompts ask for text inside the image (titles, labels, captions, signage)
30.0% state an aspect ratio explicitly — 4:5 and 9:16 are tied for most common, far ahead of 16:9
18.8% include camera settings (focal length, aperture, ISO or shutter)
73% of the resulting images are portrait-oriented

Takeaway: prompts are being written for phones and social feeds. If your output is for Instagram or Shorts, say the ratio — and if you want words in the picture, quote them exactly.

4. Templates are everywhere

24.6% of prompts contain fill-in slots like [subject] or {{brand}}, and 4.4% are fully structured JSON. Creators increasingly share prompts as reusable templates rather than one-offs.

A template-style brief that works across models:

[Purpose: profile photo / product ad / poster] of [subject + action],
[setting + time of day], [camera distance + lens feel], [light direction],
[colour grade / film look], [aspect ratio]. Keep [anything that must not change].

5. Portraits are the default use case

Share of prompts by use case (a prompt can have several tags, so these add up to more than 100%):

Use case	Share
Portraits	56.3%
Posters & visuals	42.6%
Illustration & 3D	34.4%
Ads & product	23.5%
Brand & logo	10.1%
Video	7.8%

Most shared prompts are about people — headshots, editorial portraits, "keep my face" edits. Product, logo and infographic prompts have far less public prior art, which makes a good template more valuable.

6. Sharing is bursty

Of the 3,231 prompts with a known post date, June and July 2026 alone account for 76.6%. Model launches drive waves of sharing; the long tail in between is thin.

Get the data

The full edition — methodology, metric definitions, limitations and CSV/JSON downloads — is here: AI Image Prompt Statistics 2026. Each edition is dated and immutable (2026-10-04 edition), so you can cite a specific snapshot.

If you want to see the prompts behind these numbers, the ChatGPT photo prompts and Nano Banana prompts collections show each image next to the prompt that made it.

What patterns have you noticed in the prompts that work for you? I'm especially curious whether people have found prompts that transfer cleanly between GPT Image 2 and Nano Banana.

DEV Community: Mt13