VQ-VAE vs KL-based VAE

[samuruph]

Hi! Thank you for your work!
I was wondering why you choose VQ-VAE over KL VAE, since I have noticed you had implemented both versions here.
Which was the reason why you used VQ-VAE? I presume model complexity and/or inference time?
Moreover, have you pre-trained AutoencoderKL on nuscenes, for LiDAR reconstruction?

Thank you in advance for you help!