Key points
- A seed is the number that decides the very first blur a picture starts from. Start from the same number and you get the same blur.
- Keep the number and every other setting the same, and the same picture comes back no matter how many times you run it. It's how a picture you liked gets found again.
- Fix the seed when you want to turn just one dial and see what changes. It tells you the difference came from that dial alone.
- Whether a number is big or small means nothing. A number that worked well before has no guarantee of keeping that edge once the model or settings change.
- Even with the same seed, changing the sentence, the step count, or the size changes the picture. A small tweak to the sentence or the steps often keeps the overall layout and shifts the details, but changing the size changes the whole thing.
The analogy
Stand in front of an unmanned parcel locker and you have to enter a number before the door opens. Press 3 and compartment 3 always opens, and what's inside is always the same. Press it again later and nothing changes.
Change the number to 7 and a completely different compartment opens. That doesn't make 7 a better number than 3 — it's just a different compartment, and being next to 3 doesn't mean compartment 4 holds anything similar. A seed is that compartment number. Start from the same number and the same picture comes out; change the number and a different picture comes out.
Leave the number blank and let the machine open whatever compartment it likes, and something different comes out every time. That's what it means to leave the seed field empty.
In detail
A picture starts from a single sheet of blur
An image generator doesn't start from a blank screen — it starts from a screen packed with tiny blur. Drawing that blur takes a long run of random-looking numbers. Instead of truly random numbers, an image generator takes one starting number and repeats a fixed calculation to spin out a long string of numbers that only looks random, and draws on that. The seed is exactly that starting number.
Same starting number, same string of numbers that gets drawn. Same first blur, so the steps that peel it away line up the same way too. That's why writing down the seed lets a picture that turned out well get made again months later.
Some methods also mix in a bit more randomness at every step along the way. Those numbers trace back to the same starting number too, so holding onto the seed reproduces even that jitter exactly.
Holding the number steady makes comparison possible
The moment to fix the seed is when there's one dial you want to test in isolation. Leave the number untouched and raise only the step count, and whatever changes is entirely down to the steps. Let the seed drift too and the whole picture shifts, and there's no way to tell what actually caused the improvement.
Going the other way, checking whether a sentence is well written calls for running several different numbers. A sentence that only works with one particular seed just got lucky. A sentence is solid only once it produces reasonably good pictures across three or four different seeds.
The number itself has no power
People sometimes trade tips about seeds that "work well." But the number itself carries no power. Change the sentence or the rest of the settings and the same number produces a different picture. A story about a seed that worked well comes with no guarantee once you step outside that particular sentence and those particular settings.
There's no difference between a small number and a large one, and neighboring numbers don't produce similar pictures either. 1 and 2 aren't next-door neighbors — they're complete strangers. Bump the number up by just one and the entire starting blur gets redrawn from scratch.
Same number, different picture
If the seed matches but the result doesn't, something else must be different. A single space in the sentence, the step count, the guidance strength, the image size — any one of these changes the picture on its own. Changing the size in particular changes even the size of the starting blur, so a completely different picture comes out.
Even with every setting matched, differences can still creep in. Different tools or chips can draw their string of numbers in different ways, so the same number can give an entirely different blur; and even with the same method, a different software version can produce a difference in a very small digit, and that gap can grow large enough to see as it carries through step after step. A seed is a solid way to find a picture again, but finding it back means writing down every other condition too, not just the number.
More precisely
A seed is the integer that sets the starting state of a random-number generator. The numbers used here aren't truly random — they're a string of values calculated by a fixed rule from that starting state, which is why they're called pseudorandom. The same starting state always produces the same string. An image generator fills the first screen with this string of numbers, and the sampler also draws the jitter it mixes in at every step from that same string. Depending on the tool, leaving the seed field blank pulls a number automatically from the clock or the device's state, and many tools display the number used alongside the finished picture.
The analogy breaks down in places. A parcel locker at least gives some clue about what's inside before the door opens, but a seed gives no way at all to know what picture will come out just by looking at the number — the only way to find out is to actually generate it. And a parcel locker gives the same result every time in front of the same machine, but with a seed, switching to different computing hardware can shift the result slightly, sometimes completely, even with the same number.
Try it yourself
Common misconceptions
It's easy to think a number that worked well works everywhere, but actually one model may favor certain numbers a little, yet nothing guarantees that advantage carries over once the model or the settings change.
It's easy to think matching the seed alone brings back the same picture, but actually the model, the sentence, the step count, the guidance, the size, and the rest of the settings all have to match too.
It's easy to think a small change in the number makes only a small change in the picture, but actually neighboring numbers have no relationship at all, so the whole picture can flip.
One-line summary
A seed is the starting number that sets a picture's first blur, and it has to be written down together with every other setting to meet the same picture again.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-17