Guide
Why the same prompt gives a different picture
Nothing changed and everything changed. This is not a bug and not a setting someone forgot — randomness is the mechanism the whole process runs on. What matters is which parts of it you can pin down, and which pin down less than they appear to.
The mechanism
Generation starts from noise, not from your words
An image model does not build a picture up from your description. It starts with a field of random static and removes that static step by step, using your prompt as a steer for which direction to remove it in.
The static is therefore the first draft. A different field of static is a different draft, and it resolves into a different picture even though the steering was identical. Your prompt never described one image; it described a region of acceptable images, and each run lands somewhere else inside that region.
The number that generates the static is the seed. Fix it and you fix the starting point, which is why the same seed with the same prompt reliably returns the same picture — provided nothing else moved.
Why a seed is not enough
Four other things move underneath it
A seed fixes where the path starts. These decide the path taken from there, and changing any one is enough to break reproducibility while the seed sits there looking reassuring.
The sampler and step count. Different samplers walk from noise to image by different routes and arrive somewhere slightly different. More steps is not simply more detail; it is a different destination.
The guidance scale. How hard the prompt pulls against the model’s own tendencies. Small changes here move composition, not only fidelity.
The model version. The one that silently ruins archives. Weights get updated, and a hosted model can change behind a name that did not change. Yesterday’s seed against today’s weights is a different generator.
Batch size and hardware, occasionally. Generating four at once can give a different first image than generating one, because of how the random numbers are drawn across the batch. Floating-point arithmetic also differs slightly between GPU types, and across a long chain of operations those differences compound into a visible one.
In practice
Record the whole recipe, or let variation work for you
If reproducibility matters, a seed alone is not a record. The recipe is the seed plus the exact prompt, the negative prompt, the sampler, the steps, the guidance, the resolution and the pinned model version. Anything left out is a variable you have agreed to let move.
That is why generations here store their parameters alongside the result rather than leaving you to write them down, and why model references are pinned to a specific version rather than to a moving name.
But seeds are the wrong tool for the job most people reach for them to do. A seed holds a whole image, not the person inside it. Change the prompt from standing to sitting and the same seed returns a different face in a different room, because you moved the steering and the noise resolved elsewhere. Holding an identity across genuinely different images is a different technique entirely, and reaching for a seed to do it is the most common wasted afternoon in this field.
The other half of the answer is to stop fighting it. Variation is free exploration: generate several, pick the composition that works, and only then fix the seed and refine around it. Pinning everything from the first attempt means climbing a hill from a random starting point you had no particular reason to prefer.
Asked often
Short answers
- Why does the same prompt give different images every time?
- Generation starts from random noise and removes it step by step, so the starting noise decides which of many valid images you land on. The prompt describes a space of acceptable results rather than one result, and each run samples a different point in that space.
- Does using the same seed guarantee the same image?
- Only if everything else is identical too. A seed fixes the starting noise, but the sampler, the step count, the guidance scale, the model version and sometimes the batch size all change the path taken from that noise. Change any one and the same seed produces a different picture.
Keep reading
The rest of the series
- How long can an AI video actually be?The per-clip ceiling is real; length comes from assembly, and assembly has its own failure mode.
- What AI video actually costsWhat drives the per-second number, and why paying for failures is the thing to check before you buy.
- Who owns what you generateUsage rights and copyright are not the same question. Why the second one is unsettled almost everywhere.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.