art_gan
sideGAN that generates modern art.
A DCGAN trained from scratch to generate Cubist-style paintings, built as a full technical walkthrough rather than a demo repo.
Training images came from WikiArt’s Cubism collection, resized to 128×128 and normalized to [-1, 1]. The discriminator is a five-layer convolutional stack (32→64→128→256→512 filters, stride 2, LeakyReLU, batch norm, dropout) collapsing to a single sigmoid output. The generator starts from a 100-dimensional noise vector, projects to a 4×4×256 volume, and upsamples back to 128×128 through repeated UpSampling2D + Conv2D + batch-norm blocks ending in tanh. Both trained adversarially with Adam (lr 1.5e-4, β1 0.5) on a V100 through Spell’s hosted GPU platform, checkpointing generated samples every 100 of 10,000 epochs.
Published as a full walkthrough on Towards Data Science, covering dataset prep, the architecture, the training loop, and running the whole thing on rented GPU hardware.