Entering identical wording twice into an image generator produces two different pictures, often with different composition and subject. The variation comes from where the process starts rather than from any ambiguity in the words.
Generation starts from noise
These systems begin with a field of random values and repeatedly remove noise from it, guided at each step by the prompt.
The starting field determines the rough arrangement of everything that follows: where masses sit, how the frame is divided, which direction a figure faces.
Two different starting fields therefore lead to two different images even when every other input is identical.
The seed is the controllable part
That random field is produced from a number called a seed, and holding the seed fixed reproduces the same image.
Most interfaces generate a new seed for each run without showing it, which is why the process appears non-deterministic to users.
Tools that expose the seed convert the generator into something repeatable, and that single control is what makes iterative work possible.
The prompt underdetermines the picture
A description of a few dozen words specifies a tiny fraction of the information in an image.
Everything unstated — lens length, time of day, exact posture, what is behind the subject — is filled from the model's sense of what usually accompanies such a description.
The noise field selects among those possibilities, so vaguer prompts produce wider variation and heavily specified ones produce narrower variation at the same seed spread.
Fixing the seed does not guarantee reproduction elsewhere
The same seed reproduces an image only within the same model version, sampler, step count and resolution.
Changing any of those alters how the noise is interpreted, and a hosted service may update its model without changing its name.
Images generated on a service months apart are therefore frequently not reproducible even with every visible setting recorded.
Working with variation instead of against it
The productive workflow generates a batch at varying seeds, selects the composition that works, then holds that seed while adjusting the prompt.
This separates the two things being chosen: the arrangement, which the noise controls, and the content, which the words control.
Anyone treating regeneration as a way to get a better version of the same picture is fighting the mechanism, since a new seed does not refine an image — it produces an unrelated one that happens to match the same description.