{"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/joint-autoregressive-and-hierarchical-priors","title":"Joint Autoregressive and Hierarchical Priors for Learned Image Compression","arxiv_id":"1809.02736","date":"2018-09-08","proceeding":"NeurIPS 2018 12","authors":["David Minnen","Johannes Ballé","George Toderici"],"abstract":"Recent models for learned image compression are based on autoencoders,\nlearning approximately invertible mappings from pixels to a quantized latent\nrepresentation. These are combined with an entropy model, a prior on the latent\nrepresentation that can be used with standard arithmetic coding algorithms to\nyield a compressed bitstream. Recently, hierarchical entropy models have been\nintroduced as a way to exploit more structure in the latents than simple fully\nfactorized priors, improving compression performance while maintaining\nend-to-end optimization. Inspired by the success of autoregressive priors in\nprobabilistic generative models, we examine autoregressive, hierarchical, as\nwell as combined priors as alternatives, weighing their costs and benefits in\nthe context of image compression. While it is well known that autoregressive\nmodels come with a significant computational penalty, we find that in terms of\ncompression performance, autoregressive and hierarchical priors are\ncomplementary and, together, exploit the probabilistic structure in the latents\nbetter than all previous learned models. The combined model yields\nstate-of-the-art rate--distortion performance, providing a 15.8% average\nreduction in file size over the previous state-of-the-art method based on deep\nlearning, which corresponds to a 59.8% size reduction over JPEG, more than 35%\nreduction compared to WebP and JPEG2000, and bitstreams 8.4% smaller than BPG,\nthe current state-of-the-art image codec. To the best of our knowledge, our\nmodel is the first learning-based method to outperform BPG on both PSNR and\nMS-SSIM distortion metrics.","url_abs":"http://arxiv.org/abs/1809.02736v1","url_pdf":"http://arxiv.org/pdf/1809.02736v1.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":"joint-autoregressive-and-hierarchical-priors","repo_url":"https://github.com/Nikolai10/Checkerboard-Context-Model","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":null},{"paper_slug":"joint-autoregressive-and-hierarchical-priors","repo_url":"https://github.com/mandt-lab/improving-inference-for-neural-image-compression","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"MIT"}},{"paper_slug":"joint-autoregressive-and-hierarchical-priors","repo_url":"https://github.com/InterDigitalInc/CompressAI","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"image-compression","task_name":"Image Compression"},{"task_slug":"ms-ssim","task_name":"MS-SSIM"},{"task_slug":"ssim","task_name":"SSIM"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1809.02736","atlas_url":"https://app.syntology.ai/?focus=1809.02736","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}