Inspired from @hwalsuklee github repository and video, this presents the investigation/reimplementation of VAE/CVAE with different latent space size.
- MNIST
- CIFAR10 (#TODO)
- Pytorch + wandb
| Input | latent_dim=2 | latent_dim=5 | latent_dim=10 | latent_dim=20 | |
|---|---|---|---|---|---|
| VAE |  |
 |
 |
 |
 |
| CAVE |  |
 |
 |
 |
 |
| Manifold / Conditional Generation | Distribution | |
|---|---|---|
| VAE |  |
 |
| CVAE |  |
 |
| Video | VAE | CVAE |
|---|---|---|
| Manifold |  |
 |
Below result shows CVAE result of fixed latent space vector z, but with different condition (0-9 for MNIST)
 |
 |
|---|---|
 |
 |
 |
 |
 |
 |
 |
 |