pith. machine review for the scientific record. sign in

arxiv: 2604.20130 · v1 · submitted 2026-04-22 · 💻 cs.LG · cs.CV

Recognition: unknown

Pairing Regularization for Mitigating Many-to-One Collapse in GANs

Kuan-Yu Lin , Yu-Chih Huang , Tie Liu
Authors on Pith no claims yet

Pith reviewed 2026-05-10 00:52 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords mode collapseintra-mode collapsepairing regularizationGAN traininggenerative adversarial networkslatent consistencyprecision recall
0
0 comments X

The pith

A pairing regularizer mitigates many-to-one collapse in GANs by enforcing local consistency between latent variables and generated samples, with effects depending on the dominant training failure mode.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a pairing regularizer added to the generator objective to target intra-mode collapse, the case where many latent points map to nearly identical outputs. This regularizer promotes local consistency, and experiments show its benefits shift with training conditions: it drives structured exploration and better coverage when collapse is likely, while it sharpens the output distribution and raises precision once training has stabilized. The method is presented as a lightweight addition that works alongside existing stabilization tricks rather than replacing them. A reader would care because mode collapse still limits how reliably GANs produce diverse, high-quality samples even after other fixes are applied.

Core claim

We introduce a pairing regularizer jointly optimized with the generator that mitigates many-to-one collapse by enforcing local consistency between latent variables and generated samples. In collapse-prone regimes with limited exploration the regularizer encourages structured local exploration and raises recall; under stabilized training with sufficient exploration it discourages redundant mappings and improves precision without loss of recall. Experiments on toy distributions and real-image benchmarks confirm that the regularizer complements existing stabilization techniques by directly addressing intra-mode collapse.

What carries the argument

The pairing regularizer, which enforces local consistency between latent variables and generated samples to reduce many-to-one mappings.

If this is right

  • In collapse-prone regimes the regularizer improves coverage and raises recall.
  • Under stabilized training the regularizer improves precision without reducing recall.
  • The regularizer can be added to existing GAN stabilization methods to address intra-mode collapse directly.
  • The same pairing approach yields measurable gains on both toy distributions and real-image datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The regime-dependent behavior implies that monitoring collapse indicators during training could decide when to activate the regularizer.
  • Pairing ideas might transfer to other latent-variable models that suffer from output redundancy.
  • The local-consistency term could be combined with latent-space regularization techniques to further control mapping diversity.

Load-bearing premise

The pairing regularizer can be jointly optimized with the standard GAN objective without introducing new training instabilities or requiring regime-specific hyperparameter schedules.

What would settle it

Run identical GAN training runs with and without the pairing term on a benchmark known to exhibit intra-mode collapse and measure whether recall rises in unstable settings or precision rises in stable settings while recall stays flat.

Figures

Figures reproduced from arXiv: 2604.20130 by Kuan-Yu Lin, Tie Liu, Yu-Chih Huang.

