Where can I find a pre-trained variational autoencoder that's trained on natural images?

Does anyone know a pre-trained variational autoencoder (VAE) or a VAE-GAN that's trained on natural images?

I have been searching for a variational autoencoder that is trained on natural images (e.g. ImageNet) where I can extract latent features from. It could be implemented in any framework, preferably MXNet, Chainer or PyTorch.