The same prompt run twice produces two different images. The variation is designed in rather than accidental, and it is controllable once its source is understood.
Where the randomness enters
Generation begins with a field of random noise and progressively removes it, guided by the prompt, until an image remains.
The starting noise determines the composition. Different noise leads to a different arrangement of the same requested content.
The prompt constrains what appears, and it constrains where and how far less tightly than most users expect. Composition is largely inherited from the noise rather than dictated by the words.
What a seed actually fixes
The seed is the number that generates the starting noise, so reusing it reproduces the same starting point.
With an identical seed, prompt, model and settings, the output is reproducible, which is what makes systematic experimentation possible.
Change any one of those and reproducibility ends, and provider-side model updates break it without any action from the user. A seed recorded six months ago is not a reliable way to recover an image.
Why a fixed seed is the right way to test
Comparing two prompts on different seeds compares two starting compositions as much as two prompts.
Holding the seed constant while changing one term in the prompt isolates that term's effect, which is the only way to learn what a word actually does.
Running the final version across several seeds then shows whether the improvement holds generally or worked once by accident.
How the other settings interact
Guidance strength controls how closely the result follows the prompt, and raising it produces literal, harder-edged and often over-saturated images.
Step count governs how many refinement passes run, with diminishing returns past a point that varies by model.
Dimensions affect content rather than just framing, since composition conventions differ between portrait and landscape in the training data and the model reproduces them.
Why variation is worth keeping
Practitioners generate batches deliberately, because reviewing twenty options is faster than refining one prompt towards an imagined ideal.
The randomness supplies compositions the requester would not have specified, and selecting from options is a task people perform far better than describing.
Locking a seed too early narrows that search to a single composition, which is why experienced users fix the seed only once they have found a starting point worth refining.