{"url":"/sota/image-generation-on-imagenet-32x32","task":{"name":"Image Generation","url":"/task/image-generation","note":null},"dataset":{"name":"ImageNet 32x32","url":null},"category":"Computer Vision","categories":["Computer Vision","Medical","Miscellaneous","Natural Language Processing"],"category_note":null,"description":"**Image Generation** (synthesis) is the task of generating new images from an existing dataset.\r\n\r\n- **Unconditional generation** refers to generating samples unconditionally from the dataset, i.e. $p(y)$\r\n- **[Conditional image generation](/task/conditional-image-generation)** (subtask) refers to generating samples conditionally from the dataset, based on a label, i.e. $p(y|x)$.\r\n\r\nIn this section, you can find state-of-the-art leaderboards for **unconditional generation**. For conditional  generation, and other types of image generations, refer to the subtasks.\r\n\r\n<span style=\"color:grey; opacity: 0.6\">( Image credit: [StyleGAN](https://github.com/NVlabs/stylegan) )</span>","description_from":"task","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","rank":"the archive's row order at snapshot; not re-ranked","rows_end_at":"2025-07-28","rows_withheld_as_spam":0,"metric_values":"the archive's strings, untouched"},"metrics":["FID","bpd","Inception score"],"metric_direction":{"note":"inferred from the metric name only (the archive records no direction); null = not inferred, chart draws points only","by_metric":{"FID":"lower","bpd":null,"Inception score":"higher"}},"counts":{"rows":35,"rows_with_code":29,"rows_with_paper_page":35,"rows_dated":35,"rows_using_additional_data":0},"rows":[{"rank_in_archive_order":1,"model":"PaGoDA","metrics":{"FID":"0.79"},"uses_additional_data":false,"paper_date":"2024-05-23","paper":"/paper/pagoda-progressive-growing-of-a-one-step","paper_url":"https://arxiv.org/abs/2405.14822v2","paper_title":"PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher","code":"https://github.com/sony/pagoda","n_code_links":1,"syntology":{"n_ran":15,"n_unverified":2,"n_samples":17,"n_pointer_only_licence":0}},{"rank_in_archive_order":2,"model":"StyleGAN-XL","metrics":{"FID":"1.10"},"uses_additional_data":false,"paper_date":"2022-02-01","paper":"/paper/stylegan-xl-scaling-stylegan-to-large-diverse","paper_url":"https://arxiv.org/abs/2202.00273v2","paper_title":"StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets","code":"https://github.com/autonomousvision/stylegan-xl","n_code_links":2,"syntology":{"n_ran":14,"n_unverified":5,"n_samples":19,"n_pointer_only_licence":0}},{"rank_in_archive_order":3,"model":"R3GAN","metrics":{"FID":"1.27"},"uses_additional_data":false,"paper_date":"2025-01-09","paper":"/paper/the-gan-is-dead-long-live-the-gan-a-modern","paper_url":"https://arxiv.org/abs/2501.05441v1","paper_title":"The GAN is dead; long live the GAN! A Modern GAN Baseline","code":"https://github.com/brownvc/r3gan","n_code_links":1,"syntology":{"n_ran":9,"n_unverified":2,"n_samples":11,"n_pointer_only_licence":11}},{"rank_in_archive_order":4,"model":"DDPM-IP","metrics":{"FID":"2.66"},"uses_additional_data":false,"paper_date":"2023-01-27","paper":"/paper/input-perturbation-reduces-exposure-bias-in","paper_url":"https://arxiv.org/abs/2301.11706v3","paper_title":"Input Perturbation Reduces Exposure Bias in Diffusion Models","code":"https://github.com/forever208/ddpm-ip","n_code_links":1,"syntology":{"n_ran":2,"n_unverified":1,"n_samples":3,"n_pointer_only_licence":2}},{"rank_in_archive_order":5,"model":"FM","metrics":{"FID":"5.02","bpd":"3.53"},"uses_additional_data":false,"paper_date":"2022-10-06","paper":"/paper/flow-matching-for-generative-modeling","paper_url":"https://arxiv.org/abs/2210.02747v2","paper_title":"Flow Matching for Generative