Image Color Noise (Legacy)

Sometimes you just need a canvas of noise: a background pass, a texture seed for compositing, a test pattern, or the base for a grain effect you're about to grade on top of. Image Color Noise (Legacy) generates it on the fly - no file, no model, just a seeded noise image at whatever size you ask for.

Mechanically it's Fourier-domain noise synthesis. You pick a noise_type and it shapes a noise spectrum accordingly: white is flat random, red/pink/green/blue tilt the power spectrum toward those color names (red noise is heavy in low frequencies - think slow drift - blue is the spiky high-frequency stuff), and grey is Gaussian luma noise. The frequency and attenuation knobs scale the spectrum and amplitude, and seed makes the whole thing reproducible. Same seed, same texture, forever.

The inputs that matter

• width / height - output size, 64 to 4096.
• noise_type - the character of the noise. grey for monochrome luma grain, white for full-color static, red/pink/green/blue for the filtered flavors, mix for a blend. This is the input you'll change most.
• frequency - scales the noise spectrum. Lower values give broader, smoother variation; higher gives finer detail. 0.5 default.
• attenuation - the amplitude/contrast of the noise. 0.5 default.
• seed - reproducibility. Keep it fixed to replay the exact same texture; randomize for variety.

The output

A single Image. It's a raw texture, so what you do with it is up to you: blend it into a generated image for grain, use it as a placeholder, feed it to a sampler as image-to-image noise, or study what filtered noise looks like before you go tuning a grain pipeline. Because it's seeded, you can regenerate the identical texture in a later session - useful when the exact grain is part of a locked workflow.

Installing it

Part of silveroxides/ComfyUI-UtilsCollection. ComfyUI Manager → search ComfyUI-UtilsCollection → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart after. Dependencies are opencv-python and typing-extensions; the noise math is numpy/torch, both already present. No models. "(Legacy)" means it's the compat alias for the pack's canonical UC_Image_Color_Noise.

Where people get burned

The noise type names are the trap, because "red noise" is about the frequency spectrum, not the color you'll see - red noise looks like slow-moving smoky variation, not red pixels. Don't expect red to produce a red image; it produces low-frequency noise. And the output is normalized to the full 0–255 range every run, so attenuation doesn't behave like a simple "how loud is this" knob - it shapes the distribution, not the final brightness. If your grain looks washed out, check whether you're blending at the right opacity rather than cranking attenuation.