Papers › Efficient-VDVAE: Less is more

Efficient-VDVAE: Less is more

25 Mar 2022arXiv:2203.13751archive 2025-07-28

Louay Hazami, Rayhane Mama, Ragavan Thurairatnam

Hierarchical VAEs have emerged in recent years as a reliable option for maximum likelihood estimation. However, instability issues and demanding computational requirements have hindered research progress in the area. We present simple modifications to the Very Deep VAE to make it converge up to 2.6× faster, save up to 20× in memory load and improve stability during training. Despite these changes, our models achieve comparable or better negative log-likelihood performance than current state-of-the-art models on all $7$ commonly used image datasets we evaluated on. We also make an argument against using 5-bit benchmarks as a way to measure hierarchical VAE's performance due to undesirable biases caused by the 5-bit quantization. Additionally, we empirically demonstrate that roughly 3% of the hierarchical VAE's latent space dimensions is sufficient to encode most of the image information, without loss of performance, opening up the doors to efficiently leverage the hierarchical VAEs' latent space in downstream tasks. We release our source code and models at https://github.com/Rayhane-mamah/Efficient-VDVAE .

Paper PDF Code

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

Rayhane-mamah/Efficient-VDVAE officialmentioned in papermentioned on GitHubjaxMIT report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Image Generation Quantization

Results from the paper archive 2025-07-28

Task	Dataset	Model	Metric	Value	Rank at snapshot	Leaderboard	Report
Image Generation	Binarized MNIST	Efficient-VDVAE	nats	79.09	#4 of 10	Archive leaderboard	report
Image Generation	CelebA 256x256	Efficient-VDVAE	bpd	0.51	#1 of 17	Archive leaderboard	report
Image Generation	CelebA 256x256	Efficient-VDVAE	bpd (8-bits)	1.35	#1 of 17	Archive leaderboard	report
Image Generation	CelebA 64x64	Efficient-VDVAE	bits/dimension	1.83	#36 of 39	Archive leaderboard	report
Image Generation	CelebA-HQ 1024x1024	Efficient-VDVAE	bits/dimension	1.01	#10 of 10	Archive leaderboard	report
Image Generation	FFHQ 1024 x 1024	Efficient-VDVAE	bits/dimension	2.30	#19 of 20	Archive leaderboard	report
Image Generation	FFHQ 256 x 256	Efficient-VDVAE	FID	34.88	#40 of 51	Archive leaderboard	report
Image Generation	FFHQ 256 x 256	Efficient-VDVAE	bits/dimension	0.53	#40 of 51	Archive leaderboard	report
Image Generation	FFHQ 256 x 256	Efficient-VDVAE (DINOv2)	FD	514.16	#47 of 51	Archive leaderboard	report
Image Generation	FFHQ 256 x 256	Efficient-VDVAE (DINOv2)	Precision	0.86	#47 of 51	Archive leaderboard	report
Image Generation	FFHQ 256 x 256	Efficient-VDVAE (DINOv2)	Recall	0.14	#47 of 51	Archive leaderboard	report
Image Generation	ImageNet 64x64	Efficient-VDVAE	Bits per dim	3.30 (different downsampling)	#37 of 65	Archive leaderboard	report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

1x1 Convolution AdaMax Convolution Cosine Annealing Hierarchical VAE Mixture of Logistic Distributions Sigmoid Activation Stochastic Depth

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections