What is a seed in AI image and video generation, and when should you fix it?
A seed is the starting point of the noise a model renders from. When to pin it to a number, when to leave it random, and why it is not a consistency button.
Open the Advanced section under a model's settings in the studio and on many models you will find a field called seed, usually set to -1. Most people never touch it. A few treat it like a magic number that will make every future picture look the same. The truth sits between the two: a seed is a small, precise tool for running controlled experiments, and once you know what it holds still, it saves real money on re-rolls.
What the number actually is
Image and video models of the kind in the catalog do not paint from a blank canvas. They start from a field of random noise and refine it, step by step, towards something that matches your prompt. The seed is the number that decides what that starting noise looks like. Same seed, same starting noise; different seed, different noise, and so a different picture even with an identical prompt.
That is why pressing Generate twice gives you two different results: with the seed at -1, the model draws a fresh random one for every render. Put a fixed number in the field and, on most models that offer it, the same prompt with the same settings comes back the same or very nearly the same.
Think of it as a lab notebook, not a style
The useful way to use a seed is the way a scientist uses a control: hold everything still and change one thing. With the seed pinned, any difference between two renders is caused by the one edit you made, not by luck. Without it, you cannot tell whether a better picture came from your new wording or from a friendlier roll of the dice.
- Render with the seed left random until one result has a composition you like.
- Pin a number of your own (say 1234) and render again. Keep this as your baseline.
- Change one thing in the prompt: one colour, one light word, one object. Render with the same seed.
- Compare against the baseline. Keep the edit if it helped, undo it if not, and move to the next one.
Step two matters more than it looks. A random seed is drawn on the spot and is not shown back to you, so the picture from step one cannot be reproduced exactly. Choose your own number before the render you want to build on.
A lighthouse on a rocky headland at dusk, waves breaking below, warm light in the lamp room, wide shot from the beachWhen to fix the seed
- Tuning a prompt word by word. The case above: every edit is judged against the same starting point.
- Comparing settings. One render per resolution or quality option, same seed, and you see what the setting itself does.
- A family of variations. A product shot in three colourways, the same scene at three times of day. A fixed seed keeps the layout closer between them than a random one would.
- Video experiments. Seeds are even more common on video models than on image ones, and a video render costs more, so being able to test one motion word at a time pays back quickly.
When to leave it random
When you are still exploring. Early on you want variety: five different compositions to choose from, not five near-copies of one. A pinned seed also has a quiet failure mode: if the starting noise happens to produce an awkward pose or a cramped layout, no amount of prompt editing will rescue it, because every render starts from the same awkwardness. When a direction stalls, set the seed back to -1 and roll a few times before pinning a new one.
What a seed does not hold still
The seed fixes the starting noise, and the starting noise only means something together with everything else. Change the model and the same number produces an unrelated picture. Change the aspect ratio or the resolution and the canvas changes shape, so the result moves too, sometimes a little and sometimes completely. A long edit to the prompt can drift as far from the baseline as a new seed would.
Most of all, a seed is not a character lock. Using the same number for two different prompts does not give you the same face in two scenes; it gives you two pictures that happen to share a starting pattern. For a recurring character, product or style, a reference image is the tool that works, and the reference guide shows how to use one.
Where you will not find one
Not every model has a seed field. Some keep it internal, and the simplest ones expose almost nothing: Fast Image, for instance, takes a prompt and an aspect ratio and that is all. Even where the field exists, a few models are not perfectly repeatable, and the same seed can come back with small differences in fine detail. Treat a fixed seed as a strong steadying hand, not a guarantee of a pixel-identical copy. And a fixed seed does not make a render cheaper: each one costs what the Generate button shows, the same as any other. What it saves is the re-rolls spent guessing, which, as the re-roll post works out, is where most of the money goes.
Questions
What does seed -1 mean?
It means random. The model draws a new seed for every render, so the same prompt gives a different picture each time. Enter your own number to make results repeatable.
Does the same seed give the same image on a different model?
No. A seed only means something together with the model, its settings and the prompt. On another model the same number produces an unrelated picture.
Can I use a seed to keep a character consistent?
Not reliably. A seed keeps the starting noise the same, not the person. For a recurring character, use a reference image of them and describe what should stay unchanged.
