Conditional Variational Autoencoder for facial attribute manipulation

A Beta-CVAE for facial attribute manipulation using the CelebA dataset, incorporating a beta-variational objective to balance reconstruction fidelity and latent disentanglement. The model enabled conditional generation of facial images based on user-specified attributes. Its latent space exhibited controlled attribute-specific transformations, facilitating interpretability and targeted image synthesis. The model was trained on a subset of 10,000 out of 200,000+ images from the CelebA dataset. The training was conducted on Google Colab Pro using an NVIDIA A100 GPU.

A sufficiently disentangled representation while maintaining acceptable reconstruction quality.

Disentanglement was emphasized, enabling more interpretable latent spaces but at the cost of reduced reconstruction quality. Beta = 4

Generated faces to exhibit gradual addition of facial features.