Larger, more compressed 3-sclae vqvae, trained on large dataset (including pixel arts), with deeper architecture, 4x spatial compression and latent dimension=32.
It efficently provides a compression of input image by a factor of 1.5, meanwhile keeping same performance as before