Figure 1
Figure 1. Figure 1: Two-dimensional Gaussian mixture illustrating inter- and intra-mode collapse. (a) Standard GAN exhibits inter-mode collapse, failing to cover all mixture components. (b) Adding gradient penalty (R1) avoids mode dropping but results in highly concentrated samples within each mode, indicating persistent many-to-one (intra-mode) collapse. (c) Pairing regularization increases intra-mode diversity and mitigates… view at source ↗
Figure 2
Figure 2. Figure 2: Two-dimensional ring distribution illustrating many-to-one collapse in high-recall regime. All methods achieve similar coverage of the data manifold, yet differ in how probability mass is distributed along the ring. (a) Standard GAN produces uneven and concentrated samples, indicating many-to-one collapse. (b) Gradient penalty improves training stability but does not fully resolve intra-mode collapse. (c) … view at source ↗
Figure 3
Figure 3. Figure 3: R3GAN on synthetic data examples. R3GAN still suffers from many-to-one collapse in both examples. which augments the baseline with the proposed pairing reg￾ularization. Unless otherwise specified, all methods share the same generator and discriminator architectures and are trained using identical optimization settings. For image ex￾periments, we adopt the StyleGAN2 training framework and apply regularizers… view at source ↗
Figure 4
Figure 4. Figure 4: Training trajectories from 1M to 8M images without data augmentation. Results are averaged over three random seeds. Pairing-GAN yields competitive precision while maintaining similar recall (left), and achieves noticeably higher coverage (right), reducing many-to-one collapse effects in the no-aug, collapse-prone regime [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training trajectories from 1M to 25M images with data argumentation. Results are averaged over three random seeds. Pairing￾GAN and StyleGAN-ADA show highly similar trajectories in both Precision–Recall and Precision–Coverage spaces. Pairing-GAN achieves comparable recall and coverage, with marginally higher precision at each training stage [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Toy 2D Gaussian grid task with 25 modes. Standard GAN and R1/R2 regularization suffer from many-to-one collapse, resulting in severe mode dropping and poor intra-mode coverage. Pairing regularization encourages latent–sample correspondence, leading to improved mode coverage and more uniform support across modes. (a) Target (b) R3GAN (c) MS-GAN (d) Pairing-GAN [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Toy 2D Gaussian grid task with 25 modes arranged in a 5x5 layout. (a) Target distribution. (b) R3GAN (Huang et al., 2024). (c) MS-GAN (Mao et al., 2019). (d) Pairing-GAN (ours) [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Pairing-GAN generated image on conditional CIFAR-10. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
read the original abstract

Mode collapse remains a fundamental challenge in training generative adversarial networks (GANs). While existing works have primarily focused on inter-mode collapse, such as mode dropping, intra-mode collapse-where many latent variables map to the same or highly similar outputs-has received significantly less attention. In this work, we propose a pairing regularizer jointly optimized with the generator to mitigate the many-to-one collapse by enforcing local consistency between latent variables and generated samples. We show that the effect of pairing regularization depends on the dominant failure mode of training. In collapse-prone regimes with limited exploration, pairing encourages structured local exploration, leading to improved coverage and higher recall. In contrast, under stabilized training with sufficient exploration, pairing refines the generator's induced data density by discouraging redundant mappings, thereby improving precision without sacrificing recall. Extensive experiments on both toy distributions and real-image benchmarks demonstrate that the proposed regularizer effectively complements existing stabilization techniques by directly addressing intra-mode collapse.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a pairing regularizer that is jointly optimized with the standard GAN generator objective to mitigate intra-mode (many-to-one) collapse. It claims that the regularizer produces qualitatively different benefits depending on the dominant training failure mode: in collapse-prone regimes it promotes structured local exploration and raises recall, while in stabilized regimes it discourages redundant mappings and raises precision without harming recall. The method is evaluated on toy distributions and real-image benchmarks and is presented as complementary to existing stabilization techniques.

Significance. Intra-mode collapse has received less attention than inter-mode collapse; a simple, jointly optimized regularizer that demonstrably improves coverage or fidelity according to training regime would be a useful practical contribution. If the regime-dependent mechanism is shown to be robust and the improvements are supported by quantitative metrics and ablations, the work could influence how practitioners combine regularization with existing GAN stabilizers.

