Dynamic Contextual Attention Networks for Improved Single Image Generation

Abstract

Recently, single generative models that generate various samples from a single image without training on large datasets are of great interest. However, training the model to generate realistic images from a single sample remains a challenging task. Previous research in the field of single-image generation has also exhibited sensitivity to fine-grained image features and encountered limitations in terms of the quality of the generated images. To overcome the limitations, we introduce an attention module and block architecture that leverages latent information within a single image. The proposed method aims to enhance the capabilities of a single image generation model to generate a diverse range of data samples based on a single input image. Our proposed model shows innovative results in a variety of tasks and performs particularly well in image generation and outpainting. Experimental results show that our proposed model provides better image quality compared to the traditional single image generation, which represents an important advance for future image processing applications.