Modeling","code":"https://github.com/shivammehta25/Matcha-TTS","n_code_links":5,"syntology":{"n_ran":1,"n_unverified":3,"n_samples":4,"n_pointer_only_licence":3}},{"rank_in_archive_order":6,"model":"MDM-Prime","metrics":{"FID":"6.98","Inception score":"11.65"},"uses_additional_data":false,"paper_date":"2025-05-24","paper":"/paper/beyond-masked-and-unmasked-discrete-diffusion","paper_url":"https://arxiv.org/abs/2505.18495v1","paper_title":"Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":7,"model":"MDM","metrics":{"FID":"7.91","Inception score":"11.60"},"uses_additional_data":false,"paper_date":"2025-05-24","paper":"/paper/beyond-masked-and-unmasked-discrete-diffusion","paper_url":"https://arxiv.org/abs/2505.18495v1","paper_title":"Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":8,"model":"DDPM++ (VP, NLL) + ST","metrics":{"FID":"8.42","Inception score":"11.82","bpd":"3.85"},"uses_additional_data":false,"paper_date":"2021-06-10","paper":"/paper/score-matching-model-for-unbounded-data-score-1","paper_url":"https://arxiv.org/abs/2106.05527v5","paper_title":"Soft Truncation: A Universal Training Technique of Score-based Diffusion Model for High Precision Score Estimation","code":"https://github.com/Kim-Dongjun/Soft-Truncation","n_code_links":1,"syntology":{"n_ran":4,"n_unverified":8,"n_samples":12,"n_pointer_only_licence":0}},{"rank_in_archive_order":9,"model":"i-DODE","metrics":{"FID":"9.09","bpd":"3.43"},"uses_additional_data":false,"paper_date":"2023-05-06","paper":"/paper/improved-techniques-for-maximum-likelihood","paper_url":"https://arxiv.org/abs/2305.03935v4","paper_title":"Improved Techniques for Maximum Likelihood Estimation for Diffusion ODEs","code":"https://github.com/thu-ml/i-dode","n_code_links":1,"syntology":{"n_ran":4,"n_unverified":3,"n_samples":7,"n_pointer_only_licence":0}},{"rank_in_archive_order":10,"model":"MSGAN","metrics":{"FID":"12.3"},"uses_additional_data":false,"paper_date":"2019-11-16","paper":"/paper/self-supervised-gan-analysis-and-improvement-1","paper_url":"https://arxiv.org/abs/1911.06997v2","paper_title":"Self-supervised GAN: Analysis and Improvement with Multi-class Minimax Game","code":"https://github.com/tntrung/msgan","n_code_links":1,"syntology":null},{"rank_in_archive_order":11,"model":"DDPM","metrics":{"FID":"16.18","bpd":"3.89"},"uses_additional_data":false,"paper_date":"2020-06-19","paper":"/paper/denoising-diffusion-probabilistic-models","paper_url":"https://arxiv.org/abs/2006.11239v2","paper_title":"Denoising Diffusion Probabilistic Models","code":"https://github.com/labmlai/annotated_deep_learning_paper_implementations","n_code_links":70,"syntology":{"n_ran":178,"n_unverified":75,"n_samples":253,"n_pointer_only_licence":62}},{"rank_in_archive_order":12,"model":"NDM","metrics":{"FID":"17.02","bpd":"3.55"},"uses_additional_data":false,"paper_date":"2023-10-12","paper":"/paper/neural-diffusion-models","paper_url":"https://arxiv.org/abs/2310.08337v3","paper_title":"Neural Diffusion Models","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":13,"model":"QC-NCSN++","metrics":{"FID":"19.62","Inception score":"9.94"},"uses_additional_data":false,"paper_date":"2022-09-26","paper":"/paper/quasi-conservative-score-based-generative","paper_url":"https://arxiv.org/abs/2209.12753v3","paper_title":"On Investigating the Conservative Property of Score-Based Generative Models","code":"https://github.com/chen-hao-chao/qcsbm","n_code_links":1,"syntology":{"n_ran":2,"n_unverified":8,"n_samples":10,"n_pointer_only_licence":0}},{"rank_in_archive_order":14,"model":"NFDM","metrics":{"bpd":"3.34"},"uses_additional_data":false,"paper_date":"2024-04-19","paper":"/paper/neural-flow-diffusion-models-learnable","paper_url":"https://arxiv.org/abs/2404.12940v2","paper_title":"Neural Flow Diffusion Models: Learnable Forward Process for Improved