Why change only one thing at a time?
Changing one thing at a time is the simplest way to know which change caused which difference. OpenAI's image prompting guide says: "Refine one thing at a time and inspect the result." If you change the lighting, the camera and the colour together, a better image tells you nothing about which change helped.
Name the parts first; see the anatomy of an AI image prompt. The table below is a worked example with invented values. Each row changes one part of a base prompt, and the other parts stay as they were.
| Run | Changed part | Change from base | What to check |
|---|---|---|---|
| Base | none | The starting prompt | Three samples as the reference |
| Run 2 | Lighting | "at dawn" to "at dusk" | Does the light shift while the boat and harbour stay put? |
| Run 3 | Colour | "muted blue and rust" to "warm amber" | Does the palette change without the subject changing? |
| Run 4 | Camera | "low camera angle" to "bird's-eye view" | Does the framing change without the palette changing too? |
Settings count as variables too. Black Forest Labs' technical guide says "Changing the resolution makes a new generation, so the composition can differ from the draft." A resolution change is therefore a second change, even if you only meant to change the words. Keep resolution fixed while you test wording, and change it at the end.
Which design should you use for a set of variants?
Three simple designs make a practical start: one change at a time, a pair grid and a seeded shuffle. Use one change at a time when you want to learn what each part does on its own. The tool calls this the seeded random walk. Use a pair grid when you suspect two parts interact. Use a seeded shuffle when you want a wide, varied set that you can draw again later.
| Design | How to build it | Runs in the example | Use it to |
|---|---|---|---|
| One change at a time | Start from the base. For each part, swap in each alternative while everything else stays fixed. | 9 (3 parts times 3 alternatives) | Learn what each part does alone |
| Pair grid | Pick two parts and cross every option of one with every option of the other. | 9 (3 times 3) | Check whether two parts interact, such as lighting and mood |
| Seeded shuffle | Draw random combinations of alternatives from a fixed list, using a seed so the same list can be drawn again. | You choose, for example 12 | Explore widely while keeping the list repeatable |
The Image Prompt DNA tool can build a one-change set from a labelled prompt, up to the count you set. The one-change design grows with the number of parts, not with every combination. Three parts with three options each take nine runs, while every combination would take 27. A pair grid can hide the effect of a single part if you look only at its corners, so keep the single-part runs too. A seeded shuffle can change several parts at once, so it answers a different question: which mixes look good, not what each part does.
How do you keep a record you can trust?
Keep one row per run, and write the row before you look at the image. The row names the changed part, its old and new values, every setting you held constant, the seed and sample count, and what you were looking for.
| Run | Change, from to | Held constant, seed and samples | Looking for, and verdict |
|---|---|---|---|
| R3 | Colour: muted blue and rust to warm amber | Same model, settings and aspect ratio 3:2. Seed 42, 3 samples. | Palette shifts while the boat and harbour stay the same. Verdict: keep, drop or retest. |
Fill in the "looking for" column before you generate. Without it, you will tend to read whatever appears as the result you hoped for.
What does a seed do, and what does it not guarantee?
A seed is a number you can fix so that a run can be repeated. Fixing it can make a comparison cleaner, but it is not a promise of the same picture. Liu and Chilton note that "people often use seeds to have reproducible results and behavior", but in their VQGAN+CLIP experiments "using different seeds ... resulted in generations that would differ in composition". In their VQGAN+CLIP runs, a different seed changed the composition with the same words.
The AUTOMATIC1111 wiki warns about a parser change: "this means that your saved seeds may produce different pictures". The same wiki shows a prompt matrix rendered "all with the same seed and each with a corresponding prompt", which is a fixed-seed grid built into that tool.
Other tools differ. Black Forest Labs' reference lists no seed field for FLUX 3 Image, so a seed cannot be your control on that model. It does not say whether FLUX.2 takes one. OpenAI's image pages do not mention seeds at all. Their advice is "Repeat requests to measure consistency." No readable source says a seed reproduces an image across model versions, settings or parser versions. Stability AI's and Midjourney's seed documentation could not be read for this guide.
How do you judge results without fooling yourself?
Judge each variant against the base prompt, with the same settings, using at least three samples per variant, and with the look you are testing written down in advance. Three is this guide's own floor, not a rule from a vendor. Liu and Chilton suggested that you "generate between 3 to 9 different seeds" to see what a prompt can return, but that advice came from VQGAN+CLIP at 256x256 pixels and 300 steps, so treat it as a starting point.
