A Comprehensive Survey on Data-Efficient GANs in Image Generation

Abstract

Generative Adversarial Networks (GANs) have achieved remarkable achievements in image synthesis. These successes of GANs relies on large scale datasets, requiring too much cost. With limited training data, how to stable the training process of GANs and generate realistic images have attracted more attention. The challenges of Data-Efficient GANs (DE-GANs) mainly arise from three aspects: (i) Mismatch Between Training and Target Distributions, (ii) Overfitting of the Discriminator, and (iii) Imbalance Between Latent and Data Spaces. Although many augmentation and pre-training strategies have been proposed to alleviate these issues, there lacks a systematic survey to summarize the properties, challenges, and solutions of DE-GANs. In this paper, we revisit and define DE-GANs from the perspective of distribution optimization. We conclude and analyze the challenges of DE-GANs. Meanwhile, we propose a taxonomy, which classifies the existing methods into three categories: Data Selection, GANs Optimization, and Knowledge Sharing. Last but not the least, we attempt to highlight the current problems and the future directions.

PCE-Palm: Palm Crease Energy Based Two-Stage Realistic Pseudo-Palmprint Generation

“…With limited data, GAN-based models often face challenges with discriminator overfitting and the imbalance between discrete data spaces and continuous latent distributions (Li et al 2022;Yang and Wang 2023). These issues result in reduced fidelity and unstable training processes.…”

WeditGAN: Few-Shot Image Generation via Latent Space Relocation

“…The discriminator provides supervision to ∆w so that the relocated W + tgt aligns with the target dataset. However, there might be a gap between the dataset distribution P data and the domain distribution P domain due to the potentially biased samples in few-shot datasets (Li et al 2022). As a result, Wedit-GAN may relocate W + tgt where the distribution of the generated images P gen is closed to P data instead of the expected P domain , by unnecessarily editing in-domain attributes in the source latent space.…”

FakeCLR: Exploring Contrastive Learning for Solving Latent Discontinuity in Data-Efficient GANs

“…there are just 10 examples per artist in the Artistic-Faces dataset [11]. Therefore, Data-Efficient Generative Adversarial Networks (DE-GANs [12]) are important to many real-world practices and attract more and more attention. More analyses can be found in the survey [12].…”

“…Previous studies attribute the degradation of DE-GANs to the overfitting of the discriminator. Many data augmentation, regularization, architectures, and pre-training techniques [12] have been proposed to mitigate this issue. (i) Data Augmentation [14,26,15,27,28] is a striking method for mitigating overfitting of the discriminator and orthogonal to other ongoing researches on training, architecture, and regularization.…”