What does the seed do in Stable Diffusion?

By Updated 3 min read

The seed is an integer that initialises the random noise every generation starts from. With the same seed, prompt, model, settings and software, Stable Diffusion produces the same image; change the seed and you get a different composition for the same prompt. In Automatic1111 and on this site, -1 picks a random seed.

Why the seed matters

Diffusion starts from noise and every image is a deterministic function of that noise plus your settings. The seed is the recipe for the noise. That makes it the most useful tool for learning: fix the seed, change one thing, compare. It is the method behind our artist style comparison: 1,731 artists, four fixed prompts and one fixed seed per prompt. Only the artist name changes, so the differences you see are the artist's effect and nothing else.

How to use the seed

  1. Generate with seed -1 until a composition you like appears.
  2. Read the seed in the image info (Automatic1111 writes it in the PNG metadata and under the image) and paste it into the seed field.
  3. Now refine the prompt or the negative prompt word by word: the composition usually stays, the details move.
  4. With a non-ancestral sampler (Euler, DDIM, DPM++ 2M), the same seed still gives a close image when you change the steps or the CFG, which lets you compare settings fairly. Ancestral samplers such as Euler a add new noise at every step and drift.

The prompt examples on this site list the seed with nearly every image so you can start from a known result.

What breaks reproducibility

  • A different model or LoRA: the noise is the same but the denoiser is not, so the image changes: slightly with a light LoRA, completely with another model.
  • A different image size: the noise tensor has another shape, so the picture is unrelated even with the same seed.
  • A different GPU or software version: tiny floating-point differences can shift details; between GPU brands Automatic1111 even draws different noise, unless its random number source is set to CPU.
  • Automatic1111 has extra settings that change the result, such as the variation seed, Clip skip and Eta noise seed delta; keep them identical for exact repeats.

Variations around a seed

Automatic1111 has a variation seed and a variation strength slider: at 0.1 to 0.3 you get siblings of the image, same pose and palette with small changes, useful for picking the best hands or face. An X/Y/Z plot with seeds on one axis and a prompt word on the other produces a comparison grid in one run.

Try it now, freeSee 1,731 artists rendered with the same seeds →

Questions people also ask

Is a bigger seed better?

No. The seed is a label for a noise pattern, not a quality setting; 7 and 4,294,967,295 are equally good. Only its repeatability matters.

Can I get the same image on another website with the same seed?

Only if the site uses the same model, size, sampler, steps, CFG and software. Between two Automatic1111 installations that is usually true; between different services almost never.

Why does the same seed give a different image on this site's tools?

The public tools apply their own negative prompt, detail LoRAs, steps and sampler to every request, and inpainting resizes your image to fit 1024 px, so the result differs from what your local setup produces. Inside the site, the same request with the same seed is reproducible.

All guides