Upload vqvae-4x-32.pt

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

vqvae-4x-32.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5a2b00c52cce2db1f11f3045258f2e397391c88962103d0c873cc92520aa470
+size 10022621