Random Noise Variation Simple 🍌

If you've used Automatic1111, you know the "variation seed + variation strength" pair - a way to get a sibling of an image you liked, keeping the composition roughly intact while nudging details around, instead of rerolling to something totally unrelated. ComfyUI's advanced sampling pipeline doesn't have a built-in equivalent; Random Noise Variation Simple is cgem156's version of it, built as a NOISE generator for that pipeline.

What the two controls do

seed sets your starting point, same as any seed. similarity (0–1, step 0.001) is the blend dial: at 0 you're effectively getting fresh random noise regardless of the seed you picked, and turning it up blends toward a fixed, seed-tied pattern. The exact interpolation math isn't documented anywhere - this reading comes from how the params are named and how A1111's version of the same idea behaves, not from a spec - so treat similarity as a dial to explore rather than a value with a precise, guaranteed meaning. In practice: start from a seed whose composition you like, dial similarity up in small steps (0.1–0.3 first), and watch how much changes.

This is the "Simple" node in the pair - its sibling, VariationNoise, adds a separate base_seed and a batch_index for more control. Reach for this one when you just want a single seed plus one blend knob and don't need the extra structure.

Using it

Output is a NOISE object - wire it into SamplerCustomAdvanced or this pack's SamplerCustomAdvancedPreview/SamplerCustomXY. It won't plug into a plain KSampler's seed widget; that's a different type.

Installing it

Via ComfyUI Manager: search cgem156-ComfyUI, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/laksjdjf/cgem156-ComfyUI

Restart afterward. No model downloads, no extra dependencies.

Common issues & troubleshooting

At similarity=0 it just looks like a normal random seed every time. That's expected - 0 is the "no similarity" end of the dial, functionally close to picking a fresh seed on every queue regardless of what seed is set to.

Even near similarity=1, results barely resemble each other. Check what sampler is downstream. Ancestral samplers (Euler A, DPM++ SDE) inject their own fresh randomness at every step on top of whatever noise you feed in, which can wash out a subtle variation effect entirely. If you want a clean before/after comparison, pair this with a converging sampler (Euler, DPM++ 2M) - those settle toward a stable image instead of adding their own noise throughout, so the effect of similarity actually shows.

You need per-batch-item control, or want to keep an explicit "base" seed separate from the variation seed. That's exactly the gap VariationNoise fills - same pack, same idea, one more knob.

Type errors wiring it up. NOISE only fits ComfyUI's advanced sampler nodes (SamplerCustomAdvanced and this pack's variants) - it won't connect to a classic KSampler's seed field.