Diffusion Modelling","code":"https://github.com/GrigoryBartosh/neural_diffusion","n_code_links":1,"syntology":null},{"rank_in_archive_order":15,"model":"BSI","metrics":{"bpd":"3.44"},"uses_additional_data":false,"paper_date":"2025-02-11","paper":"/paper/generative-modeling-with-bayesian-sample","paper_url":"https://arxiv.org/abs/2502.07580v2","paper_title":"Generative Modeling with Bayesian Sample Inference","code":"https://github.com/martenlienen/bsi","n_code_links":1,"syntology":null},{"rank_in_archive_order":16,"model":"DenseFlow-74-10","metrics":{"bpd":"3.63"},"uses_additional_data":false,"paper_date":"2021-06-08","paper":"/paper/densely-connected-normalizing-flows","paper_url":"https://arxiv.org/abs/2106.04627v3","paper_title":"Densely connected normalizing flows","code":"https://github.com/matejgrcic/DenseFlow","n_code_links":4,"syntology":null},{"rank_in_archive_order":17,"model":"VDM","metrics":{"bpd":"3.72"},"uses_additional_data":false,"paper_date":"2021-07-01","paper":"/paper/variational-diffusion-models","paper_url":"https://arxiv.org/abs/2107.00630v6","paper_title":"Variational Diffusion Models","code":"https://github.com/google-research/vdm","n_code_links":5,"syntology":{"n_ran":29,"n_unverified":9,"n_samples":38,"n_pointer_only_licence":2}},{"rank_in_archive_order":18,"model":"Hourglass","metrics":{"bpd":"3.74"},"uses_additional_data":false,"paper_date":"2021-10-26","paper":"/paper/hierarchical-transformers-are-more-efficient","paper_url":"https://arxiv.org/abs/2110.13711v2","paper_title":"Hierarchical Transformers Are More Efficient Language Models","code":"https://github.com/labmlai/annotated_deep_learning_paper_implementations","n_code_links":3,"syntology":{"n_ran":2,"n_unverified":2,"n_samples":4,"n_pointer_only_licence":2}},{"rank_in_archive_order":19,"model":"Reflected Diffusion","metrics":{"bpd":"3.74"},"uses_additional_data":false,"paper_date":"2023-04-10","paper":"/paper/reflected-diffusion-models","paper_url":"https://arxiv.org/abs/2304.04740v3","paper_title":"Reflected Diffusion Models","code":"https://github.com/louaaron/Reflected-Diffusion","n_code_links":1,"syntology":{"n_ran":0,"n_unverified":2,"n_samples":2,"n_pointer_only_licence":0}},{"rank_in_archive_order":20,"model":"ScoreFlow","metrics":{"bpd":"3.76"},"uses_additional_data":false,"paper_date":"2021-01-22","paper":"/paper/on-maximum-likelihood-training-of-score-based","paper_url":"https://arxiv.org/abs/2101.09258v4","paper_title":"Maximum Likelihood Training of Score-Based Diffusion Models","code":"https://github.com/yang-song/score_flow","n_code_links":3,"syntology":{"n_ran":5,"n_unverified":0,"n_samples":5,"n_pointer_only_licence":2}},{"rank_in_archive_order":21,"model":"Image Transformer","metrics":{"bpd":"3.77"},"uses_additional_data":false,"paper_date":"2018-02-15","paper":"/paper/image-transformer","paper_url":"http://arxiv.org/abs/1802.05751v3","paper_title":"Image Transformer","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":22,"model":"δ-VAE","metrics":{"bpd":"3.77"},"uses_additional_data":false,"paper_date":"2019-01-10","paper":"/paper/preventing-posterior-collapse-with-delta-vaes","paper_url":"http://arxiv.org/abs/1901.03416v1","paper_title":"Preventing Posterior Collapse with delta-VAEs","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":23,"model":"MRCNF","metrics":{"bpd":"3.77"},"uses_additional_data":false,"paper_date":"2021-06-15","paper":"/paper/multi-resolution-continuous-normalizing-flows","paper_url":"https://arxiv.org/abs/2106.08462v5","paper_title":"Multi-Resolution Continuous Normalizing Flows","code":"https://github.com/voletiv/mrcnf","n_code_links":1,"syntology":null},{"rank_in_archive_order":24,"model":"Very Deep