{"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/polarity-sampling-quality-and-diversity","title":"Polarity Sampling: Quality and Diversity Control of Pre-Trained Generative Networks via Singular Values","arxiv_id":"2203.01993","date":"2022-03-03","proceeding":"CVPR 2022 1","authors":["Ahmed Imtiaz Humayun","Randall Balestriero","Richard Baraniuk"],"abstract":"We present Polarity Sampling, a theoretically justified plug-and-play method for controlling the generation quality and diversity of pre-trained deep generative networks DGNs). Leveraging the fact that DGNs are, or can be approximated by, continuous piecewise affine splines, we derive the analytical DGN output space distribution as a function of the product of the DGN's Jacobian singular values raised to a power $\\rho$. We dub $\\rho$ the $\\textbf{polarity}$ parameter and prove that $\\rho$ focuses the DGN sampling on the modes ($\\rho < 0$) or anti-modes ($\\rho > 0$) of the DGN output-space distribution. We demonstrate that nonzero polarity values achieve a better precision-recall (quality-diversity) Pareto frontier than standard methods, such as truncation, for a number of state-of-the-art DGNs. We also present quantitative and qualitative results on the improvement of overall generation quality (e.g., in terms of the Frechet Inception Distance) for a number of state-of-the-art DGNs, including StyleGAN3, BigGAN-deep, NVAE, for different conditional and unconditional image generation tasks. In particular, Polarity Sampling redefines the state-of-the-art for StyleGAN2 on the FFHQ Dataset to FID 2.57, StyleGAN2 on the LSUN Car Dataset to FID 2.27 and StyleGAN3 on the AFHQv2 Dataset to FID 3.95. Demo: bit.ly/polarity-samp","url_abs":"https://arxiv.org/abs/2203.01993v2","url_pdf":"https://arxiv.org/pdf/2203.01993v2.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":"polarity-sampling-quality-and-diversity","repo_url":"https://github.com/AhmedImtiazPrio/magnet-polarity","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok"}}],"tasks":[{"task_slug":"diversity","task_name":"Diversity"},{"task_slug":"image-generation","task_name":"Image Generation"},{"task_slug":"unconditional-image-generation","task_name":"Unconditional Image Generation"}],"methods":[{"method_slug":"1x1-convolution","method_name":"1x1 Convolution"},{"method_slug":"adam","method_name":"Adam"},{"method_slug":"average-pooling","method_name":"Average Pooling"},{"method_slug":"batch-normalization","method_name":"Batch Normalization"},{"method_slug":"biggan-deep","method_name":"BigGAN-deep"},{"method_slug":"bottleneck-residual-block","method_name":"Bottleneck Residual Block"},{"method_slug":"conditional-batch-normalization","method_name":"Conditional Batch Normalization"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"depthwise-convolution","method_name":"Depthwise Convolution"},{"method_slug":"depthwise-separable-convolution","method_name":"Depthwise Separable Convolution"},{"method_slug":"early-stopping","method_name":"Early Stopping"},{"method_slug":"feedforward-network","method_name":"Feedforward Network"},{"method_slug":"gan-hinge-loss","method_name":"GAN Hinge Loss"},{"method_slug":"linear-layer","method_name":"Linear Layer"},{"method_slug":"nvae","method_name":"NVAE"},{"method_slug":"nvae-encoder-residual-cell","method_name":"NVAE Encoder Residual Cell"},{"method_slug":"nvae-generative-residual-cell","method_name":"NVAE Generative Residual Cell"},{"method_slug":"non-local-block","method_name":"Non-Local Block"},{"method_slug":"non-local-operation","method_name":"Non-Local Operation"},{"method_slug":"normalizing-flows","method_name":"Normalizing Flows"},{"method_slug":"off-diagonal-orthogonal-regularization","method_name":"Off-Diagonal Orthogonal Regularization"},{"method_slug":"path-length-regularization","method_name":"Path Length Regularization"},{"method_slug":"pointwise-convolution","method_name":"Pointwise Convolution"},{"method_slug":"projection-discriminator","method_name":"Projection Discriminator"},{"method_slug":"r1-regularization","method_name":"R1 Regularization"},{"method_slug":"relu","method_name":"ReLU"},{"method_slug":"residual-connection","method_name":"Residual Connection"},{"method_slug":"residual-normal-distribution","method_name":"Residual Normal Distribution"},{"method_slug":"sagan","method_name":"SAGAN"},{"method_slug":"sigmoid-activation","method_name":"Sigmoid Activation"},{"method_slug":"softmax","method_name":"Softmax"},{"method_slug":"spectral-normalization","method_name":"Spectral Normalization"},{"method_slug":"squeeze-and-excitation-block","method_name":"Squeeze-and-Excitation Block"},{"method_slug":"ttur","method_name":"TTUR"},{"method_slug":"truncation-trick","method_name":"Truncation Trick"},{"method_slug":"weight-demodulation","method_name":"Weight Demodulation"},{"method_slug":"weight-normalization","method_name":"Weight Normalization"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-generation-on-afhqv2","task":"Image Generation","dataset":"AFHQV2","model":"Polarity-StyleGAN3","rank_in_archive_order":1,"of":7,"metrics":{"FID":"3.95"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-celeba-hq-1024x1024","task":"Image Generation","dataset":"CelebA-HQ 1024x1024","model":"Polarity-ProGAN","rank_in_archive_order":6,"of":10,"metrics":{"FID":"7.28"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-ffhq-1024-x-1024","task":"Image Generation","dataset":"FFHQ 1024 x 1024","model":"Polarity-StyleGAN2","rank_in_archive_order":3,"of":20,"metrics":{"FID":"2.57"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-imagenet-256x256","task":"Image Generation","dataset":"ImageNet 256x256","model":"Polarity-BigGAN","rank_in_archive_order":91,"of":94,"metrics":{"FID":"6.82"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-lsun-car-512-x-384","task":"Image Generation","dataset":"LSUN Car 512 x 384","model":"Polarity-StyleGAN2","rank_in_archive_order":1,"of":2,"metrics":{"FID":"2.27"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-lsun-cat-256-x-256","task":"Image Generation","dataset":"LSUN Cat 256 x 256","model":"Polarity-StyleGAN2","rank_in_archive_order":4,"of":8,"metrics":{"FID":"6.34"},"uses_additional_data":false},{"leaderboard":"/sota/image-generation-on-lsun-churches-256-x-256","task":"Image Generation","dataset":"LSUN Churches 256 x 256","model":"Polarity-StyleGAN2","rank_in_archive_order":11,"of":27,"metrics":{"FID":"3.92"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/2203.01993","atlas_url":"https://app.syntology.ai/?focus=2203.01993","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}