major comments (2)
  1. [Abstract] The central claim that the pairing regularizer yields recall gains in collapse-prone regimes and precision gains in stabilized regimes is load-bearing, yet the manuscript provides no a-priori diagnostic for identifying the dominant failure mode nor a schedule for the pairing coefficient. Without such a mechanism, the reported improvements cannot be unambiguously attributed to the claimed interaction rather than to post-hoc hyper-parameter selection (see Abstract and the description of joint optimization).
  2. [Abstract] The weakest assumption—that the regularizer can be added to the standard GAN objective without introducing new instabilities or requiring regime-specific tuning—is not supported by any analysis of training dynamics or sensitivity to the pairing strength. If the coefficient is fixed across all experiments, the regime-dependent narrative remains non-operational.
minor comments (2)
  1. [Abstract] The abstract states that 'extensive experiments' were performed but supplies no quantitative metrics, ablation tables, or implementation details; these should be summarized with specific numbers and baselines even in the abstract.
  2. Notation for the pairing loss and its integration into the generator objective should be introduced with an equation in the main text rather than left implicit.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting key aspects of our claims on the pairing regularizer. We address each major comment below and propose revisions to clarify the regime-dependent effects and provide supporting analyses.

read point-by-point responses

  1. Referee: [Abstract] The central claim that the pairing regularizer yields recall gains in collapse-prone regimes and precision gains in stabilized regimes is load-bearing, yet the manuscript provides no a-priori diagnostic for identifying the dominant failure mode nor a schedule for the pairing coefficient. Without such a mechanism, the reported improvements cannot be unambiguously attributed to the claimed interaction rather than to post-hoc hyper-parameter selection (see Abstract and the description of joint optimization).

    Authors: We agree that an explicit a-priori diagnostic would strengthen attribution of the observed effects. In the revised manuscript, we will add a dedicated subsection discussing practical identification of the dominant failure mode, for example via monitoring of recall or sample diversity metrics during training. The pairing coefficient is fixed across all reported experiments to emphasize general applicability without regime-specific schedules; we will include a new ablation study demonstrating robustness of the benefits over a range of coefficient values. These additions will help distinguish the regularizer's interaction with training regimes from hyperparameter selection. revision: yes

  2. Referee: [Abstract] The weakest assumption—that the regularizer can be added to the standard GAN objective without introducing new instabilities or requiring regime-specific tuning—is not supported by any analysis of training dynamics or sensitivity to the pairing strength. If the coefficient is fixed across all experiments, the regime-dependent narrative remains non-operational.

    Authors: We acknowledge the absence of explicit training dynamics analysis in the current version. The revised manuscript will add training curve visualizations and sensitivity ablations on the pairing strength to confirm that the regularizer integrates without introducing instabilities and that a single fixed coefficient produces the reported benefits in both collapse-prone and stabilized setups. The regime-dependent narrative is operationalized through separate experimental protocols (base GAN with vs. without stabilization), where quantitative recall and precision metrics show the distinct effects; the added analyses will further substantiate this without requiring per-regime retuning. revision: yes

Circularity Check

0 steps flagged

Empirical proposal with no derivation chain or self-referential predictions

full rationale

The paper introduces a pairing regularizer as a practical addition to the GAN objective and reports its observed effects on recall and precision across different training regimes. These effects are presented as empirical outcomes from experiments on toy distributions and image benchmarks, not as quantities derived from equations, fitted parameters renamed as predictions, or self-citations that close a logical loop. No mathematical derivation, uniqueness theorem, or ansatz is invoked in the provided text, so the central claims remain independent of the inputs by construction. The regime-dependent distinction is framed observationally rather than as a load-bearing prediction that reduces to the regularizer itself.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no mathematical formulation, free parameters, axioms, or new entities are specified.

pith-pipeline@v0.9.0 · 5461 in / 971 out tokens · 56278 ms · 2026-05-10T00:52:58.501697+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

86 extracted references · 11 canonical work pages · 2 internal anchors

  1. [2]

    2009 , journal=

    Learning multiple layers of features from tiny images , author=. 2009 , journal=

    2009

  2. [3]

    Advances in neural information processing systems , volume=

    UniGAN: Reducing mode collapse in GANs using a uniform generator , author=. Advances in neural information processing systems , volume=

  3. [4]

    International Conference on Artificial Intelligence and Statistics , pages=

    Testing Generated Distributions in GANs to Penalize Mode Collapse , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2024 , organization=

