What are some good ChatGPT prompts for pictures? One sentence of your own

What was measured. Forty real files — 39 stock photographs at 640 × 360 and 800 × 450, plus one code screenshot — box-downsampled to the 192 × 192 grid the live tool measures on, then composed into all four shapes at all three lengths. No model was called and no image was generated; every number below is arithmetic over pixels and over the strings the composer returns, run on this machine.

Here is the whole answer in one line. The generated sentence is a description of the photograph, not an instruction about the picture. It knows the light was low-key and the palette ran to charcoal and slate; it does not know whether the frame holds fireworks, a bicycle or a person. So a good prompt for a picture is the sentence you write about the subject, followed by the sentence the file writes about the light. Below is what is in that second sentence, how long it is, how it grows, and what it costs you in room.

The sentence, before you add anything to it

This is a real output — a strongly backlit field of grass flowers, 640 × 360, measured on the 192 × 192 grid, longest setting, DALL·E shape:

A widescreen 16:9 image, dark, low-key lighting, with a palette of black,
charcoal, silver, white. Vivid, saturated colour, slightly warm tone, moderate
detail and texture. Moody and dramatic atmosphere.

That is 203 characters: the shortest long line the forty files produced. Nothing in it is a parameter, a weight or a tag. It is three plain sentences, each closed with a full stop, in the order the composer appends them: framing and exposure and palette, then saturation and temperature and detail, then mood.

The sentence count is fixed by the length setting and by nothing else. Not one of the 40 photographs produced two sentences at the short setting or four at the long one, and not one produced a fragment: every line ends on a full stop. Compare that with the Flux shape run over the same forty files, which returns three or four sentences at short, five or six at medium and seven or eight at long, because it writes one sentence per measured value.

Three settings, one line that only grows

The most useful thing about this shape is that the three settings are not three different drafts. They are the same line with clauses appended. Take the short line and ask a blunt question: is it a literal prefix of the medium line for the same photograph?

So the length setting adds characters and never rearranges them: 74 to 101 characters (median 89.5) going from short to medium, and 26 to 33 (median 31) going from medium to long. Read the long line and you are reading the short sentence, then a second sentence, then a third. Nothing you already read gets rewritten behind your back, which makes the line cheap to edit: you can cut the last sentence off a long output and the remainder is exactly what the medium setting would have given you.

The last clause has four possible endings

Because clauses are only ever appended, the final sentence of a long line is always the mood clause—and across forty photographs that clause turned out to be one of just four strings.

That is the whole range: 25 to 32 characters, four strings, one of them landing on 16 of the 40 photographs. The middle sentence is where the variety actually lives — at the medium setting the second sentence came out 24 distinct ways in 40 photographs, running 73 to 100 characters, because it carries three measured values (saturation, temperature, detail) rather than one. If the mood of your generated prompt is wrong, that last clause is four words you can replace in one stroke; if the detail is wrong, you have a wider sentence to argue with.

One capital per sentence, and six hundred colons that all belong to a ratio

Counting capitals is a fast way to see how mechanical the line is. This shape capitalises exactly one word per sentence — one capital at short, two at medium, three at long, in 40 out of 40 files — and every one of those capitals is the first letter of a sentence. The composer does run a capitalising function over two clauses (the saturation clause and the mood clause), but in this shape both of those clauses open a sentence, so the capital lands where a capital would go anyway.

Run the same colon count across all four shapes and 480 lines and you get 600 colons, every one of them inside an aspect ratio. Midjourney carries two per line — one in the framing phrase widescreen 16:9, one in --ar 16:9 — and the other three shapes carry the single one in the framing phrase. Not one colon in 480 lines is a weight. There is no weight syntax here to collide with, which is the practical reason this shape survives being pasted into a chat box untouched.

What one word costs you, in pixels

The forty files carry between 230,400 and 360,000 pixels each; the tool measures all of them on the same 36,864-pixel grid, a reduction of 6.3 to 9.8 times before anything is described. Divide the original pixel count by the number of words in the finished line and you get the exchange rate between your photograph and your prompt.

At the short setting, roughly fourteen thousand pixels of your photograph are behind every word you get back — and a median 27.5% of that short line is the four colour names. Stretching to the long setting buys more words per pixel but cheaper words: the pixel cost per word nearly halves while the colour names drop to 12.4% of the line. That is the trade the length setting actually makes.

The vocabulary is small enough to state outright. Across all 120 lines there are 72 distinct words, and seven of them appear in every single line: A, image, of, palette, widescreen, with, and the 16:9 ratio inside the framing phrase. Seven words that are true of every photograph in the set, because they describe the shape of the frame rather than anything in it.

The parts you have to write, and how much room they take

Now the part that answers the question. Below is what the generated line never carries, counted over 120 lines of this shape and, where another shape does carry it, that shape’s count for contrast.

Three of those are worth spelling out. There is no aspect token, so the size and shape of the image you want are yours to state. There is no weight syntax, so emphasis is yours to state in words. And there is no word for the thing in the picture: over the same forty files the slot that could name a subject stayed empty on 36 of them, and the four times it fired it wrote black and white photograph — a fact about the file, not about the picture. That count is laid out in full in what the best AI prompts for images have in common; the consequence is what matters here.

So the division of labour is measurable. Write your sentence, then let the measured sentence ride behind it:

1. Your sentence goes first, and it names a thing. Eight words is enough to be 28% of the finished prompt at the short setting and 15.2% at the long one. Nothing in the generated line will name the subject for you.
2. Use the long setting if you want the picture to leave a mark. Short collapses fastest: the median pixel cost per word is 14,400 at short against 7,432 at long, and the four colour names shrink from 27.5% of the line to 12.4%.
3. Overwrite the last clause when the mood is wrong. It is one of four strings, 25 to 32 characters long, and it is appended last, so replacing it changes nothing else in the line.
4. Do not paste the line into anything that expects a weight. Six hundred colons across 480 lines, all of them inside an aspect ratio; zero brackets, zero AND, zero negative. A parser looking for (term:1.3) will find nothing and will read the whole sentence as one term.
5. Say the size yourself. The only ratio in the line is the one measured off your file. If you want a square or a portrait render, that instruction does not exist in the generated text.

One thing worth settling before you write any of it: the line is a description of the picture you already have, not a request for a new one, so how to prompt ChatGPT for images when the generated line contains no verb comes down to the sentence you add yourself. Across 480 lines from these same forty files, twelve imperative verbs, nine pronouns and six modals were each searched for as whole words, and not one of them appeared a single time.

Forty files, four hundred and eighty lines

Forty files — 39 stock photographs at 640 × 360 and 800 × 450, plus one code screenshot — were decoded on this machine and box-downsampled to 192 × 192, the grid the live tool measures on, so every figure matches what the site would tell you for the same file. Each was described once and composed into four shapes at three lengths: 480 lines. Sentence counts come from splitting on full stops; the prefix test is a literal string comparison between two length settings of the same file and shape; the pixel-per-word figures divide the original pixel count by the word count of the finished line; the colon and capital counts are character counts over those 480 strings; the palette share is the length of the palette clause over the length of the whole line. No model was called, no image was generated, and nothing here is taken from another site’s wording.

For the shape itself and why it is built as sentences, see how a ChatGPT image prompt is put together; for the closed vocabulary behind every phrase in the line, see what an image describer actually writes; for the measurements each phrase is derived from, see what an image to prompt tool measures; and for turning a picture into something you can paste, see how to generate a prompt from an image.