Power-Law Noise 🦚

Every Stable Diffusion image starts from noise - but not all noise is equal. The default in ComfyUI (and almost everywhere) is white noise: every pixel independent, all frequencies equally loud. Power-Law Noise from Power Noise Suite generates the other colors instead - pink, brown, blue, green, grey - where the energy is spread across frequencies according to a power law. It's the difference between static on a dead TV and the rumble of ocean waves, and when you feed the latter into a sampler, the texture of the output changes with it.

This is the "revamped" core of the pack. WASasquatch rebuilt the noise system here for "more accurate noise representation," and this node is the standalone version of what the pack's Power KSampler Advanced uses internally every step. Give it a size, pick a noise type, and it hands you back a LATENT full of that noise plus an IMAGE preview so you can see what you're about to sample from.

The inputs that matter

• noise_type - the whole point. The list: white (regular ComfyUI noise), grey, pink, green, blue, plus random_mix, brownian_fractal, velvet, and violet. White = random pixels; pink and brown lean into low frequencies (smooth, cloudy, organic); blue and violet do the opposite and are sharper, finer, more high-frequency. random_mix rolls three random types and composites them. If you don't know what these look like, you will after five minutes of generating previews - this is a "screenshot the output and write down what you liked" situation, because nobody's written a good guide for it.
• alpha_exponent - steepness of the power law. Higher means softer, more low-frequency noise; negative values push the other way. This is the single most useful knob for tuning "how much do I want this to look like clouds vs. static."
• scale - overall amplitude. Leave at 1 to start.
• resampling - how the noise is scaled down to latent size (the node generates at full pixel size, then downsamples 8×). nearest-exact keeps values chunky; bilinear/bicubic smooth them.
• device - cpu or cuda. CPU is fine for noise and keeps your VRAM for the sampler.

optional_vae deserves a warning. If you leave it empty, the node just downsamples the noise straight into a latent - fast, and what you want for feeding a sampler. If you plug a VAE in, it runs a real VAE encode on every batch item, which is slower and changes the noise's character. Only do that if you specifically want the noise to live in encoded latent space.

Outputs are latents (wire into a sampler or Blend Latents) and previews (look at this).

What you'd use it for

The obvious one is img2img and latent mixing: generate pink noise as your starting latent, then run a KSampler at low denoise for dreamy, structured variation. The other is education - once you've seen what each noise color does to a generation, the Power KSampler Advanced's noise_type menu stops being a mystery.

Install

It's part of Power Noise Suite, so install once and all the nodes come along. In ComfyUI Manager, search "Power Noise Suite", or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/PowerNoiseSuite

Restart. No downloads, no models - the only dependency is torch, which every ComfyUI already has.

Gotchas

The README's own advice applies here: don't crank things blindly. A noise type that looks great in the preview can behave differently once the sampler is walking across it, and the pack's own troubleshooting notes blame green-tinted or nonsense output on tolerances being too high or input noise being too strong. Keep alpha_exponent and scale modest, preview as you go, and treat this node as the test bench for the sampler's noise settings rather than a set-and-forget.