StyleGAN Exploration

Compact StyleGAN implementation in TensorFlow/Keras for image generation on CIFAR-10, in a single Jupyter notebook. An 8-layer mapping network and AdaIN style modulation drive synthesis, with style mixing and latent interpolation experiments. An educational implementation, not a reproduction of StyleGAN2 or StyleGAN3.

StyleGAN is hard to learn from its papers alone, so a small, readable implementation helps show how the parts fit.

A compact StyleGAN in TensorFlow and Keras, trained on 20,000 CIFAR-10 images for 10 epochs. It has an 8-layer mapping network, AdaIN style modulation and a discriminator with softplus loss. After training, it runs style-mixing and latent-interpolation experiments.

An educational implementation. It does not reproduce StyleGAN2 or StyleGAN3, and the repository says so.