• Corpus ID: 246996534

When, Why, and Which Pretrained GANs Are Useful?

@article{Grigoryev2022WhenWA,
  title={When, Why, and Which Pretrained GANs Are Useful?},
  author={Timofey Grigoryev and Andrey Voynov and Artem Babenko},
  journal={ArXiv},
  year={2022},
  volume={abs/2202.08937}
}
The literature has proposed several methods to finetune pretrained GANs on new datasets, which typically results in higher performance compared to training from scratch, especially in the limited-data regime. However, despite the apparent empirical benefits of GAN pretraining, its inner mechanisms were not analyzed indepth, and understanding of its role is not entirely clear. Moreover, the essential practical details, e.g., selecting a proper pretrained GAN checkpoint, currently do not have… 
FakeCLR: Exploring Contrastive Learning for Solving Latent Discontinuity in Data-Efficient GANs
TLDR
This paper proposes FakeCLR, which only applies contrastive learning on perturbed fake samples, and devises three related training techniques: Noise-related Latent Augmentation, Diversity-aware Queue, and Forgetting Factor of Queue.
A Comprehensive Survey on Data-Efficient GANs in Image Generation
TLDR
This paper revisits and analyzes DE-GANs from the perspective of distribution optimization, and proposes a taxonomy, which classifies the existing methods into three categories: Data Selection, GANs Optimization, and Knowledge Sharing.
StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets
TLDR
The final model, StyleGAN-XL, sets a new state-of-the-art on large-scale image synthesis and is the first to generate images at a resolution of 10242 at such a dataset scale.

References

SHOWING 1-10 OF 31 REFERENCES
On Leveraging Pretrained GANs for Limited-Data Generation
TLDR
It is revealed that low-level filters of both the generator and discriminator of pretrained GANs can be transferred to facilitate generation in a perceptually-distinct target domain with limited training data.
MineGAN: Effective Knowledge Transfer From GANs to Target Domains With Few Images
TLDR
This work proposes a novel knowledge transfer method for generative models based on mining the knowledge that is most beneficial to a specific target domain, either from a single or multiple pretrained GANs, and shows that it effectively transfers knowledge to domains with few target images, outperforming existing methods.
Top-k Training of GANs: Improving GAN Performance by Throwing Away Bad Samples
TLDR
A simple modification to the Generative Adversarial Network (GAN) training algorithm that materially improves results with no increase in computational cost is introduced: when updating the generator parameters, the gradient contributions from the elements of the batch that the critic scores as `least realistic' are zeroed out.
Transferring GANs: generating images from limited data
TLDR
The results show that using knowledge from pretrained networks can shorten the convergence time and can significantly improve the quality of the generated images, especially when the target data is limited and it is suggested that density may be more important than diversity.
Freeze Discriminator: A Simple Baseline for Fine-tuning GANs
TLDR
It is shown that simple fine-tuning of GANs with frozen lower layers of the discriminator performs surprisingly well, and a simple baseline, FreezeD, significantly outperforms previous techniques used in both unconditional and conditional GAns.
Image Generation From Small Datasets via Batch Statistics Adaptation
TLDR
This work proposes a new method for transferring prior knowledge of the pre-trained generator, which is trained with a large dataset, to a small dataset in a different domain, and can generate higher quality images compared to previous methods without collapsing.
Few-shot Image Generation with Elastic Weight Consolidation
TLDR
This work adapts a pretrained model, without introducing any additional parameters, to the few examples of the target domain, in order to best preserve the information of the source dataset, while fitting the target.
Demystifying MMD GANs
TLDR
The situation with bias in GAN loss functions raised by recent work is clarified, and it is shown that gradient estimators used in the optimization process for both MMD GANs and Wasserstein GAns are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters.
The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
TLDR
A new dataset of human perceptual similarity judgments is introduced and it is found that deep features outperform all previous metrics by large margins on this dataset, and suggests that perceptual similarity is an emergent property shared across deep visual representations.
Unsupervised Discovery of Interpretable Directions in the GAN Latent Space
TLDR
This paper introduces an unsupervised method to identify interpretable directions in the latent space of a pretrained GAN model by a simple model-agnostic procedure, and finds directions corresponding to sensible semantic manipulations without any form of (self-)supervision.
...
...