Generative Adversarial Networks (GANs) can synthesize high-quality images by estimating the latent distribution. However, when face with few-shot image datasets, they often suffer from severe overfitting. Previous solutions have primarily focused on data augmentation, model architecture, and loss functions. This paper proposes to address the instability and overfitting issues from the perspective of the convolution process. To tackle these problems, FewConv is introduced as a plug-and-play alternative to traditional convolutions. FewConv independently learns spatial and channel information, reducing the spatial information that needs to be learned while complexifying the channel information. Specifically, FewConv calculates the variance of channel features at each layer to assess their importance and selects the significant portions for depthwise convolution. For channel information, spatial-to-channel feature transformation is performed before pointwise convolution. This makes pointwise convolution need to learn more diverse channel information. The diverse feature input of FewConv enhances its capacity to combat overfitting. Moreover, using FewConv also reduces network parameters and FLOPs, making the network more compact. To validate the effectiveness of FewConv, extensive experiments were conducted on diverse datasets. Models using FewConv achieved better FID scores and exhibited more stable training processes. FewConv is also applied to recognition training on ResNet and MobileNet, with experimental results demonstrating its effectiveness in recognition tasks.