{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/efficient-vdvae-less-is-more","title":"Efficient-VDVAE: Less is more","arxiv_id":"2203.13751","date":"2022-03-25","proceeding":null,"authors":["Louay Hazami","Rayhane Mama","Ragavan Thurairatnam"],"abstract":"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 $2.6\\times$ faster, save up to $20\\times$ 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 .","url_abs":"https://arxiv.org/abs/2203.13751v2","url_pdf":"https://arxiv.org/pdf/2203.13751v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"efficient-vdvae-less-is-more","repo_url":"https://github.com/Rayhane-mamah/Efficient-VDVAE","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"jax","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"image-generation","task_name":"Image Generation"},{"task_slug":"quantization","task_name":"Quantization"}],"methods":[{"method_slug":"1x1-convolution","method_name":"1x1 Convolution"},{"method_slug":"adamax","method_name":"AdaMax"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"cosine-annealing","method_name":"Cosine Annealing"},{"method_slug":"hierarchical-vae","method_name":"Hierarchical VAE"},{"method_slug":"mixture-of-logistic-distributions","method_name":"Mixture of Logistic Distributions"},{"method_slug":"sigmoid-activation","method_name":"Sigmoid Activation"},{"method_slug":"stochastic-depth","method_name":"Stochastic Depth"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-generation-on-binarized-mnist","task":"Image Generation","dataset":"Binarized MNIST","model":"Efficient-VDVAE","rank_in_archive_order":4,"of":10,"metrics":{"nats":"79.09"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-celeba-256x256","task":"Image Generation","dataset":"CelebA 256x256","model":"Efficient-VDVAE","rank_in_archive_order":1,"of":17,"metrics":{"bpd":"0.51","bpd (8-bits)":"1.35"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-celeba-64x64","task":"Image Generation","dataset":"CelebA 64x64","model":"Efficient-VDVAE","rank_in_archive_order":36,"of":39,"metrics":{"bits/dimension":"1.83"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-celeba-hq-1024x1024","task":"Image Generation","dataset":"CelebA-HQ 1024x1024","model":"Efficient-VDVAE","rank_in_archive_order":10,"of":10,"metrics":{"bits/dimension":"1.01"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-ffhq-1024-x-1024","task":"Image Generation","dataset":"FFHQ 1024 x 1024","model":"Efficient-VDVAE","rank_in_archive_order":19,"of":20,"metrics":{"bits/dimension":"2.30"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-ffhq-256-x-256","task":"Image Generation","dataset":"FFHQ 256 x 256","model":"Efficient-VDVAE","rank_in_archive_order":40,"of":51,"metrics":{"FID":"34.88","bits/dimension":"0.53"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-ffhq-256-x-256","task":"Image Generation","dataset":"FFHQ 256 x 256","model":"Efficient-VDVAE (DINOv2)","rank_in_archive_order":47,"of":51,"metrics":{"FD":"514.16","Precision":"0.86","Recall":"0.14"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-imagenet-64x64","task":"Image Generation","dataset":"ImageNet 64x64","model":"Efficient-VDVAE","rank_in_archive_order":37,"of":65,"metrics":{"Bits per dim":"3.30 (different downsampling)"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=2203.13751","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}