SIC-GAN: A Self-Improving Collaborative GAN for Decoding Sketch RNNs

Chi-Chun Chuang, Zheng-Xin Weng, Shan-Hung Wu

Abstract: Variational RNNs are proposed to output “creative” sequences. Ideally, a collection of sequences produced by a variational RNN should be of both high quality and high variety. However, existing decoders for variational RNNs suffer from a trade-off between quality and variety. In this paper, we seek to learn a variational RNN that decodes high-quality and high-variety sequences. We propose the Self-Improving Collaborative GAN (SIC-GAN), where there are two generators (variational RNNs) collaborating with each other to output a sequence and aiming to trick the discriminator into believing the sequence is of good quality. By deliberately weakening one generator, we can make another stronger in balancing quality and variety. We conduct experiments using the QuickDraw dataset and the results demonstrate the effectiveness of SIC-GAN empirically.

Keywords: RNNs, GANs, Variational RNNs, Sketch RNNs

Data: [Quick, Draw! Dataset](https://paperswithcode.com/dataset/quick-draw-dataset)