This repository contains the implementation and experiments for the course project “Towards Controllable High-Quality Image Generation.”
The project explores improving both image quality and fine-grained control in generative models by combining insights from VAEs, GANs, and contrastive learning.
Generative models like VAEs, GANs, and Diffusion models have achieved impressive quality but often lack controllability in their latent spaces.
This project surveys advanced architectures and implements methods to enhance controllability while preserving fidelity.
Key areas explored:
- Variational Autoencoders (VAEs) and extensions (DiffuseVAE, NoisyVAE).
- Generative Adversarial Networks: GAN, Conditional GAN, ContraGAN, StyleGAN.
- Latent-space manipulation using PCA and contrastive losses.
- Surveyed literature on GANs, Conditional GANs, and ContraGAN, analyzing techniques for controllable image generation.
- Implemented a Contrastive VAE with contrastive loss (inspired by ContraGAN) to improve latent space disentanglement.
- Developed a Conditional GAN for class-controlled image generation on MNIST, enabling attribute-level control.
- Contrastive VAE produced more separated and interpretable latent vectors compared to standard VAE.
- Conditional GAN generated sharp, class-specific MNIST digits with stable training.
- Demonstrated controllable manipulations (e.g., stroke thickness, position) in latent space.
- Rajiv Chitale
- Vedant Bhandare
- Umanshiva Ladva
- Nitya Bhamidipaty