Models & Parameters

Seed Number: How to Make a Result Repeatable

Also called image seed, random seed, seed value, fixed seed

A seed number is the integer that initialises the random noise a diffusion model starts from. Hold the seed and every other setting fixed and you get the same output again, which is what turns generation from a slot machine into a controlled experiment where one variable changes at a time.

What the number is attached to

Diffusion starts from a field of pure noise and removes it step by step until an image appears. That noise has to come from somewhere, and a pseudorandom generator produces it from a single integer. That integer is the seed number. Same integer, same starting noise, same path through the denoising process, same picture.

This is why it is the most useful setting on the page and the least interesting one to tune. There is nothing to optimise. A seed is a bookmark, not a quality dial.

The workflow that makes it worth using

Almost every wasted afternoon of generation comes from changing two things at once. A fixed seed removes that problem entirely.

  1. Explore on random. Run the prompt four or five times with a random seed and look at the spread. This tells you what the prompt actually means to the model, which is usually broader than you assumed.
  2. Lock the best one. Copy the seed number from the run that came closest.
  3. Change exactly one thing. Guidance, a single clause, one style word, the sampler. Regenerate. Now the difference you see is caused by the thing you changed.
  4. Unlock at the end. Once the prompt is right, go back to random and generate a batch. You are now sampling variations of a prompt you trust rather than guessing.

Step three is where the whole value sits. Without a fixed seed, tuning a prompt is indistinguishable from re-rolling it.

What a seed does not do

It is not a style, an aesthetic, or a shared resource. The number means nothing outside the exact configuration that produced the image: same model, same version, same resolution, same everything. Somebody else's image seed pasted into your interface is just an arbitrary integer.

It is also not a fix for reproducibility across time. Hosted endpoints get updated, and a model version bump behind the same name will break your saved results without warning. If a look genuinely matters to a project, save the output file rather than trusting that the seed will regenerate it in six months.

Using it to get variation instead of repetition

The inverse trick is worth knowing. Once you have a composition you like, keeping the seed and making tiny prompt edits gives you near siblings: the same layout and lighting with a different jacket colour or time of day. Change the random seed instead and you get a genuinely different picture. So one number controls both halves of the job, holding a frame still while you iterate, or releasing it when you have iterated enough.

The prompt for this

A starting point that reliably produces the effect. Adjust the subject and setting; keep the technical clauses.

A lone red telephone box on a foggy moor at dawn, low contrast, 35mm, muted palette

Try Seed yourself

Open the generator with a starting point already filled in.

Frequently asked questions

What is a good seed number?
There is no such thing. Every value is an equally arbitrary starting point, and a seed that produced a great portrait carries nothing over to your next prompt. Lists of lucky seeds are folklore. What matters is not which number you use but that you write it down.
Why did the same seed give me a different image?
Something else moved. The usual suspects are a changed sampler or step count, a different resolution, a prompt edited by one character, or the provider updating the model version behind the endpoint. Bitwise reproducibility also depends on hardware and numeric precision, so the same seed on different GPUs can differ slightly.
Does an image seed transfer between models?
No. The number indexes noise, not content, and each model interprets that noise through its own weights. The same seed on two checkpoints gives two unrelated pictures. Seeds also do not survive a resolution or aspect ratio change, because the noise tensor itself changes shape.
Should I keep the seed random or fixed?
Random while you are still exploring, because variety is the point and a fixed seed hides how much range the prompt has. Fixed the moment you like something and start tuning, since otherwise you cannot tell whether an improvement came from your edit or from luck.
Does the seed affect motion in video?
Yes, and more than people expect. In video the noise initialises the whole clip, so it shapes the movement pattern as well as the first frame. Two seeds on an identical prompt can give the same subject two completely different actions.
How do I find the seed of a result I already generated?
Most tools return it with the output, in the result metadata or a details panel, and many write it into the file. If a run gave you something you want to build on, copy that number before generating again, since a random run will not hand it back twice.

Related terms