Noise math

Most people treat noise as a black box: pick a seed, get whatever the sampler hands you. Noise math from More Math opens the box. It's the node you use when you want to manipulate the initial latent noise itself - generate it from a formula, blend two noise tensors, add procedural structure - instead of accepting the default Gaussian draw.

How it works

The node evaluates your expression over the noise tensor, and it exposes the noise-specific variables: I (the input latent used to generate the noise), plus the image-style X, Y, W, H, C, B, T. The default expression is the familiar lerp a*(1-w)+b*w (in autogrow terms I0*(1-F0)+I1*F0) - a blend between two noise inputs. Where it gets interesting is the random-generation functions in the shared library: noise(seed), randn, randu, randc, randb, all seeded and deterministic for a given seed, so your handcrafted noise is reproducible. There's also perlin, voronoi, plasma, and ridged for procedural noise that isn't Gaussian at all.

Why would you want that? Structured noise - like a perlin base mixed into the initial latent - is a real technique for guiding composition before the sampler even starts, and blending noise tensors is how you get reproducible variation between otherwise-identical generations.

Inputs and outputs

V is the autogrow list of noise inputs (V0, V1, ...), F the floats, and Noise is the expression field (that's the one you type into, default a*(1-w)+b*w). remember_stack and the optional stack input are the pack's cross-node state, only relevant if you're threading values.

Outputs are NOISE and STACK. The noise feeds a sampler's noise input - in ComfyUI terms, the NOISE type plugs into sampler nodes that accept an explicit initial noise.

Installing it

Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/mcDandy/more_math
cd more_math
pip install -r requirements.txt

Restart ComfyUI, or install "More math" from ComfyUI Manager. Dependency: antlr4-python3-runtime on top of torch. No models.

Gotchas

Noise tensors are unbounded - Gaussian draws live around mean 0, variance 1, so clamp(noise, 0, 1) destroys the distribution and the sampler will happily amplify the damage. If you blend noise, keep weights in a range that preserves roughly unit variance. And remember the seeded generators need an explicit seed: use the same seed with the same expression and you get the same noise, change either and everything shifts. It's a brand-new, solo-maintained pack, so structured-noise workflows are mostly unexplored territory - which is exactly the fun of this node, but bring your own experiments.