AI image prompt breakdown: see the DNA in your prompt

How to use Image Prompt DNA

1. Paste a prompt. Flags such as --ar 16:9, weights such as (word:1.3), a Negative prompt line and plain sentences all work. You can also load a sample.
2. Read the genes. The annotated prompt labels each span with its gene. The Genes table shows the value, its source and any weight. Check the list of checks below it.
3. Keep a gene fixed. Setting, camera, lighting, style, colour and mood start free to change. Tick a gene's lock to keep it fixed. Subject, action, quality words and detail have no alternatives, so they stay fixed.
4. Choose a mode and generate. One change at a time gives one gene per variant. The pair grid crosses two genes. The seeded walk changes one to three genes per variant.
5. Export the result. Copy the variants, download a TXT or CSV list, or save the JSON DNA card with the genes, locks, seed and dialect.

What the nine genes are

The tool reads a prompt as a set of genes. A gene is one kind of instruction an image model can act on: what the picture shows, where it happens, how it is framed, how it is lit, what medium it imitates, which colours it uses and what feeling it carries. There are nine genes plus a tenth bucket for everything else. Parameters and a negative prompt sit outside the genes, and they are kept exactly as you typed them.

A gene gets its text in two passes. First, the tool searches the prompt for known phrases such as "golden hour" or "oil painting" in a fixed vocabulary. Where phrases overlap, the longest one wins, so "golden hour light" is read as one lighting phrase and not as "golden hour" with a loose word after it. Second, the first clause gives the subject: the text before the first comma, cut at a prose word such as "in" or "on". Every other word that is not part of a phrase becomes detail. The tool does not guess beyond these rules.

The vocabulary follows the kinds of modifier that prompt writers use. Oppenlaender's taxonomy describes six types of prompt modifier used by practitioners: subject term, style modifier, image prompt, quality booster, repeating term and magic term (Oppenlaender, arXiv 2204.13988). This tool reads subject terms, style modifiers and quality boosters directly, and it flags repeated terms. It cannot read an image prompt, which needs the picture itself, or a magic term, which is defined by surprise and needs a judgement the tool cannot make.

Other tools on this site cover neighbouring jobs. Idea to Agent Spec is one, and the LLM Token Counter counts prompt tokens and compares API costs for language models.

How the three dialects are read

Image prompts come in three shapes, and the same words can mean different things in each. The tool reads the shape first and then the genes. If the guess is wrong, choose the dialect from the menu and the same text is read again with that dialect's rules.

This tool keeps flags as typed and makes no claim about what any flag does. The Midjourney documentation could not be read while this page was prepared, so no claim is made about its flags or versions. The markers in the table are what the tool looks for. They are not a statement about how any model reads them.

Weights are the clearest case of tool-specific syntax. In the AUTOMATIC1111 web UI, (word:1.5) raises the attention given to that word by a factor of 1.5, and square brackets lower it. The same wiki says NovelAI uses a multiplier of 1.05 and braces instead, so one tool's weight does not carry over to another (AUTOMATIC1111 wiki, Features). The tool marks any weight above 1.5 as a possible issue. It does not treat 1.5 as a limit.

Negative prompts are the second clear case. The AUTOMATIC1111 wiki describes its negative prompt as a prompt of things the model should avoid. FLUX 3 Image has no negative prompt field at all, so Black Forest Labs advises describing the scene you want, for example "an empty promenade" instead of "no crowds" (Black Forest Labs, Technical Parameters). Its FLUX.2 [pro] and [max] rows read "No negative prompts supported" (Black Forest Labs, Prompt Reference). A negative line may therefore be ignored, depending on the model, and the tool says so when it sees one.

How controlled variations work

A variation keeps most of the prompt still and changes one thing. OpenAI's image prompting guide gives the same advice: "Refine one thing at a time and inspect the result" (OpenAI, Image prompting guide). When only one gene moves, a change in the picture is easier to trace to that gene. A gene can hold more than one phrase, and a seed can change the picture even when the words do not; Liu and Chilton saw this in their VQGAN+CLIP runs.

Three modes are available.

• One change at a time. Each variant differs from the base in exactly one gene that is not locked. The label names the gene and the change, for example "lighting: golden hour → blue hour". Each one-change variant is built by changing a single gene, and the test suite checks this.
• Pair grid. Two genes are crossed, so every cell changes both. Use it to see whether two choices interact. The grid is at most six columns wide, and the count sets its overall size.
• Seeded random walk. Each variant changes one to three genes that are not locked, chosen by a seeded generator. The same seed and the same genes give the same list.

Locks decide what can change. Setting, camera, lighting, style, colour and mood start free to change, and ticking a gene's lock keeps it fixed. The subject has no alternatives in this tool, so it stays fixed: a variation that changed the subject would be a different image. A gene with no pool of alternatives, such as quality words, is fixed too. A locked gene never changes in any mode, and the tests check that.

The seed belongs to this tool. It makes the choice of alternatives repeatable, so you can share a list. It is not a model seed. The AUTOMATIC1111 wiki warns that after a change to its prompt parser "your saved seeds may produce different pictures", and FLUX 3 Image has no seed field (Black Forest Labs, Prompt Reference). Liu and Chilton advise generating "between 3 to 9 different seeds" to see what a prompt can return, but their experiments used VQGAN+CLIP at 256 by 256 pixels, so treat that as a historical result rather than a rule for current models (Liu and Chilton, arXiv 2109.06977). For a current model, generate a few seeds yourself and compare them.

For the full method, read the controlled AI image prompt variations guide. For what each part of a prompt is for, read the anatomy of an AI image prompt guide.

Limits and what it cannot tell you

This is a rule-based reader. It does not run an image model, it cannot see the picture, and it does not know how any particular model will respond. Treat every check as a question to look at, not a verdict.

• The vocabulary is finite. A phrase that is not in it becomes detail, so a gene can be missed. The Genes table shows the exact text and character range behind every value, so you can check each one.
• The subject is the first clause, cut at a prose word. A prompt that opens with a setting, such as "in a forest, …", has no subject, and the tool reports a gap. A style-led start can still yield a subject, as in "oil painting of a lighthouse".
• An artist name is read as a style reference only when the name is capitalised. Lower-case names are not detected.
• The conflict and repeat rules are small and fixed. They catch a photograph next to an oil painting, or a golden hour next to a night setting. They do not weigh how strong each choice is.
• Quality words such as "masterpiece" are not documented as switches in the sources. Black Forest Labs advises omitting such tags for FLUX 3 Image (Black Forest Labs, Building a Good Prompt).
• The token figure is a rule-based estimate, not a vendor tokenizer.

The sources are mixed in age. Liu and Chilton's experiments used VQGAN+CLIP. Google's Vertex AI image prompt guide carries the banner "Vertex AI documentation is no longer being updated", and its examples are Imagen 3 and Imagen 2. Check the date of any advice before you rely on it for a specific model.

Privacy: what happens to your prompt

The tool's code never sends your prompt anywhere and never stores it unless you download it. The site-wide promo banner is separate and never receives your text.

Three actions take something out of the tab, and only when you press them. Copy places the text on your clipboard. Download saves a TXT, CSV or JSON file to your device. The settings link copies a URL that holds the dialect, mode, locks, count, seed and sentence form. It never holds your prompt or the variants.

The box holds up to 20,000 characters. A longer paste is cut at that limit, and the page says so. The tool's code keeps nothing between visits, so download the variants you want to keep. If you prepare reference images for a model, the Image Compressor also works in the browser and uploads nothing.