Prompting

Same words, different order: AI image prompts

Reorder a prompt without adding a word and the render can change. Run a simple swap test to see what your model reads first, with before and after examples.

4 min read
A row of small glass prisms on a dark table, a violet beam lighting the first brightest while the rest fade into haze

Most prompt advice is about which words to add. This post is about a cheaper move: keep every word you already have and change only where it sits. Order is free, it costs no extra vocabulary, and it is the fastest way to find out what a model treats as the main idea of your sentence.

The short version of the theory: image models tend to give the opening of a prompt the most weight and let the tail fade, and a modifier tends to stick to the noun nearest to it. Neither is a hard rule, so do not take our word for it. Run the experiment below and watch your own results.

The swap experiment

  1. Write a prompt with four parts: subject, setting, light, style.
  2. Generate it once and keep the result as your baseline.
  3. Move exactly one part to the front, change nothing else, and generate again.
  4. Repeat for each part. Four renders, and you know which part your model obeys when it has to choose.

Compare the pictures by asking one question: what is the biggest thing in the frame? Whatever you moved to the front usually takes over that role. If the picture barely changes, the model is already reading your prompt as a whole and order is not your problem.

Prompt
Baseline: A ceramic teapot on a windowsill, rainy afternoon, soft grey window light, muted film photograph
Prompt
Light first: Soft grey window light, a ceramic teapot on a windowsill, rainy afternoon, muted film photograph
Expect the glow and the rain to become the subject, and the teapot to shrink into the scene.
Prompt
Style first: Muted film photograph, a ceramic teapot on a windowsill, rainy afternoon, soft grey window light
Expect grain and colour grade to dominate. Handy when the look matters more than the object.

Before and after: three rewrites

1. The prompt that opens with a mood

Prompt
Before: Moody, atmospheric, dramatic, a fisherman mending a net on a wooden pier at dawn
Prompt
After: A fisherman mending a net on a wooden pier at dawn, low mist over the water, moody and dramatic
The mood words were not wrong, they were early. Moved to the end they colour the scene instead of replacing it.

2. The adjective that sticks to the wrong noun

Prompt
Before: A red scarf and a blue door and a wooden chair in a hallway
Prompt
After: A hallway with a blue door. In front of it, a wooden chair with a red scarf draped over the back
Each colour now sits right next to the one thing it belongs to, and the layout is spelled out as a sentence instead of a list.

3. The setting that came first and swallowed the subject

Prompt
Before: A vast desert at sunset with huge dunes and a big sky, a small red fox sitting on a rock
Prompt
After: A small red fox sitting on a rock, vast desert dunes behind it, big sunset sky
When the fox goes first, the desert becomes a backdrop. If you want the landscape to win, reverse it on purpose.

A default order, and when to break it

  1. Subject, with its one defining detail.
  2. Setting, short.
  3. Light, one source and a direction.
  4. Style and medium, last.

Break it deliberately. A product shot on black can open with the light, because the light is the point. A poster in a named print style can open with the style. A wide landscape can open with the place. The default is a starting point for the case where you are not sure what matters most, and that is what the swap experiment is for. The full structure, with a worked example, is in our guide to writing prompts.

What order will not fix

  • Piles of buzzwords. Moving them around does not make them mean anything; see why quality buzzwords do little.
  • Things you want left out. Reordering will not remove an object. That is the job of negative prompts, where the model supports them.
  • Contradictions. If the light says dawn and the setting says midnight, one of them has to lose. Pick one.
  • Exact layouts and counts. Order nudges emphasis. It does not place objects on a grid, and models still miscount.

Different models also read prompts differently: one follows a long sentence closely, another rewards a short list of nouns. Run your swap test on the model you will actually use, and compare the options on the models page. Ready to try it? Open the studio, paste the baseline prompt above and move one phrase at a time.

Questions

Does the first word of a prompt really matter most?

Usually it carries the most weight, but not in every model. Some read the whole prompt more evenly. The swap experiment tells you in four renders which kind you are working with.

Should the style come first or last?

Last for most scenes, so it modifies a subject that is already clear. Put it first only when the look matters more than the object, such as a poster in a named print style.

Is reordering better than rewriting a prompt?

It is cheaper to try first, because it changes one variable and leaves your wording alone. If the picture still is not right after a reorder, the problem is the words, not their order.

#Prompting#Workflow

Keep reading

All articles →