    2024

  4. [6]

    Artificial Intelligence , volume=

    DivGAN: A diversity enforcing generative adversarial network for mode collapse reduction , author=. Artificial Intelligence , volume=. 2023 , publisher=

    2023

  5. [7]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Seeing what a gan cannot generate , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  6. [10]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Scaling up gans for text-to-image synthesis , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  7. [11]

    Advances in neural information processing systems , volume=

    Differentiable augmentation for data-efficient gan training , author=. Advances in neural information processing systems , volume=

  8. [12]

    Advances in neural information processing systems , volume=

    Improved precision and recall metric for assessing generative models , author=. Advances in neural information processing systems , volume=

  9. [13]

    International conference on machine learning , pages=

    Reliable fidelity and diversity metrics for generative models , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  10. [15]

    International conference on machine learning , pages=

    Which training methods for GANs do actually converge? , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  11. [16]

    Advances in neural information processing systems , volume=

    Infogan: Interpretable representation learning by information maximizing generative adversarial nets , author=. Advances in neural information processing systems , volume=

  12. [17]

    Advances in neural information processing systems , volume=

    Improved training of wasserstein gans , author=. Advances in neural information processing systems , volume=

  13. [18]

    2017 , eprint=

    Wasserstein GAN , author=. 2017 , eprint=

    2017

  14. [19]

    arXiv preprint arXiv:1807.00734 , year =

    The Relativistic Discriminator: A Key Element Missing from Standard GAN , author =. arXiv preprint arXiv:1807.00734 , year =

    work page arXiv

  15. [20]

    Advances in Neural Information Processing Systems , volume=

    The gan is dead; long live the gan! a modern gan baseline , author=. Advances in Neural Information Processing Systems , volume=

  16. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  17. [22]

    On the Convergence and Stability of

    Mescheder, Lars and Geiger, Andreas and Nowozin, Sebastian , booktitle =. On the Convergence and Stability of. 2018 , editor =

    2018

  18. [23]

    Advances in Neural Information Processing Systems , year=

    Assessing Generative Models via Precision and Recall , author=. Advances in Neural Information Processing Systems , year=

  19. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Training Generative Adversarial Networks with Limited Data , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  20. [26]

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , booktitle=

  21. [27]

    Consistency Models

    Consistency Models , author=. arXiv:2303.01469 , year=

    work page internal anchor Pith review arXiv

  22. [28]

    ICLR , year=

    Flow Matching for Generative Modeling , author=. ICLR , year=

  23. [29]

    Large Scale GAN Training for High Fidelity Natural Image Synthesis

    Large Scale GAN Training for High Fidelity Natural Image Synthesis , author=. arXiv preprint arXiv:1809.11096 , year=

    work page internal anchor Pith review arXiv

  24. [30]

    Jiang, Yifan and Chang, Shiyu and Wang, Zhangyang , booktitle=

  25. [31]

    Lee, Kwonjoon and Chang, Huiwen and Jiang, Lu and Zhang, Han and Tu, Zhuowen and Liu, Ce , booktitle=

  26. [32]

    Tackling the Generative Learning Trilemma with Denoising Diffusion

    Xiao, Zhisheng and Kreis, Karsten and Vahdat, Arash , booktitle=. Tackling the Generative Learning Trilemma with Denoising Diffusion

  27. [33]

    Wang, Zhendong and Zheng, Huangjie and He, Pengcheng and Chen, Weizhu and Zhou, Mingyuan , booktitle=

  28. [34]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Alias-Free Generative Adversarial Networks , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  29. [35]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Denoising Diffusion Probabilistic Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  30. [36]

    International Conference on Learning Representations (ICLR) , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations (ICLR) , year=

  31. [37]

    International Conference on Learning Representations (ICLR) , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations (ICLR) , year=

  32. [38]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Improved Techniques for Training Score-Based Generative Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  33. [39]

