FreGAN: Exploiting Frequency Components for Training GANs under Limited Data

Authors: mengping yang, Zhe Wang, Ziqiu Chi, Yanbing Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments
Researcher Affiliation Academia Department of Computer Science and Engineering, ECUST, China Key Laboratory of Smart Manufacturing in Energy Chemical Process, ECUST, China wangzhe@ecust.edu.cn mengpingyang@mail.ecust.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our codes are available at https://github.com/kobeshegu/Fre GAN_Neur IPS2022.
Open Datasets Yes We test the effectiveness of our method on low-shot datasets from various domains with different resolutions. On 256 256 resolution, we use Animal Face Dogs and Cat [33], as well as 100-shot-Panda, Obama, and Grumpy_cat [51]. On 512 512 resolution, we use Anime-Face, Art Paintings, Moongate, Flat-colored, and Fauvism-still-life [24]. On 1024 1024 resolution, we use Pokemon, Skulls, Shells, Met Face [16] and Breca HAD [1].
Dataset Splits No The paper does not provide specific details about validation dataset splits (percentages, sample counts, or explicit standard split references).
Hardware Specification No The paper states that implementation details in Section 4 provide information on the total amount of compute and type of resources, but Section 4 and its subsections do not explicitly mention specific hardware details such as GPU/CPU models, processor types, or memory amounts used for experiments.
Software Dependencies No The paper does not explicitly list specific software dependencies with their version numbers, such as programming languages, libraries, or solvers.
Experiment Setup Yes We choose the current state-of-the-art few-shot generative model Fast GAN [24] as the backbone and implement our proposed techniques upon it. all other settings remain the same as [24]. We decompose the intermediate 8 8, 16 16, 32 32 features of G and D into frequency components for our frequency skip connection, high-frequency discriminator, and high-frequency alignment.