{"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/generating-high-fidelity-images-with-subscale","title":"Generating High Fidelity Images with Subscale Pixel Networks and Multidimensional Upscaling","arxiv_id":"1812.01608","date":"2018-12-04","proceeding":"ICLR 2019 5","authors":["Jacob Menick","Nal Kalchbrenner"],"abstract":"The unconditional generation of high fidelity images is a longstanding\nbenchmark for testing the performance of image decoders. Autoregressive image\nmodels have been able to generate small images unconditionally, but the\nextension of these methods to large images where fidelity can be more readily\nassessed has remained an open problem. Among the major challenges are the\ncapacity to encode the vast previous context and the sheer difficulty of\nlearning a distribution that preserves both global semantic coherence and\nexactness of detail. To address the former challenge, we propose the Subscale\nPixel Network (SPN), a conditional decoder architecture that generates an image\nas a sequence of sub-images of equal size. The SPN compactly captures\nimage-wide spatial dependencies and requires a fraction of the memory and the\ncomputation required by other fully autoregressive models. To address the\nlatter challenge, we propose to use Multidimensional Upscaling to grow an image\nin both size and depth via intermediate stages utilising distinct SPNs. We\nevaluate SPNs on the unconditional generation of CelebAHQ of size 256 and of\nImageNet from size 32 to 256. We achieve state-of-the-art likelihood results in\nmultiple settings, set up new benchmark results in previously unexplored\nsettings and are able to generate very high fidelity large scale samples on the\nbasis of both datasets.","url_abs":"http://arxiv.org/abs/1812.01608v1","url_pdf":"http://arxiv.org/pdf/1812.01608v1.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":[],"tasks":[{"task_slug":"decoder","task_name":"Decoder"},{"task_slug":"image-generation","task_name":"Image Generation"},{"task_slug":"high","task_name":"Vocal Bursts Intensity Prediction"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-generation-on-celeba-256x256","task":"Image Generation","dataset":"CelebA 256x256","model":"SPN Menick and Kalchbrenner (2019)","rank_in_archive_order":2,"of":17,"metrics":{"bpd":"0.61"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-imagenet-32x32","task":"Image Generation","dataset":"ImageNet 32x32","model":"SPN Menick and Kalchbrenner (2019)","rank_in_archive_order":26,"of":35,"metrics":{"bpd":"3.85"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-imagenet-64x64","task":"Image Generation","dataset":"ImageNet 64x64","model":"SPN","rank_in_archive_order":47,"of":65,"metrics":{"Bits per dim":"3.52"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/1812.01608","atlas_url":"https://app.syntology.ai/?focus=1812.01608","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}