Why do I get a different image from the same prompt?

Because the seed was not fixed. Generation starts from random noise and the starting point changes on every run, which affects the result more than most prompt edits do.

Does the same seed work across different models?

No. A seed indexes into that specific model’s noise space, so the same number in another model or version produces something unrelated. Always record the model with the seed.

There is no such thing. No seed is inherently better; it is simply a starting point. What matters is fixing one so your comparisons are meaningful.

How do I find the seed of an image I already made?

If the tool exposed it, it will be in the generation metadata alongside the prompt. If it was never recorded, it cannot be recovered from the image file.

Should I always fix the seed?

Fix it when refining or comparing. Leave it random when exploring, because variety is the point at that stage.

Does the seed affect quality?

Not systematically, but some seeds suit a given prompt better than others. That is why generating a batch and choosing is standard practice.

AI image glossary

What is a seed in AI image generation?

The number that determines the starting noise a generation begins from. The same seed with the same prompt and model reproduces the same image.

A seed is the number that determines the random noise an image generation starts from. Because diffusion begins with noise and denoises toward your prompt, the starting noise shapes the result enormously — so the same seed, prompt and model produce the same image every time, and a different seed produces a different one.

A single small numbered brass token alone at the centre of a wide warm paper surface

Why the same prompt gives a different image every time

Nothing about your instruction changed between runs; the starting point did. Diffusion begins from random noise and denoises toward your prompt, so the noise it began with shapes the result enormously. Left unfixed, that number is chosen at random on every generation.

Fixing the seed turns guessing into testing

Change one word with the seed unfixed and the difference you observe is mostly the new starting noise rather than your edit, so you learn nothing about the word. Fix the seed and the only variable is the thing you changed, which is the entire basis of learning what prompt terms actually do.

A seed only means something inside one model

The same number in a different model, or often a different version of the same model, produces something completely unrelated. A seed indexes into one specific model’s noise space rather than describing an image, so recording the model alongside it is not optional.

What to save when a result is worth keeping

Prompt, seed and model together. That triple is the closest thing generation has to a photographic negative, and without all three a good image cannot be recovered, varied deliberately, or handed to someone else to reproduce.

Seed at a glance

Type: An integer
Controls: The starting noise
Same seed + prompt + model: Same image
Change the model: Reproducibility is lost
Typical use: Locking a result you liked
Random each run: Unless you fix it

Related terms

What a seed sits next to

What is a checkpoint?The other half of reproducibility.

What is a sampler?How the steps are taken.

What is guidance scale?How hard the prompt is enforced.

120 images and their prompts120 images with their real seeds.

AI prompt generatorBuild a prompt worth locking.

AI Image GlossaryEvery term, defined.

Put it to use

Where this control appears

the main generatorThe main generator — text to finished image.Open browse real prompts120 images with the exact prompts that made them.Open Image to ImageStart from a picture you already have.Open edit part of a photoChange part of a photo, keep the rest.Open upscale to 4KEnlarge to 4K without the smearing.Open restore an old photoExposure, noise and old-photo restoration.Open masked object removalMasked removal that rebuilds the background.Open blur a photo backgroundDepth-aware, not a uniform filter.Open Colorize PhotoPlausible colour for black and white.Open make a markMarks that still read at 16 pixels.Open design a tattooDesigns that survive being tattooed.Open AI Art GeneratorIllustration and painting, not photographs.Open studio headshotsStudio-grade portraits without a studio.Open Passport PhotoThe exact specification, and how to meet it.Open interior visualisationRestyle a room you can photograph.Open terms explainedSeed, denoise, inpainting, explained.Open

Seed — questions people ask

Why do I get a different image from the same prompt?: Because the seed was not fixed. Generation starts from random noise and the starting point changes on every run, which affects the result more than most prompt edits do.
Does the same seed work across different models?: No. A seed indexes into that specific model’s noise space, so the same number in another model or version produces something unrelated. Always record the model with the seed.
What is a good seed?: There is no such thing. No seed is inherently better; it is simply a starting point. What matters is fixing one so your comparisons are meaningful.
How do I find the seed of an image I already made?: If the tool exposed it, it will be in the generation metadata alongside the prompt. If it was never recorded, it cannot be recovered from the image file.
Should I always fix the seed?: Fix it when refining or comparing. Leave it random when exploring, because variety is the point at that stage.
Does the seed affect quality?: Not systematically, but some seeds suit a given prompt better than others. That is why generating a batch and choosing is standard practice.

Related AI image terms

Start creating today

Generate your first image with the best AI image generator.

Turn a sentence into a finished, photorealistic image in seconds — then refine every detail. No setup, no Discord, no GPU.

Generate with a fixed seed Get Started

Join 4,200+ creators using LaFoto