🌫️ 智绘_生成噪波

Most nodes in a ComfyUI graph consume an image and make another image. ZH_GenerateNoise (🌫️ 智绘_生成噪波) produces an image from nothing - just dimensions and a seed. That sounds pointless until you want a pure noise texture for something: a base layer for image-to-image work, a test pattern, a cloud-smoke tile for a compositor, or a full-frame grain source you'll blend over footage later. It's the from-scratch sibling of ZH_AddNoise, which only adds noise to an image you already have.

What it actually generates

The dropdown gives you five procedural noise families, all computed with numpy so there's zero model download and zero VRAM:

• 白噪声 (White noise) - pure random per-pixel static. The flat, honest kind.
• 柏林噪声 (Perlin) - smooth, gradient-based noise. This is your clouds, smoke, and terrain texture.
• 分形噪声 (Fractal) - Perlin layered over itself at multiple frequencies (the octaves parameter controls how many layers). This is where you get the interesting, mountain-range-looking stuff.
• 梯度噪声 (Gradient) - smooth value-noise gradients, great for soft transitions and masks.
• 云朵噪声 (Cloud) - the soft, puffy one. Perfect as a starting cloud/atmosphere map.

Two parameters shape the character. scale (0.1–10) controls feature size - crank it up for big soft blobs, keep it near 1 for fine detail. octaves (1–8) stacks detail layers for the fractal types; 4 is a solid default, higher gives busier textures at a small speed cost.

The inputs that matter

For the simple case you only touch three: width and height (both default 512, and they take any multiple of 8 up to 8192 - so yes, it'll make an 8K noise image if you're patient), and seed to lock or randomize the pattern (0 = random each run, same convention as ZH_AddNoise). batch_size (1–64) emits multiple independently-generated frames in one shot, which is handy for building a noise video sequence - though note each frame in the batch uses the same seed stream, so they'll correlate; if you need uncorrelated frames, run it once per seed. Output is a single 噪波图像 (IMAGE) socket.

Install

It's part of the 智绘灵箱 (ComfyUI-ZhiHui) pack, so install the pack, not the node:

cd ComfyUI/custom_nodes
git clone https://github.com/zhuyungen/ComfyUI-ZhiHui.git

Restart ComfyUI (or use ComfyUI Manager, searching "智绘灵箱" / "ComfyUI-ZhiHui"). You get this node along with ~40 others in the pack; all the noise code here is numpy-only, so no extra install weight beyond what the pack's requirements.txt drags in for its other, heavier nodes.

Where people get burned

• "Scale" is not "size". New users read scale as image dimensions and wonder why nothing got bigger. It's noise feature size, not resolution - the resolution lives in width/height.
• Random seed makes it un-reproducible. Same convention as everywhere: 0 = randomize. Set a real seed if the texture needs to match a previous run.
• It outputs a noisy IMAGE, not a latent. You can feed it to an img2img node or blend it downstream, but it won't plug into a KSampler's latent port. If you actually wanted latent-space noise, that's the KSampler's own noise_seed you're after, not this node.

Reach for this when a workflow needs any image-shaped thing to test against - a mask matcher, a blend base, a texture layer - and you don't want to hunt down or generate a real photo. It's a plumbing-y utility, but it's one of the rare ones that can actually surprise you: a fractal noise layer at the right scale looks a lot more like terrain than "noise" sounds.