Abstract Image Noise Layer

If you're picturing TV static when you read "noise layer," stop. Abstract Image Noise Layer generates soft, cloud-like organic gradients - the sort of thing you'd get from Perlin noise - by summing a handful of sine waves. It's the fog, the marble, the smoky undertone of an abstract composition, and it's arguably the most useful single layer in this pack for generating img2img bases.

Why? Because it fills the canvas with smooth tonal variation instead of edges. Shapes and lines give the model geometry to interpret; noise gives it atmosphere and lighting cues. Feed a soft two-color sine-field through VAE Encode into a sampler at 0.8–0.9 denoise and the model turns that wash into clouds, water, fabric, or just a gorgeous gradient-rich backdrop. It's also the layer that keeps your abstract image from looking flat - throw a noise layer under your shapes and suddenly everything has depth.

The inputs you'll actually set

• num_layers - stack 2–3 noise layers and the cloud fields start interacting like actual weather. One layer reads flat; several read dimensional.
• alpha - opacity of the noise, 0–255. The node already keeps internal alpha modest, so you mostly adjust this to pull the noise forward or push it back.
• blend_mode - the standard 12 (normal, multiply, screen, overlay, add, …). overlay and multiply are magic here: they let the noise tint and shade the layers beneath without erasing them.
• noise_color_mode - random RGB, hex via noise_color_hex, or the named colors. Two-color noise looks best; random picks the two from the same mode.

shape_position can confine the noise field to a quadrant or center if you want a localized haze, and randomize delegates scale, colors, position, and blend to the seed.

How it works

The source calls it "fast Perlin noise," and it earns the name honestly. Instead of expensive fractal interpolation, it builds a meshgrid, layers 3–8 sine waves with randomized frequency, phase, and amplitude, normalizes the sum to 0–1, and maps that through two colors to make an RGBA layer. The output is smooth and organic - no blocky pixels, no salt-and-pepper static. That's a deliberate design choice that separates this from a plain "add random noise" filter: the goal is a usable painting layer, not a data corruption effect.

One honest caveat: because it's a simple sine sum rather than true multi-octave Perlin, don't expect perfect cloud gradients in every seed. Some seeds give lovely swirling marble; others give broad stripes. That's fine for a base image - reseed until the field pleases you, then lock it.

Install

Part of the AbstractImaGen pack - ComfyUI Manager, search AbstractImaGen, install, restart, refresh. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/wakattac/ComfyUI-AbstractImaGen

Dependencies: Pillow, numpy, scipy, torch - all present in a normal ComfyUI install. No models to fetch. After restart, the node appears under AbstractImage/Layers.

Gotchas

• noise_color_hex requires noise_color_mode set to hex, same contract as the other layers.
• This node expects a single image frame. Batches or videos won't process per-frame.
• Noise layers are cheap, so don't be shy - but the node's own internal alpha is already semi-transparent, so if your noise looks invisible over a busy base, check alpha and the blend mode before assuming it's broken.