Small PyTorch implementation of Compositional Pattern-Producing Networks (CPPNs). Supports two modes: random generation (procedural art from untrained weights) and single-image overfitting (CPPN as image compression demo).

A Compositional Pattern-Producing Network is a neural network used as a function from coordinates to values. Instead of being trained on data, it takes a point (x, y) and outputs a color (r, g, b). The full image is produced by querying every pixel independently.

These are the 3 ideas that make CPPNs more interesting:

• No training. Weights are initialized randomly. The network is never shown any image, the patterns emerge solely from the network's structure.
• Symmetric activations. Functions like sin, cos, tanh and Gaussians produce spatial regularities such as periodicity, symmetry and smooth gradients that pure ReLU networks cannot produce.
• Resolution-independent. Because the network maps coordinates rather than pixels, you can render the same pattern at any resolution. The 2048x2048 version is not an upscale of the 256x256, both are sampled directly from the same function. Note: vanilla CPPNs exhibit spectral bias, quality degrades when sampling at resolutions higher than the training grid.

This implementation also adds a latent vector z that acts as a seed for pattern variations and an optional radial input r that biases the network toward radial compositions.

CPPNs were introduced by Kenneth Stanley in 2007 as a representation for evolutionary art and neural network topology generation.

Training using recommended settings (default values) demonstrates compression (19x smaller than PNG target) at degraded but recognizable quality (PSNR 19.68 dB, still below JPEG quality).

This implementation uses Stanley's original 2007 architecture with Normal(0, 1) weight initialization. While effective for shallow networks (4-6 layers), this setup fails to train at greater depths.

I observed dead network behavior when configuring hidden_dim=128 with hidden_layers=8 in cppn.py:

• Training loss plateaus around iteration ~200 at a degenerate value
• Model output collapses to near-zero (near-black image with isolated colored pixels)
• PSNR drops to 15.91 dB (vs 19.68 dB at recommended hidden_dim=64, hidden_layers=6)

This is the dead network problem: signal saturation through deep sin/tanh layers causes vanishing gradients in the final sigmoid, preventing further learning. Modern INR architectures address this through specialized initialization (e.g., SIREN's bounded uniform init with ω₀=30) and Fourier feature encoding.

In short: vanilla CPPN scales cleanly up to ~6 layers. Going deeper requires architectural changes covered in recent INR literature.

git clone https://github.com/NFAsylum/cppn.git

pip install -r requirements.txt

Random mode

cd cppn
python main.py

Training mode

cd cppn
python train.py

• Generated images in random mode are saved to output/.
• Training snapshots and logs are saved to training_output/.

tileable: if image is tileable in all directions, default value is True.

sigma: affects how chaotic is the image, default value is random float between 0.3 and 3.

r_strength: affects how radial is the image, default value is random float between 0 and 10 (not used when tileable is active)

quantity: affects how many images are generated (all images will use the same model)

size presets (square images): 'xxxsmall':32, 'xxsmall':64, 'xsmall':128, 'small':256, 'medium':512, 'large':1024, 'huge':2048. Default preset is large

target_path: path for target image used in training

target_size: target image size used in training

render_size: image size for final rendering, (when bigger than target_size, exposes spectral bias)

iterations: amount of iterations for training

snapshot_iters: which iterations generate snapshot images

• Stanley, K. O. (2007). Compositional Pattern Producing Networks: A Novel Abstraction of Development. Genetic Programming and Evolvable Machines, 8(2), 131–162. DOI
• Ha, D. (2016). Generating Large Images from Latent Vectors. otoro.net — practical CPPN implementation reference.
• Sitzmann, V., Martel, J. N. P., Bergman, A. W., Lindell, D. B., & Wetzstein, G. (2020). Implicit Neural Representations with Periodic Activation Functions. NeurIPS 2020. arXiv:2006.09661 — SIREN paper, addresses the depth/initialization limitations discussed above.
• Vaidyanathan, K. et al. (2023). Random-Access Neural Compression of Material Textures. SIGGRAPH 2023. Project page — NVIDIA Neural Texture Compression, applied INR for game-ready textures.

MIT, see LICENSE.