VAE","metrics":{"bpd":"3.8"},"uses_additional_data":false,"paper_date":"2020-11-20","paper":"/paper/very-deep-vaes-generalize-autoregressive-1","paper_url":"https://arxiv.org/abs/2011.10650v2","paper_title":"Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images","code":"https://github.com/openai/vdvae","n_code_links":8,"syntology":{"n_ran":8,"n_unverified":1,"n_samples":9,"n_pointer_only_licence":4}},{"rank_in_archive_order":25,"model":"Gated PixelCNN","metrics":{"bpd":"3.83"},"uses_additional_data":false,"paper_date":"2016-06-16","paper":"/paper/conditional-image-generation-with-pixelcnn","paper_url":"http://arxiv.org/abs/1606.05328v2","paper_title":"Conditional Image Generation with PixelCNN Decoders","code":"https://github.com/openai/pixel-cnn","n_code_links":14,"syntology":{"n_ran":4,"n_unverified":10,"n_samples":14,"n_pointer_only_licence":3}},{"rank_in_archive_order":26,"model":"SPN Menick and Kalchbrenner (2019)","metrics":{"bpd":"3.85"},"uses_additional_data":false,"paper_date":"2018-12-04","paper":"/paper/generating-high-fidelity-images-with-subscale","paper_url":"http://arxiv.org/abs/1812.01608v1","paper_title":"Generating High Fidelity Images with Subscale Pixel Networks and Multidimensional Upscaling","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":27,"model":"Flow++","metrics":{"bpd":"3.86"},"uses_additional_data":false,"paper_date":"2019-02-01","paper":"/paper/flow-improving-flow-based-generative-models","paper_url":"https://arxiv.org/abs/1902.00275v2","paper_title":"Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design","code":"https://github.com/aravind0706/flowpp","n_code_links":4,"syntology":{"n_ran":0,"n_unverified":9,"n_samples":9,"n_pointer_only_licence":0}},{"rank_in_archive_order":28,"model":"PixelRNN","metrics":{"bpd":"3.86"},"uses_additional_data":false,"paper_date":"2016-01-25","paper":"/paper/pixel-recurrent-neural-networks","paper_url":"http://arxiv.org/abs/1601.06759v3","paper_title":"Pixel Recurrent Neural Networks","code":"https://github.com/EugenHotaj/pytorch-generative/blob/master/pytorch_generative/models/autoregressive/pixel_cnn.py","n_code_links":20,"syntology":{"n_ran":19,"n_unverified":10,"n_samples":29,"n_pointer_only_licence":18}},{"rank_in_archive_order":29,"model":"NVAE w/ flow","metrics":{"bpd":"3.92"},"uses_additional_data":false,"paper_date":"2020-07-08","paper":"/paper/nvae-a-deep-hierarchical-variational","paper_url":"https://arxiv.org/abs/2007.03898v3","paper_title":"NVAE: A Deep Hierarchical Variational Autoencoder","code":"https://github.com/NVlabs/NVAE","n_code_links":10,"syntology":{"n_ran":23,"n_unverified":18,"n_samples":41,"n_pointer_only_licence":23}},{"rank_in_archive_order":30,"model":"ANF Huang et al. (2020)","metrics":{"bpd":"3.92"},"uses_additional_data":false,"paper_date":"2020-02-17","paper":"/paper/augmented-normalizing-flows-bridging-the-gap","paper_url":"https://arxiv.org/abs/2002.07101v1","paper_title":"Augmented Normalizing Flows: Bridging the Gap Between Generative Flows and Latent Variable Models","code":"https://github.com/mj-will/augmented-flows","n_code_links":1,"syntology":{"n_ran":1,"n_unverified":0,"n_samples":1,"n_pointer_only_licence":0}},{"rank_in_archive_order":31,"model":"BIVA Maaloe et al. (2019)","metrics":{"bpd":"3.96"},"uses_additional_data":false,"paper_date":"2019-02-06","paper":"/paper/biva-a-very-deep-hierarchy-of-latent","paper_url":"https://arxiv.org/abs/1902.02102v3","paper_title":"BIVA: A Very Deep Hierarchy of Latent Variables for Generative Modeling","code":"https://github.com/vlievin/biva-pytorch","n_code_links":2,"syntology":null},{"rank_in_archive_order":32,"model":"Residual Flow","metrics":{"bpd":"4.01"},"uses_additional_data":false,"paper_date":"2019-06-06","paper":"/paper/residual-flows-for-invertible-generative","paper_url":"https://arxiv.org/abs/1906.02735v6","paper_title":"Residual Flows for Invertible Generative