A subseed is a second seed that the sampler blends with your main seed. Subseed strength controls the blend ratio, so a small strength produces a slight variation of the original while keeping its structure.

How is this different from just changing the seed?

Changing the main seed gives a completely unrelated image. Using a fixed seed plus a subseed lets you interpolate between two noise patterns, so you get a controlled family of related images instead of random ones.

What subseed strength gives subtle variations?

Strengths around 0.1 to 0.25 give subtle, recognisable variations. 0.3 to 0.5 produces noticeable changes while keeping the gist, and above 0.6 the image starts to diverge significantly from the original.

Does ComfyUI have the same controls?

ComfyUI does not expose subseed directly, but you achieve the same effect by adding a small random delta to the noise seed or using a variation/noise-mixing node. The tool gives you a seed offset range to step through.

Seed Variation Strategy Guide

Explore variations without losing the image you love

You generated something great, but it is not quite right — the pose is perfect but you want a different expression, or the colours need to shift a touch. Re-rolling the seed throws away everything good. The fix is seed plus subseed variation: anchor to your winning seed and blend in a controlled amount of a second noise pattern. This guide turns a “how much change do I want” decision into concrete settings for Stable Diffusion WebUI or ComfyUI.

How seed variation works

Stable Diffusion starts every image from a field of random noise generated from a seed. The same seed plus the same prompt and settings always gives the same image. A subseed is a second noise field; subseed strength is how much of it gets mixed into the main one:

noise = lerp(seed_noise, subseed_noise, subseed_strength)

0.0 — pure main seed, identical to the original.
0.1–0.25 — subtle variation; same composition, small detail changes.
0.3–0.5 — noticeable shift; pose and layout mostly intact.
0.6–1.0 — strong divergence; closer to a fresh seed.

In WebUI this lives under the “Extra” checkbox next to the seed field. In ComfyUI you replicate it by stepping the noise seed by a small integer offset or using a noise-mixing node.

Tips for a clean variation workflow

Lock everything else. Keep prompt, steps, CFG, sampler and resolution fixed so the only variable is the variation amount.
Sweep a grid. Generate a row at strengths 0.1 / 0.2 / 0.3 / 0.4 to see the family of options at a glance, then pick a direction.
Combine with img2img. Once you have a near-perfect variation, take it into img2img at low denoising to polish without re-rolling.
Save winning subseeds. A seed + subseed + strength triple is fully reproducible — note it down so you can return to that exact image.