CNN-Based Local Tone Mapping in The Perceptual Quantization Domain

This paper introduces VG-GAN, a novel conditional GAN for graphical design generation tasks with applications in background design, layout and scene generation. VG-GAN utilizes vector-based methods to achieve scale invariance of the generated layouts. Concretely, the GAN model outputs only relevant vector layout information, and the final layout image is rendered in a post-processing step to allow layouts to be scaled arbitrarily. in contrast to existing vector-based generation models that require a choice of initial class relationships, VG-GAN proposes a selection module to automatically learn the class relationships in target applications, presenting a novel application of vector-based generation not addressed by existing literature. VG-GAN is applied on three generation tasks given optional design conditions, namely, banner background generation, document layout generation and clipart scene generation. The results demonstrate the model's effectiveness in learning and generating conditional graphical design.