Why the same prompt gives you a different picture every time

You write a description, you get a picture you like, and you close the window. The next day you type the same words and get something else entirely. Nothing is broken. The randomness is deliberate, and once you know where it lives you can switch it off.

The part that changes

Every image starts from a patch of random noise, a bit like television static. The description you wrote is the instruction for what to pull out of that static. Change the starting noise and you get a different picture from the same instruction, in the same way that two illustrators given one brief will hand back two different drawings, both correct.

That starting point is a number, usually called the seed. Most tools pick a new one each time so that pressing the button again gives you something fresh rather than the picture you already have. That is the behaviour you want ninety percent of the time, and the reason the tenth time is so frustrating.

Getting yesterday's picture back

If the seed is what changed, then fixing the seed is what brings the picture back. Somewhere in the tool there will be a seed field, often with a dice or refresh icon beside it that means pick a new one. Set it to the number the picture you liked was made with and generate again. Same seed and same description gives you the same image.

Which means the number is worth keeping. Good tools record it against every image they make, so you can look back at a result from last week and find the seed that produced it. If yours does not, write it down before you close the window, because it is not recoverable afterwards.

Same seed, still not the same

If you have fixed the seed and the picture still differs, something else moved. The usual culprits, roughly in order of how often they catch people out:

• The description is not identical. A missing comma or an extra word counts as a change.
• The image size is different. The same seed at a different shape is effectively a different starting point.
• A style or preset is applied that was not before, or is applied at a different strength.
• The number of steps or the guidance setting has moved, including by loading a different preset.
• You are on a different computer. Results can differ slightly between machines even with everything else matched.

That last one surprises people, so it is worth stating plainly: a seed is a reliable way to reproduce a picture on the machine that made it, and a good but not perfect way to reproduce it somewhere else.

Using the randomness on purpose

Once you can control it, the randomness becomes the useful part rather than the annoying part. The practical method is to work in two passes.

First, leave the seed random and write loosely. Generate a batch and look for a composition that is close to what you had in mind. You are not looking for the finished thing here, only for a starting point that is pointing the right way.

Then lock the seed of whichever one you liked, and start editing the words. Because the starting noise is now fixed, each change you make to the description shows you the effect of that change and nothing else. Swap the time of day, change the colour of a coat, add a detail to the background. You are adjusting one variable at a time instead of rolling the dice again, which is the difference between steering and guessing.

Tools that offer several variations from one seed are doing a version of this for you, holding the starting point steady while nudging everything else, so you can compare like with like.

Keep the settings, not just the picture

The mistake almost everyone makes early on is saving the image and nothing else. Months later you have a folder of pictures you cannot recreate, and no way to make a matching one for the second half of a pair.

What you want to keep, alongside the file, is the description you used, the seed, the size, and whatever style or preset was applied. Some tools save all of that into the image file itself, so dragging it back in restores the settings. Some offer to save a look as a reusable preset, which is the same idea with a friendlier name. Either is better than a note in a text file, but a note in a text file is much better than nothing.

It matters most when you need consistency rather than a single good image: a set of illustrations that belong together, a series of headers in one style, a character who should look the same in the second picture as in the first. All of that rests on being able to hold something steady while you change something else.

The short version

Different results from identical wording are the tool working as designed, not a fault. The seed is the random part, it is usually a field you can type into, and fixing it makes results repeatable. Explore with it unlocked, refine with it locked, and save the settings next to the picture so that today's good result is still available to you next month.

If your tool does not show a seed at all, that is worth knowing before you rely on it for anything that needs to match. Being able to reproduce your own work is the difference between a toy and something you can plan around.