Generic and Privacy-free Synthetic Data Generation for Pretraining GANs

Kyungjune Baek, Hyunjung Shim

Keywords: GANs, low-shot, pretraining, synthetic dataset

TL;DR: Proposing ways to generate synthetic datasets mimicking the natural images and to utilize them as a pretraining dataset for GANs under low-shot learning manner.

Abstract:

Transfer learning for GANs successfully improves low-shot generation performance. However, existing studies show that the pretrained model using a single benchmark dataset is not generalized to various datasets. More importantly, the pretrained model can be vulnerable to copyright or privacy risks. To resolve both issues, we propose an effective and unbiased data synthesizer, namely Primitives-PS, inspired by the generic characteristics of natural images.
Since Primitives-PS only considers the generic properties of natural images, the images are free from copyright and privacy issues. In addition, the single model pretrained on our dataset can be transferred to various target datasets. Extensive analysis demonstrates that each component of our data synthesizer is effective, and provides insights on the desirable nature of the pretrained model for the transferability of GANs. For better reproducibility and implementation details we provide the source code at https://github.com/FriedRonaldo/Primitives-PS.