Comparative study of DCGAN, Progressive GAN, and diffusion models on faces, synthetic colored squares, and a custom animal image dataset.
Highlights
Trained a DCGAN on standard face datasets and a custom 64×64 colored-squares dataset to study sample complexity and mode collapse.
Collected a 1k–5k image dataset of a chosen animal at 256×256 resolution and trained both DCGAN and Progressive GAN with augmentation and architecture tweaks.
Implemented a diffusion model on the same animal dataset, visualizing the forward noising and reverse denoising processes and comparing sample quality against GAN-based approaches.