{"id":"arxiv:2501.05441","title":"The GAN is dead; long live the GAN! A Modern GAN Baseline","date":"2025-01-09","section":"Self-supervised learning and vision","quality":0.03231,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/","source_url":"https://arxiv.org/abs/2501.05441","text_available":true,"authors":["Yiwen Huang","Aaron Gokaslan","Volodymyr Kuleshov","James Tompkin"],"outgoing":[{"source":"arxiv:2501.05441","target":"arxiv:1710.05941","evidence":"id","id":"arxiv:1710.05941","title":"Searching for Activation Functions","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8815985e018dbc2066450e5afbb9260b48d5b03a4f41bc23285755ccf3185454/"}],"incoming":[],"text_parts":[{"start":0,"end":15369,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html"},{"start":15369,"end":31006,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html"},{"start":31006,"end":46789,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html"},{"start":46789,"end":62595,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html"},{"start":62595,"end":78559,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html"},{"start":78559,"end":94370,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-6.html"},{"start":94370,"end":95296,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-7.html"}],"contents":[{"title":"The GAN is dead; long live the GAN! A Modern Baseline GAN","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"Abstract","offset":61,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"1 Introduction","offset":1144,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"2 Serving Two Masters: Stability and Diversity with RpGAN + R 1 + R 2 +R_{1}+R_{2}","offset":5438,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"2.1 Traditional GAN","offset":5899,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"2.2 Relativistic f f -GAN","offset":8032,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"2.3 Training Dynamics of RpGAN","offset":9755,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"Zero-centered gradient penalties.","offset":10935,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"Discussion.","offset":13672,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"2.4 A Practical Demonstration","offset":14586,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-1.html#text"},{"title":"3 A Roadmap to a New Baseline — R3GAN","offset":19519,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"Minimum baseline (Config B).","offset":20846,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"Well-behaved loss function (Config C).","offset":24336,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"General network modernization (Config D).","offset":24642,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"Neural network architecture details.","offset":26296,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"Bottleneck modernization (Config E).","offset":28233,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"4 Experiments Details","offset":29342,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"4.1 Roadmap Insights on FFHQ-256 Karras et al. [2019]","offset":29368,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-2.html#text"},{"title":"4.2 Mode Recovery — StackedMNIST Metz et al. [2016]","offset":31521,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"4.3 FID — FFHQ-256 Karras et al. [2019] (Optimized)","offset":32182,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"4.4 FID — FFHQ-64 Karras et al. [2022]","offset":32864,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"4.5 FID — CIFAR-10 Krizhevsky et al. [2009]","offset":34333,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"4.6 FID — ImageNet-32 Chrabaszcz et al. [2017]","offset":35779,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"4.7 FID — ImageNet-64 Chrabaszcz et al. [2017]","offset":36397,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"4.8 Recall","offset":37004,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"5 Discussion and Limitations","offset":38813,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"6 Conclusion","offset":41217,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Acknowledgements.","offset":41917,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Appendix A Local convergence","offset":42323,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Appendix B DiracRpGAN: A demonstration of non-convergence","offset":44298,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Summary.","offset":44360,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Definition B.1.","offset":44674,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Lemma B.2.","offset":45193,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Lemma B.3.","offset":46772,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-3.html#text"},{"title":"Lemma B.4.","offset":48032,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Appendix C General Convergence Results","offset":49466,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Summary.","offset":49509,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"C.1 Assumptions","offset":49903,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Assumption I.","offset":50206,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Assumption II.","offset":50612,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Assumption III.","offset":52326,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"C.2 Convergence","offset":53444,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Theorem.","offset":54120,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Lemma C.2.1.","offset":54642,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Lemma C.2.2.","offset":55130,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Lemma C.2.3.","offset":56085,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Lemma C.2.4.","offset":58101,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Lemma C.2.5.","offset":58730,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Lemma C.2.6.","offset":59945,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Lemma C.2.7.","offset":62361,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-4.html#text"},{"title":"Appendix D Hyperparameters, training configurations, and compute","offset":63249,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Training schedule.","offset":63718,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Dataset augmentation.","offset":64819,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Network capacity.","offset":65668,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Mixed precision training.","offset":66151,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Class conditioning.","offset":66546,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Stacked MNIST.","offset":67463,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Compute resources.","offset":67757,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Appendix E Negative Results and Future Work","offset":68132,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Appendix F Qualitative Results","offset":73702,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Appendix G Training Curves","offset":73737,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"NeurIPS Paper Checklist","offset":73768,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-5.html#text"},{"title":"Instructions for reporting errors","offset":94332,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/part-6.html#text"}],"text_code_points":95296,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a6fcdbec190497e2e52574b7cb2b998385ed1944920c17dc53b710d18f7ba152/source.md"}