Few-shot Image Generation via Non-Local Feature Constraints

Few-shot image generation aims to generate high-quality and highly diverse images for domains with limited training data. However, a small number of training samples often leads to overfitting of the model. Moreover, existing methods do not fully utilize the specific information of training images, resulting in a loss of generative diversity in the generated images. To address this issue, we propose a few-shot image generation method based on non-local feature constraints. This method employs non-local feature extraction to extract the style features of target-domain images, fully utilizing the style information of target-domain samples to guide the style transformation of generated images. To ensure that the non-local feature module extracts the common style features of real images, we propose the style consistency loss to constrain feature consistency. Extensive quantitative and qualitative experiments demonstrate the effectiveness of our method.