Modeling","code":"https://github.com/rtqichen/residual-flows","n_code_links":4,"syntology":{"n_ran":5,"n_unverified":3,"n_samples":8,"n_pointer_only_licence":0}},{"rank_in_archive_order":33,"model":"MintNet","metrics":{"bpd":"4.06"},"uses_additional_data":false,"paper_date":"2019-07-18","paper":"/paper/mintnet-building-invertible-neural-networks","paper_url":"https://arxiv.org/abs/1907.07945v2","paper_title":"MintNet: Building Invertible Neural Networks with Masked Convolutions","code":"https://github.com/ermongroup/mintnet","n_code_links":1,"syntology":null},{"rank_in_archive_order":34,"model":"Glow (Kingma and Dhariwal, 2018)","metrics":{"bpd":"4.09"},"uses_additional_data":false,"paper_date":"2018-07-09","paper":"/paper/glow-generative-flow-with-invertible-1x1","paper_url":"http://arxiv.org/abs/1807.03039v2","paper_title":"Glow: Generative Flow with Invertible 1x1 Convolutions","code":"https://github.com/openai/glow","n_code_links":27,"syntology":{"n_ran":75,"n_unverified":54,"n_samples":129,"n_pointer_only_licence":44}},{"rank_in_archive_order":35,"model":"Real NVP (Dinh et al., 2017)","metrics":{"bpd":"4.28"},"uses_additional_data":false,"paper_date":"2016-05-27","paper":"/paper/density-estimation-using-real-nvp","paper_url":"http://arxiv.org/abs/1605.08803v3","paper_title":"Density estimation using Real NVP","code":"https://github.com/tensorflow/models","n_code_links":35,"syntology":{"n_ran":43,"n_unverified":27,"n_samples":70,"n_pointer_only_licence":33}}],"since_archive":{"claim":"Results that newer papers report for their own method, placed here by Syntology. A model pointed at the cell in the paper's own table; the number was read from that cell and checked against this leaderboard's metric, dataset, split and scale; an independent check that saw this leaderboard's other rows and every other leaderboard on the same dataset accepted it. Not reviewed by the paper's authors or by the archive's editors, and not ranked against the archive rows.","extraction_file_present":true,"measurement":{"test_papers":883,"papers_with_output":881,"judged_true":108,"judged":110,"wilson95_lower":0.9361,"measured_on":"2026-09-24","frozen_commit":"0e3de0df94"},"measurement_note":"blind adjudication of accepted entries on a held-out split of archive papers, rules frozen before the test","coverage":{"sentence":"Syntology has checked 6,795 of the 9,581 papers on this site that are newer than the archive; results from the others appear after they are checked.","complete":false,"papers_newer_than_archive":9581,"papers_checked":6795,"papers_extracted_not_yet_verified":0,"boards_without_verdict":2,"papers_not_yet_extracted":2785},"order":"newest first by month (arXiv date, else the arXiv-id month), then arXiv id descending","columns":[],"entries":[]},"syntology":{"read_at":"2026-09-24T18:15:14+00:00","claim":"Per row: N of M harvested code samples from that row's paper executed on a synthesized fixture; the other M-N are unverified. Not a reproduction of the row's number; not a correctness claim. n_pointer_only_licence counts samples the site points at rather than redistributes (a licence axis, independent of ran/unverified).","rows_with_graph_line":22,"rows_with_any_sample_ran":20,"distinct_papers_with_graph_line":22,"distinct_papers_with_any_sample_ran":20,"samples_over_distinct_papers":{"n_ran":443,"n_unverified":252,"n_samples":695,"n_pointer_only_licence":209,"note":"each paper (arXiv id) counted once, however many rows it is behind; this is the page-level figure"},"samples_row_weighted":{"n_ran":443,"n_unverified":252,"n_samples":695,"n_pointer_only_licence":209,"note":"row-weighted: a paper behind several rows is counted once per row; inflated relative to samples_over_distinct_papers by design, kept for readers summing the per-row syntology blocks"}}}