Regularizing Generative Adversarial Networks under Limited Data

Tseng, Hung-Yu; Jiang, Lu; Liu, Ce; Yang, Ming–Hsuan; Yang, Wei

doi:10.1109/cvpr46437.2021.00783

Cited by 105 publications

(85 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Deep generative models [12,17,29,34,41,45,46,53] have achieved lots of successes in image synthesis tasks. GAN based methods demonstrate amazing capability in yielding high-fidelity samples [4,17,27,44,53]. In contrast, likelihood-based methods, such as Variational Autoencoders (VAEs) [29,45], Diffusion Models [12,24,41] and Autoregressive Models [34,46], offer distribution coverage and hence can generate more diverse samples [41,45,46].…”

Section: Image Synthesismentioning

confidence: 99%

MaskGIT: Masked Generative Image Transformer

Chang¹,

Zhang²,

J³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Figure 1. Example generation by MaskGIT on image synthesis and manipulation tasks. We show that MaskGIT is a flexible model that can generate high-quality samples on (a) class-conditional synthesis, (b) class-conditional image manipulation, e.g. replacing selected objects in the bounding box with ones from the given classes, and (c) image extrapolation. Examples shown here have resolutions 512ˆ512, 512ˆ512, and 512ˆ2560 in the three columns, respectively. Zoom in to see the details.

show abstract

Section: Image Synthesismentioning

confidence: 99%

MaskGIT: Masked Generative Image Transformer

Chang¹,

Zhang²,

J³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…We provide comprehensive experiments (10% and 20% data) in §C of the supplementary material. We summarize our findings as: 1) the augmentation-based ADA [20] and DA [60] leak augmentation artifacts to the generator, while ADA+LCSA and DA+LCSA alleviate this issue, 2) LCSA harmonizes with ADA, DA and LeCam loss [47]. 3) we achieve the state of the art on this limited data setting.…”

Section: Results Of Image Generationmentioning

confidence: 75%

“…We build on BigGAN [5], Omni-GAN [63], MSG-StyleGAN [18], StyleGAN2 [22] but we equip the discriminator with the manifold learner which is metacontrolled to reduce overfitting that typically happens in the discriminator rather than the generator [52]. In §C of the supplementary material, we also investigate the combination of our method with DA [60], ADA [20] and LeCam-GAN [47] in limited data scenario.…”

Section: Problem Formulationmentioning

confidence: 99%

“…However, in case of DA and ADA, augmentation artifacts leak into the generated images as shown in Figure 23. The LeCamGAN [47] restricts discriminator overfitting by reducing the real/fake loss discrepancy at the output of the discriminator. However, it is not enough to limit overfitting by just operating at the output of the discriminator (otherwise a well-designed loss would deal with all overfitting issues which is not the case in the literature).…”

Section: Data-limited Generation On Cifar-10/100mentioning

confidence: 99%

“…Below, we compare our LCSA with DA [60], ADA [20] on OmniGAN (and/or BigGAN) to variants without LCSA. We conduct experiments based on BigGAN to compare our LCSA with LeCamGAN [47] due to the issue of combining LeCam loss [47] with the multi-output loss of OmniGAN. Following [60], we augmented the real images with x-flip, and trained the OmniGAN and BigGAN for 1K epochs on the full data and 5K epochs on 10%/20% data setting.…”

Section: Data-limited Generation On Cifar-10/100mentioning

confidence: 99%

See 2 more Smart Citations

Manifold Learning Benefits GANs

Ni¹,

Koniusz²,

Hartley³

et al. 2021

Preprint

View full text Add to dashboard Cite

In this paper, we improve Generative Adversarial Networks by incorporating a manifold learning step into the discriminator. We consider locality-constrained linear and subspace-based manifolds 1 , and locality-constrained nonlinear manifolds. In our design, the manifold learning and coding steps are intertwined with layers of the discriminator, with the goal of attracting intermediate feature representations onto manifolds. We adaptively balance the discrepancy between feature representations and their manifold view, which represents a trade-off between denoising on the manifold and refining the manifold. We conclude that locality-constrained non-linear manifolds have the upper hand over linear manifolds due to their non-uniform density and smoothness. We show substantial improvements over different recent state-of-the-art baselines.

show abstract

Low-Resource Adversarial Domain Adaptation for Cross-modality Nucleus Detection

Xing

Cornish

2022

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Regularizing Generative Adversarial Networks under Limited Data

Cited by 105 publications

References 24 publications

MaskGIT: Masked Generative Image Transformer

MaskGIT: Masked Generative Image Transformer

Manifold Learning Benefits GANs

Low-Resource Adversarial Domain Adaptation for Cross-modality Nucleus Detection

Contact Info

Product

Resources

About