A worked reading shows the discipline. Suppose run 3 gives one striking image and two plain ones. The honest verdict is "sometimes better", not "better". Write that down, then run it again before you keep the change. Do not pick the best of ten outputs and call it the effect of the change, because the other nine are the evidence.
How does the tool generate variants with locked parts?
The tool works from a lock on each part of the prompt. Setting, camera, lighting, style, colour and mood start free to change, and ticking a lock keeps one fixed. Subject, action, quality words and detail have no alternatives, so they stay fixed. In one-change mode, each variant changes exactly one unlocked part, and locked parts stay identical. The pair grid changes both chosen parts in every cell, and the seeded random walk changes one to three unlocked parts per variant.
Use the Image Prompt DNA tool for the one-change set, and save each output under its variant ID. When you save a folder of outputs to compare, the Image Compressor can convert and batch-compress the files in your browser, so you can compare them side by side with the same settings, or compare the originals.
Common mistakes and what this cannot tell you
- Changing a setting and a word together, then crediting the word. Resolution alone can change the composition.
- Reading one sample as a result. Three is the floor, and more is better when the variants look alike.
- Treating a seed as a promise. None of the sources read says it reproduces an image across versions, settings or parser changes.
- Applying "one thing at a time" to seeds as a rule. OpenAI states it for iterating on images, and no source we read applies it to seeds. It is a method for your own tests.
- Copying Liu and Chilton's numbers as current advice. They come from VQGAN+CLIP at 256x256 pixels.
Only your own tests, with your settings, can show which parts matter for your model. It gives no figures for Midjourney or Stability AI, because their documents could not be read during research.
Common questions
How many samples should I generate for each variant?
Generate at least three samples of each variant with the same settings, and compare them with the base prompt's samples. That floor is this guide's own rule. Liu and Chilton suggested between 3 and 9 seeds, but their experiments used VQGAN+CLIP, an older system, so the range is a starting point. Write down the number you used so that someone else can repeat the test.
Can I get the same AI image twice from the same prompt?
Not guaranteed. A seed can help you repeat a run where the tool offers one, but no source we read says a seed reproduces an image across model versions, settings or parser versions. The AUTOMATIC1111 wiki warns that saved seeds may produce different pictures after a parser change. Where a tool has no seed field, as with FLUX 3 Image, repeat the request and compare the results. OpenAI's advice is the same: "Repeat requests to measure consistency."
What is a seed in an image generator?
A seed is a number you can fix so that a run can be repeated. Some tools expose one and some do not. Black Forest Labs' quick reference lists no seed field for FLUX 3 Image, and OpenAI's image guides do not mention seeds at all. Treat the seed as one setting to hold still during a test, not as a way to lock the final picture.
Why does changing the resolution change the picture?
Black Forest Labs says that changing the resolution makes a new generation, so the composition can differ from the draft. Treat resolution as a variable. Hold it fixed while you test words, and change it only at the end. The same guide suggests iterating at a low resolution, then using 2k or 4k when you need more detail.
Should I change the prompt or the seed first?
Change the prompt first, one part at a time, with the seed held fixed where the tool allows it. A different seed gives a different sample of the same prompt, so use seed changes to measure how much the output varies, not to test a word. Liu and Chilton's experiments each name one factor or interaction, but the paper states no one-variable rule of its own.
Does one-thing-at-a-time testing work for every model?
It is a method, not a guarantee. OpenAI's advice to refine one thing at a time concerns iterating on images. Other vendors describe iteration differently: Google's guide, marked as no longer updated, builds from a core idea, and Black Forest Labs suggests adding detail as you review results. The one-part test is most useful for learning what a part does.
Sources and further reading
- Image prompting guide, OpenAI API documentation, no date shown, read 17 October 2026
- Image generation guide, OpenAI API documentation, no date shown, read 17 October 2026
- Design Guidelines for Prompt Engineering Text-to-Image Generative Models, Vivian Liu and Lydia B. Chilton, CHI 2022 and arXiv 2109.06977 version 3, revised 28 September 2023
- Features, AUTOMATIC1111 stable-diffusion-webui wiki, community documentation, last edit 15 December 2024
- Technical Parameters, Black Forest Labs, FLUX 3 Image, no date shown
- Prompt Reference, Black Forest Labs, FLUX 3 Image and FLUX.2, no date shown
- Prompt and image attribute guide, Google Cloud for Imagen on Vertex AI, last updated 13 October 2026, marked as no longer updated