CPPN Generator

This is the node that doesn't belong in a diffusion UI - and it's great because of it. CPPN Generator produces abstract, organic, sometimes genuinely beautiful procedural images with zero weights, zero model downloads, and zero GPU requirement beyond what a sampler would hog. A CPPN - Compositional Pattern-Producing Network - is a decades-old idea from Kenneth Stanley's NEAT research, made famous to the generative-art crowd by hardmaru's blog posts. This node is a direct descendant of those, and it's exactly the thing to reach for when you need a weird abstract texture, a tiling-ish background, or an animation seed that has nothing to do with a checkpoint.

How it works

A small neural network takes coordinates as input - x, y, and radius from center r for every pixel - plus a latent vector z, and outputs pixel color. Because the network is a continuous function over coordinates, it produces smooth, structured patterns rather than noise: blobs, ripples, gradients, cells. The magic dials are the activation functions. Set everything to Sin and you get the classic CPPN look - layered interference patterns. Tanh gives soft organic gradients, ReLU gives hard-edged geometric shards, Modulo gives... well, try it.

The node runs on CPU by default (it builds its model on cuda:0 if available, else CPU), and for a single frame it's quick. The latent z is seeded, so a given seed always reproduces the same pattern. One honest caveat: framecount is declared as an input and both the README and the auto-generated docs talk about animation, but the shipped code never actually reads it - run it and you get a single image, full stop. To animate, drive z_offset or the coordinate offsets from a loop/queue node instead of hoping framecount multiplies your output.

The inputs that matter

There are roughly twenty inputs and you should touch about six of them:

• seed - the pattern identity. Everything else equal, a new seed is a new pattern.
• width / height - output size, up to 8192.
• framecount - declared, but the shipped code ignores it (see below). Don't expect it to multiply your output yet.
• mode - Grayscale or RGB. RGB gets the full three-channel treatment.
• first_activation / middle_activations / last_activation - where the character lives. Start with Sin for the classic look. last_activation defaults to Sigmoid, which is what maps raw outputs into the 0–1 range.
• scale - zoom. Below 1 zooms in, above 1 zooms out. This is the fastest way to change the "wavelength" of the pattern.

The rest - mean, std, bias_*, output_factor, output_offset, offsets - are weight-init and output-shaping knobs that tune fine detail once you've got a seed you like. normalize (default off) rescales the output to use the full 0–1 range; turn it on if your pattern comes out dark or clipped.

One quirk: the schema declares an optional images input, but the current code never actually consumes it. Don't wire anything in expecting it to do something.

Install

From the Quasimondo pack - ComfyUI Manager → search ComfyUI-QuasimondoNodes, or:

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

The node itself is pure torch/numpy - it doesn't need the pack's moderngl or even OpenCV. No checkpoint, no LoRA, no safetensors anywhere.

Where people get burned

The two traps are both about cost. First, the default is 1024×1024 with a 3-layer network - fine as a still, but if you crank framecount to 64 at full res on CPU, it stops being instant. Start small, confirm the look, then scale. Second, normalize is off by default; without it, patterns built around Sigmoid output are already in range, but switch the last activation to something like ReLU and you'll get blown-out highlights until you enable it.

The other thing to know: this is a procedural-art toy, not a diffusion model. It will never make you a photorealistic anything. As a texture/background/latent-walk machine it's a joy; expecting a model is missing the point.