Key takeaways
- A prompt is an instruction to a probabilistic system, not a blueprint — variation is the default state.
- Model family and version usually explain bigger differences than any wording change.
- Seed, sampler, dimensions and guidance each leave a visible fingerprint.
- You can’t eliminate variation, but you can converge: fix everything, then change one thing.
It’s the model, not the prompt
Each model family was trained on different data with different objectives. “Cinematic” to one model means teal-and-amber grade; to another it means letterboxing and lens flares. Even version bumps — the same model, months later — shift vocabulary mappings, which is why a prompt that sang in spring can feel flat after an update.
This is why the prompt generator offers output formats per family: the same analysis, phrased for how each model actually listens.
Seeds and sampling
Generation starts from noise, and the seed chooses which noise. Same prompt, new seed, new composition — always. On top of that, the sampler and step count decide how the noise gets resolved: some samplers are literal, others take creative liberties at low step counts. Two runs with different samplers can disagree more than two different prompts.
Dimensions and guidance
Aspect ratio isn’t just a crop — it changes what the model prioritizes. A 16:9 render invests in environment; a 4:5 invests in subject. Guidance (CFG) controls obedience: low values drift toward the model’s defaults, high values force the prompt but can bake in artifacts. Both leave fingerprints a trained eye can spot immediately.
| Factor | What it changes | Fix when drifting |
|---|---|---|
| Model family | Everything — vocabulary, style priors | Re-phrase the prompt for the target model. |
| Model version | Word mappings, defaults | Re-test saved prompts after updates. |
| Seed | Composition, detail placement | Run 3–4 seeds before judging wording. |
| Dimensions | Subject-vs-environment priority | Match ratio to the intended crop early. |
| Sampler / steps | Literalness, texture quality | Keep sampler constant while iterating. |
| Guidance (CFG) | Obedience vs drift | Nudge, don’t max out. |
Converging on a look you like
Treat variation as a search problem. Lock every dial except one. Generate a small batch. Pick the closest. Then adjust a single prompt section — light, palette, framing — and repeat. This is exactly the iteration method taught in refining prompts one variable at a time, and it turns “the AI ignored me” into a readable signal about which phrase the model weights lightly.
Editorial portrait, soft window key from camera left, muted slate-and-amber grade, 85mm at f/2, natural skin texture. [Fixed] — Run four seeds; choose the closest; then adjust only the lighting phrase next round.
When to stop iterating
If three consecutive single-variable changes don’t move the result, the model can’t hear that phrase. Re-word it in different vocabulary instead of repeating it louder.
FAQ
Does a longer prompt reduce variation?
It narrows the space but never closes it. Specificity beats length — one precise lighting phrase outweighs ten adjectives.
Can I lock a seed for consistency?
Within one model version, yes. Across versions or models, seeds aren’t portable — the noise space differs.
Is variation a bug?
No — it’s the feature that makes exploration possible. The skill is steering it, not removing it.