A flux prompt from an image: sentences, periods, and two slots that never arrive

A flux prompt generated from an image is a run of sentences rather than a comma list: three sentences at the short setting, five at medium, seven at long, each of them ending in a full stop, with no parameters and no weights anywhere in the line. Two of the slots the tool measures — contrast and vertical light direction — never reach the prompt at any length. And because the sentences are separated by full stops instead of commas, a weight parser reads four of the seven as one single term, which changes how you can edit the line.

Take the darkest of the three files — the fireworks frame — at the longest setting, on the 192 × 192 grid. Seven sentences, 273 characters, not one parameter in it:

A widescreen 16:9 image. Dark, low-key lighting. The dominant colours are navy, grey,
slate, silver. Vivid, saturated colour, cool blue colour temperature. Highly detailed with
intricate texture. Light falling on the centre with darker edges. Moody and dramatic
atmosphere.

That is 273 characters, 39 words, seven sentences, seven full stops and six commas. The Midjourney and Stable Diffusion lines built from the same photograph are 319 and 336 characters of comma-separated tags with zero full stops. Two more files sit behind every count below: a backlit field of grass flowers and a screenshot of a code editor, each decoded and pushed through the same grid. Where a table uses only one of the three, it says which.

The shape, sentence by sentence

Every sentence in the line is a fixed template with one measured phrase dropped into it. The longest setting of the night photograph, sentence by sentence:

Three of the seven sentences are more template than measurement. Sentence 1 carries four words and only the proportions are read off the file; sentence 3 carries eight words and only four of them are colours; sentence 4 is the only place where two measured slots are welded together, and the seam is a comma inside a sentence — which matters later, when a weight parser counts commas.

Sentences 6 and 7 exist only at the longest setting, and the subject sentence does not exist here at all. The template does have one: The frame is dominated by <subject>. It sits between sentences 1 and 2. None of the three photographs produces it, and the reason is in the source rather than in the pictures.

Three sentences or seven: the length ladder

Length does not add words to sentences, it adds sentences. Each step up the ladder adds exactly two, on all three pictures:

Medium adds the saturation and temperature sentence plus the detail sentence. Long adds distribution and mood. Nothing else changes: the framing sentence, the exposure sentence and the palette sentence are identical at all three settings, and a longer setting never rewrites a shorter one, it only appends to it. So the short Flux line for a picture is a strict prefix of the long one, character for character — the same is true of the Stable Diffusion and DALL·E lines, and the Midjourney line is the one exception, because its computed --ar parameter sits at the end of every length, so nothing shorter is ever a prefix of anything longer.

Two slots that never arrive

The tool measures ten slots per picture. Counting how many of them survive into the finished line, across all nine outputs (three pictures × three lengths):

Contrast is measured, written down, and then thrown away by the sentence builder. The night photograph is labelled medium contrast, the backlit field high contrast, the screenshot soft, low contrast — and none of those three phrases appears in any Flux line at any length. The comma-list shapes carry contrast from the medium setting upward, six outputs out of nine. Vertical light direction is worse: brighter upper half is measured for two of the three pictures and reaches the Midjourney and Stable Diffusion lines at the longest setting, and never reaches Flux.

This is not a rendering bug, it is what the sentence template asks for. The builder has seven sentence slots and one of them is conditional, so two measured values have nowhere to go. If the contrast of your picture is the point — a hard black shadow line, a blown-out window — you have to type it into the Flux line yourself, and you have to decide which sentence to put it in, because there is no empty one waiting.

What the subject sentence would look like

The one conditional sentence fires only when the tool judges the picture to be essentially black and white, which it decides from the share of pixels that sit close to grey. None of the three photographs qualifies: their grey share is 0.0117, 0.2016 and 0.0222 against a threshold of 0.85. Feeding the pipeline a synthetic monochrome image — 192 × 192, grey share 1.0000, mean luminance 120.49, edge score 2.20 — does fire it, and the result is worth seeing:

A square 1:1 image. The frame is dominated by black and white photograph. Moderate,
slightly dim lighting. The dominant colours are grey. Monochrome, black and white, neutral
colour temperature. Clean and minimal, smooth surfaces. Evenly lit. Balanced, natural
atmosphere.

The subject sentence is built for a noun phrase, but the only subject the tool ever produces is black and white photograph — so the sentence comes out as dominated by black and white photograph, missing its article. The same value in the comma-list shapes is just a tag, where the missing article cannot be noticed: Stable Diffusion puts black and white photograph second in the line, right behind its quality tags, and Midjourney puts it first, with no sentence around it. Sentence shape makes a defect that the tag shapes hide.

The first two sentences are not about your picture

Because sentences 1 and 2 are templates filled from two of the coarsest measurements, they come out identical for pictures that share nothing else. Three completely different images — fireworks, a backlit field, a screenshot — produce this at every length:

At the short setting, seven of the fifteen words in a Flux prompt are the same seven words for a fireworks display, a sunlit field and a screenshot. The comma-list shapes do something similar — Midjourney repeats 3 of its 9 fragments at short, Stable Diffusion 5 of 11 — but a repeated fragment in a comma list is one tag among fifteen, while a repeated sentence is a third of the line.

