Run the same prompt twice and you usually get two different pictures. That is not the model being inconsistent. Generation starts from random noise, and the seed is the number that decides which particular noise you start from. Same seed plus same prompt plus same settings means the same starting point, and therefore the same image. Change the seed and everything downstream changes with it.
It helps to think of the seed as the roll of the dice rather than a quality setting. There are no good seeds and bad seeds in the abstract. A seed that produced a wonderful portrait for one prompt has no memory of that and will do something completely unrelated for the next one.
Why locking one is useful
The value of a fixed seed is that it removes a variable. If you want to know whether adding the phrase soft backlight actually helped, you need the rest of the image to hold still while you test it. With the seed floating, you change one word and get a different composition, different pose, different everything — and you learn nothing about the word.
So the working pattern is: generate freely until something is close, note the seed of the one you liked, lock it, then edit the prompt in small steps. Now the differences you see are the differences you made.
The same trick works for series. If you need three images that feel like they belong together, holding the seed while you vary only the subject often keeps the framing and light family-resemblant. It is not a guarantee — the models are sensitive, and a big change in wording can pull the image somewhere else entirely — but it is better odds than starting fresh each time.
Where the reproducibility ends
A locked seed only reproduces an image inside the same conditions. Change the model version, the resolution, the sampler or almost any other generation parameter and the same number will lead somewhere else. Seeds are not portable across tools either, and a seed you found on one service means nothing on another.
This catches people out when a provider updates a model. The prompt and seed you saved months ago stop giving you the picture you remember, and nothing is broken — the map changed under the coordinates. If an image genuinely matters, save the file, not just the recipe for it.
When to leave it alone
Most of the time, let the seed be random. Early in a prompt, variety is the thing you want: several rolls tell you what the model thinks your words mean, and often one of them is better than the idea you had. Locking too early narrows you to one interpretation before you have seen the alternatives.
The rule of thumb is simple. Explore with the seed free, refine with it locked, and stop treating it as a dial that makes images better. It makes them repeatable, which is a different and quieter kind of useful.