    The Role of ImageNet Classes in Fr

    Kynk. The Role of ImageNet Classes in Fr. International Conference on Learning Representations (ICLR) , year=

  34. [40]

    Sauer, Axel and Schwarz, Katja and Geiger, Andreas , booktitle=

  35. [41]

    2025 , eprint=

    The GAN is dead; long live the GAN! A Modern GAN Baseline , author=. 2025 , eprint=

    2025

  36. [42]

    International Conference on Learning Representations (ICLR) , year=

    Unsupervised representation learning with deep convolutional generative adversarial networks , author=. International Conference on Learning Representations (ICLR) , year=

  37. [43]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    VEEGAN: Reducing mode collapse in GANs using implicit variational learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  38. [44]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Improved training of Wasserstein GANs , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  39. [45]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    PacGAN: The power of two samples in generative adversarial networks , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  40. [46]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Analyzing and improving the image quality of StyleGAN , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  41. [47]

    International Conference on Learning Representations (ICLR) , year=

    PresGAN: Towards probabilistic generative adversarial networks , author=. International Conference on Learning Representations (ICLR) , year=

  42. [48]

    Uncertainty in Artificial Intelligence (UAI) , year=

    Sliced score matching: A scalable approach to density and score estimation , author=. Uncertainty in Artificial Intelligence (UAI) , year=

  43. [49]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    VAEBM: A symbiosis between variational autoencoders and energy-based models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  44. [50]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    DDGAN: Denoising diffusion GANs , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  45. [51]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    MEG: Masked Energy-based Generative Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  46. [52]

    Communications of the ACM , volume=

    Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

    2020

  47. [53]

    International Conference on Learning Representations (ICLR) , year =

    Takeru Miyato and Toshiki Kataoka and Masanori Koyama and Yuichi Yoshida , title =. International Conference on Learning Representations (ICLR) , year =

  48. [54]

    Dai and Shakir Mohamed and Ian Goodfellow , title =

    William Fedus and Mihaela Rosca and Balaji Lakshminarayanan and Andrew M. Dai and Shakir Mohamed and Ian Goodfellow , title =. International Conference on Learning Representations (ICLR) , year =

  49. [55]

    Diffusion- GAN: Training gans with diffusion

    Diffusion-GAN: Training GANs with Diffusion , author =. arXiv preprint arXiv:2206.02262 , year =

    work page arXiv

  50. [56]

    Towards Principled Methods for Training Generative Adversarial Networks , booktitle =

    Mart. Towards Principled Methods for Training Generative Adversarial Networks , booktitle =. 2017 , publisher =

    2017

  51. [57]

    Amortised

    Casper Kaae S. Amortised. 5th International Conference on Learning Representations (ICLR) , year =

  52. [58]

    5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings , year=

    Towards Principled Methods for Training Generative Adversarial Networks , author=. 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings , year=

    2017

  53. [59]

    arXiv preprint arXiv:1610.04490 , year=

    Amortised MAP Inference for Image Super-resolution , author=. arXiv preprint arXiv:1610.04490 , year=

    work page arXiv

  54. [60]

    Proceedings of the 35th International Conference on Machine Learning , volume=

    Which Training Methods for GANs do actually Converge? , author=. Proceedings of the 35th International Conference on Machine Learning , volume=. 2018 , url=

    2018

  55. [61]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  56. [62]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  57. [63]

    M. J. Kearns , title =

  58. [64]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  59. [65]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  60. [66]

    Suppressed for Anonymity , author=

  61. [67]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  62. [68]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  63. [69]

    Divgan: A diversity enforcing generative adversarial network for mode collapse reduction

    Allahyani, M., Alsulami, R., Alwafi, T., Alafif, T., Ammar, H., Sabban, S., and Chen, X. Divgan: A diversity enforcing generative adversarial network for mode collapse reduction. Artificial Intelligence, 317: 0 103863, 2023

