Spatial Frequency Bias in Convolutional Generative Adversarial Networks

Authors: Mahyar Khayatkhoei, Ahmed Elgammal7152-7159

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Figure 3 shows FID Levels of GANs trained on two 128 128 image datasets: Celeb A (Liu et al. 2015) and LSUN-Bedrooms (Yu et al. 2015). The GANs exhibit an increase in FID Levels on both datasets, similar to the behavior observed in Figure 2 (middle), suggesting that the learnt high frequencies contain more mismatch than the low frequencies.
Researcher Affiliation Academia Department of Computer Science, Rutgers University New Brunswick, New Jersey {m.khayatkhoei, elgammal}@cs.rutgers.edu
Pseudocode No The paper describes mathematical formulations and processes but does not include any formal pseudocode blocks or algorithms.
Open Source Code No The paper does not provide a direct link to a code repository or explicitly state that the source code for its methodology is made publicly available.
Open Datasets Yes Figure 3 shows FID Levels of GANs trained on two 128 128 image datasets: Celeb A (Liu et al. 2015) and LSUN-Bedrooms (Yu et al. 2015).
Dataset Splits No The paper discusses evaluating GANs on datasets and comparing FID levels, but it does not specify explicit training, validation, and test dataset splits with percentages or sample counts for reproducibility of model training.
Hardware Specification No The paper does not provide any specific details about the hardware used to conduct the experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., programming languages, libraries, or frameworks like Python, PyTorch, TensorFlow) that were used in the experiments.
Experiment Setup Yes We use three popular convolutional GAN models in our studies: WGAN-GP (Gulrajani et al. 2017) serves as a simple but fundamental GAN model; and Progressively Growing GAN (PG-GAN) (Karras et al. 2018) and Style GAN2 (Karras et al. 2020) serve as stateof-the-art models with large capacity and complex structure, incorporating state-of-the-art normalization and regularization techniques. Since our goal is to compare the performance of GANs on high versus low spatial frequencies, and not to compare the overall quality of the generated samples with one another or the state-of-the-art, we chose to use PGGAN and Style GAN2 with a slightly smaller capacity in our training (corresponding to the capacity used in Section 6.1 of Karras et al. (2018) for ablation studies). See Appendix for the details of each model.