The practical consequence is that the front of a Flux prompt tells the renderer almost nothing about your picture, and the back of it tells the renderer the most: sentence 7 is the last thing before the end of the prompt and, at long, it is a shared sentence too. If you have one thing to add by hand, the end of the line is where it belongs.

What a weight parser does to a sentence

The weight syntax in the same repository splits prompts at commas. A Flux prompt has commas — six of them at medium and long — but they are in the wrong places, so the parser's units do not line up with the sentences:

Seven sentences at the longest setting, but still seven terms, and the reason is that four of the six commas sit inside two sentences. Here is the term list the parser actually returns for the night photograph at long — the whole line, cut only where the parser cuts it:

[A widescreen 16:9 image. Dark]
[low-key lighting. The dominant colours are navy]
[grey]
[slate]
[silver. Vivid]
[saturated colour]
[cool blue colour temperature. Highly detailed with intricate texture. Light falling on the
centre with darker edges. Moody and dramatic atmosphere.]

The last term is four sentences and twenty-one words long, and it is a single unit as far as any weighting tool is concerned. The first term is a sentence fragment fused to the first word of the next sentence. If you weight that last term you weight the detail of the picture, where the light falls, and the mood together, in one number. There is no syntax that reaches the mood sentence alone, because the only way to separate terms is a comma and there is no comma in front of it.

Adding the comma by hand costs you the sentence. Four edits to the same base line, each 279 characters, measured through the same parser:

Two facts hold across all nine outputs and all four edits. Nothing is ever out of range: the parser's 0.4–1.6 check returns nothing, because a sentence-shaped prompt has no parameters to be misread as weights — the Midjourney shape, which ends in --ar 16:9, reports one out-of-range item in every one of its nine lines, since the parser reads --ar 16 as a term weighted 9. And the untouched Flux line survives normalisation byte for byte: all nine outputs come back identical, so there is nothing to clean up before the first render. The edits are what cost you, not the line.

Punctuation and capital letters

Two small things follow from the template and are worth knowing before you edit by hand. Full stops are exactly the sentence count — 3, 5 and 7 — with no exceptions in nine outputs, so you can count sentences by counting periods. Commas never appear in sentence 1, and never in the closing distribution and mood sentences; they appear in sentence 2, in the palette sentence, and in the saturation/temperature sentence, which is why the parser's units drift out of alignment with the sentences as the line gets longer.

Capital letters are applied to the first character of a sentence and to nothing else: across all nine outputs there is not one capital letter after the first position of a sentence. That is why the line reads Dark, low-key lighting. and Clean and minimal, smooth surfaces. — phrases written for a comma list, with their first letter raised and a full stop bolted on. A grammar checker would flag them; a renderer reads them as text either way.

Five moves that keep a sentence intact

• Render it once before touching it. Nine of nine outputs come back identical after normalisation, and the 0.4–1.6 check has nothing to report: a sentence shape carries no parameter that could fall outside the band.
• Add the two missing slots yourself if they matter. Contrast and vertical light direction are measured and dropped. A hard shadow line or a lit upper half has to be written into an existing sentence, because no sentence is waiting for it.
• Put your own subject at the end, not the front. The first two sentences are the same for every picture that shares an aspect ratio and an exposure band. The end of the line is the least crowded place.
• Do not expect a weight to hit one sentence. Seven sentences become seven terms, and the last term swallows four of them. Weight the mood sentence, or the palette sentence, and nothing else.
• Count the commas, not the sentences, before you weight. Sentence 2 and the palette sentence carry commas of their own, so a bracket around either one lands on two or four terms rather than one.

What gets asked about a sentence-shaped prompt

Three “People also ask” boxes, captured on 2026-09-19. Each answer below is a count over the nine Flux lines this page is built from, not a rule of thumb.

How to prompt AI for images?

Match the shape of the prompt to the shape of the model. Flux reads sentences, so write sentences: a full stop between ideas, a comma only inside a list. The tool measures ten slots off your file and turns them into three to seven sentences — three at the short setting, five at medium, seven at long — and each step up the ladder appends two sentences without rewriting the ones below it.

What are some good prompts for AI images?

Below is that same fireworks line, unedited, printed so the full stops can be counted: A widescreen 16:9 image. Dark, low-key lighting. The dominant colours are navy, grey, slate, silver. Vivid, saturated colour, cool blue colour temperature. Highly detailed with intricate texture. Light falling on the centre with darker edges. Moody and dramatic atmosphere. — 273 characters, 39 words, seven sentences, no parameters. What it does not contain is as informative as what it does: no contrast, no vertical light direction, no subject.

How to prompt AI to create an image?

Decide which of the two you want from the measured part of the line. If you want a faithful render of the file, paste the sentences untouched: they round-trip byte for byte through normalisation and report nothing out of range. If you want a deliberate change, edit the sentence that carries the measurement you are changing — the exposure sentence, the palette sentence, the detail sentence — and remember that a bracket around the palette sentence splits it into four terms, not one.

Two follow-ups if you are comparing shapes. Where the weights go in a Stable Diffusion prompt writes the same fireworks frame as a comma list with a fixed quality prefix, and what a Midjourney prompt looks like from an image is the only shape of the four that carries a parameter at all. The bracket arithmetic behind either is on how to weight a prompt and write a negative, and which images an image to prompt tool handles badly covers the frames where all four fall over.