dagouchin / beta-vae-1

A Pytorch Implementation of the Beta-VAE

Face Generation Using Variational Autoencoders

This repo contains training code for two different VAEs implemented with Pytorch.
I used the CelebA Dataset for training, with 182637 training images and 19962 testing images.
Trained model can be found in /samples.

Model structures:

β-VAE [1]:

DFC-VAE [2]:

Results after 300 epochs:

Original Faces (Top) vs. Reconstructed (Bottom) :

Linear Interpolation from z1 (leftmost) to z2 (rightmost):

Vector Arithmetic from original (leftmost) to wearing sunglasses (rightmost):

Generated Images with random sampled latent z ~ N(0, 1):

(Notes: output images above are results after image dehazing using this script [3])

References

[1] β-VAE: https://openreview.net/pdf?id=Sy2fzU9gl
[2] DFC-VAE: https://arxiv.org/abs/1610.00291
[3] Dehaze: https://github.com/cssartori/image-dehazing.git