Weather Attribute-Aware Multi-Scale Image Generation with Residual Learning

We present image generation networks to generate images conforming to specified weather attributes. Taking weather attributes as the conditions, the proposed networks generate scene images with the help of a guided reference image. To generate higher-resolution images, we construct a multi-scale generation framework consisting of a global generator and a local enhancer. Furthermore, we integrate the idea of residual learning into the proposed framework, and aim at generating fine-grained texture. The evaluation shows performance comparison both from quantitative and qualitative perspectives. A comprehensive study including the impact of different attributes and extension of the proposed models is also provided. This work is kind of hybrid approach among various image generation studies.