List your company
ResearchFormerly WizWand

Generative Modeling on ImageNet 32x32 downsampled

3.77Bits Per DimensionImage Transformer
3.74483.91494.0854.2551Jul 9, 2018Nov 7, 2018Mar 9, 2019Jul 9, 2019Nov 7, 2019Mar 8, 2020Jul 8, 2020
Updated 8mo ago

Evaluation results

MethodLinks
Image Transformer2018.123.77
δ-VAEResolution=32x322020.073.77
Image TransformerResolution=32x322020.073.77
Decoder baselineSubscaling=None2018.123.79
PixelSNAIL2018.123.8
PixelSNAILResolution=32x322020.073.8
Gated PixelCNN2018.123.83
Gated PixelCNNResolution=32x322020.073.83
SPNSlices=16x162018.123.85
SPNResolution=32x322020.073.85
Flow++Resolution=32x322020.073.86
PixelRNNResolution=32x322020.073.86
SPNSlices=8x82018.123.91
NVAEResolution=32x32, Flow=With flow2020.073.92
ANFResolution=32x322020.073.92
Parallel Multiscale2018.123.95
BIVAResolution=32x322020.073.96
Residual flowResolution=32x322020.074.01
GlowK=32, L=3, resolution=32x32, permutation=invertible 1 x 1 convolution2018.074.09
Glow2018.124.09
GLOWResolution=32x322020.074.09
RealNVPK=32, L=3, resolution=32x322018.074.28
Real NVPResolution=32x322020.074.28
Conv DrawResolution=32x322020.074.4