Image Generation On Imagenet 64X64

Metrics

Bits per dim

Results

Performance results of various models on this benchmark

Model Name	Bits per dim	Paper Title	Repository
DenseFlow-74-10	3.35 (different downsampling)	Densely connected normalizing flows	-
2-rectified flow++ (NFE=1)	-	Improving the Training of Rectified Flows	-
Performer (6 layers)	3.719	Rethinking Attention with Performers	-
GDD-I	-	Diffusion Models Are Innate One-Step Generators	-
Sparse Transformer 59M (strided)	3.44	Generating Long Sequences with Sparse Transformers	-
CTM (NFE 1)	-	Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion	-
CD (Diffusion + Distillation, NFE=2)	-	Consistency Models	-
CT (Direct Generation, NFE=1)	-	Consistency Models	-
MaCow (Unf)	3.75	MaCow: Masked Convolutional Generative Flow	-
TCM	-	Truncated Consistency Models	-
Very Deep VAE	3.52	Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images	-
CDM	-	Cascaded Diffusion Models for High Fidelity Image Generation	-
Combiner-Axial	3.42	Combiner: Full Attention Transformer with Sparse Computation Cost	-
GLIDE + CLS-FREE	-	Composing Ensembles of Pre-trained Models via Iterative Consensus	-
SiD	-	Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation	-
RIN	-	Scalable Adaptive Computation for Iterative Generation	-
Logsparse (6 layers)	4.351	Enhancing the Locality and Breaking the Memory Bottleneck of Transformer on Time Series Forecasting	-
Efficient-VDVAE	3.30 (different downsampling)	Efficient-VDVAE: Less is more	-
MRCNF	3.44	Multi-Resolution Continuous Normalizing Flows	-
Gated PixelCNN (van den Oord et al., [2016c])	3.57	Conditional Image Generation with PixelCNN Decoders	-

0 of 64 row(s) selected.