The AI Image Seed, Explained — And Why It Matters for Reproducibility

What a seed is

A seed is a number. It initializes the random noise the diffusion model starts from. Same seed + same prompt + same model + same sampler + same steps + same CFG = same image, byte for byte.

Change any of those inputs and the image changes, even if the seed stays put.

Why "same seed different image" happens

Common gotchas:

• Different sampler — DPM++ 2M and Euler a will not agree on the same seed.
• Different step count — 25 steps and 30 steps produce different images.
• Different resolution — 1024×1024 and 1152×896 have different noise shapes.
• Different scheduler — Karras vs standard changes intermediate steps.
• Framework version bump — a Diffusers upgrade can quietly change the output.

If you're trying to reproduce a specific image, lock all of those, not just the seed.

The two use cases

Random seed (default): every generation gives you something new. Good for exploration.

Fixed seed: you're iterating on a prompt and want the composition to stay put while the details change. Fix the seed, tweak one word at a time.

The variant trick

Generate at seed X. Nail down the composition. Now generate at seed X+1, X+2, X+3 to get near-variants that share most of the composition but with subtle differences. Useful for producing similar ad creatives or a coherent set of illustrations.

Imagoat and seeds

Imagoat auto-picks random seeds today. Advanced seed control is on the roadmap — you'll be able to lock a seed on the Generate page and reuse it across generations.