    2023

  64. [70]

    Wasserstein gan, 2017

    Arjovsky, M., Chintala, S., and Bottou, L. Wasserstein gan, 2017

    2017

  65. [71]

    Seeing what a gan cannot generate

    Bau, D., Zhu, J.-Y., Wulff, J., Peebles, W., Strobelt, H., Zhou, B., and Torralba, A. Seeing what a gan cannot generate. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 4502--4511, 2019

    2019

  66. [72]

    Infogan: Interpretable representation learning by information maximizing generative adversarial nets

    Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., and Abbeel, P. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016

    2016

  67. [73]

    arXiv preprint arXiv:1605.09782 , year=

    Donahue, J., Kr \"a henb \"u hl, P., and Darrell, T. Adversarial feature learning. arXiv preprint arXiv:1605.09782, 2016

    work page arXiv 2016

  68. [74]

    Testing generated distributions in gans to penalize mode collapse

    Gong, Y., Xie, Z., Xie, M., and Ma, X. Testing generated distributions in gans to penalize mode collapse. In International Conference on Artificial Intelligence and Statistics, pp.\ 442--450. PMLR, 2024

    2024

  69. [75]

    Generative adversarial networks

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial networks. Communications of the ACM, 63 0 (11): 0 139--144, 2020

    2020

  70. [76]

    The gan is dead; long live the gan! a modern gan baseline

    Huang, N., Gokaslan, A., Kuleshov, V., and Tompkin, J. The gan is dead; long live the gan! a modern gan baseline. Advances in Neural Information Processing Systems, 37: 0 44177--44215, 2024

    2024

  71. [77]

    Scaling up gans for text-to-image synthesis

    Kang, M., Zhu, J.-Y., Zhang, R., Park, J., Shechtman, E., Paris, S., and Park, T. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10124--10134, 2023

    2023

  72. [78]

    Training generative adversarial networks with limited data

    Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., and Aila, T. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020 a

    2020

  73. [79]

    Analyzing and improving the image quality of stylegan

    Karras, T., Laine, S., and Aila, T. Analyzing and improving the image quality of stylegan. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020 b

    2020

  74. [80]

    Learning multiple layers of features from tiny images

    Krizhevsky, A., Hinton, G., et al. Learning multiple layers of features from tiny images. Thesis, 2009

    2009

  75. [81]

    Improved precision and recall metric for assessing generative models

    Kynk \"a \"a nniemi, T., Karras, T., Laine, S., Lehtinen, J., and Aila, T. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019

    2019

  76. [82]

    Mode seeking generative adversarial networks for diverse image synthesis

    Mao, Q., Lee, H.-Y., Tseng, H.-Y., Ma, S., and Yang, M.-H. Mode seeking generative adversarial networks for diverse image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  77. [83]

    Which training methods for gans do actually converge? In International conference on machine learning, pp.\ 3481--3490

    Mescheder, L., Geiger, A., and Nowozin, S. Which training methods for gans do actually converge? In International conference on machine learning, pp.\ 3481--3490. PMLR, 2018 a

    2018

  78. [84]

    Mescheder

    Mescheder, L., Geiger, A., and Nowozin, S. Which training methods for gans do actually converge? In Proceedings of the 35th International Conference on Machine Learning, volume 80, pp.\ 3481--3490, 2018 b . URL https://arxiv.org/abs/1801.04406

    work page arXiv 2018

  79. [85]

    Spectral Normalization for Generative Adversarial Networks

    Miyato, T., Kataoka, T., Koyama, M., and Yoshida, Y. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018

    work page Pith review arXiv 2018

  80. [86]

    F., Oh, S

    Naeem, M. F., Oh, S. J., Uh, Y., Choi, Y., and Yoo, J. Reliable fidelity and diversity metrics for generative models. In International conference on machine learning, pp.\ 7176--7185. PMLR, 2020

    2020

Showing first 80 references.