1: \begin{abstract}

2:  Modern GANs excel at generating high quality and diverse images. However, when transferring the pretrained GANs on small target data (\eg, 10-shot), the generator tends to replicate the training samples. 

3:  Several methods have been proposed to address this few-shot image generation task, but there is a lack of effort to analyze them under a unified framework. 

4:  As our first contribution, we propose a framework 

5:  to analyze existing methods during the adaptation.

6: %

7:  Our analysis 

8:  discovers that while some methods

9:  have disproportionate focus on diversity preserving which impede quality improvement, all methods achieve similar quality after convergence.

10:  Therefore, the better methods are those that can slow down diversity degradation.

11:  Furthermore, 

12:  our analysis reveals that there is still plenty of room to further slow down diversity degradation.

13: 

14:  Informed by our analysis and to 

15:  slow down the diversity degradation of the target generator during adaptation, our second contribution proposes to apply mutual information (MI) maximization to retain the source domain's 

16:  rich multi-level diversity information in the target domain generator.

17:  We propose to perform MI maximization by contrastive loss (CL), leverage the generator and discriminator as two feature encoders to extract different multi-level features for computing CL.

18:  We refer to our method as Dual Contrastive Learning (DCL).

19:  Extensive experiments on several public datasets show that, while leading to a slower diversity-degrading generator during adaptation, our proposed DCL brings visually pleasant quality and state-of-the-art quantitative performance.

20:  {Project Page:} {\color{RubineRed}\href{https://yunqing-me.github.io/A-Closer-Look-at-FSIG}{yunqing-me.github.io/A-Closer-Look-at-FSIG}}.

21: \end{abstract}

22: