“Application of Generative Adversarial Networks in Color Art Image Shadow Generation”, Hsin-Cheng Wu, Wei-Ru Lin, Shao-Kuo Tai2023-11-09 (Danbooru AI, GAN)⁠:

In this research, we propose a framework based on Generative Adversarial Networks (GANs), known as the Color Shading Frame (CSF), to address the challenge of achieving ideal shadow effects in artistic creations.

The CSF framework consists of two main components: Line Art Extraction and Shadow Composition. Line Art Extraction involves extracting line drawings from color images, while Shadow Composition aims to combine shadows with color images. Through these steps, the CSF framework enables the automatic generation of shadows with directional lighting effects.

Experimental results demonstrate that, when using neural networks as the line art extraction method, CSF outperforms traditional edge detection methods in handling noisy images with patterns resembling paper textures and images with gradients.

…For our experimentation in shadow generation, we curated a dataset from Danbooru2021, specifically filtering images labeled with “Flat_Color”, which indicate color images without shadows or variations due to lighting. We excluded images containing adult content, resulting in a final dataset of 2,692 images from a pool of 4 million images. These selected images can be categorized into various classes, including those with noise, decorative patterns, solid color blocks without contours, and gradients. We will employ 3 methods, namely Canny, SketchKeras line art extraction, and the existing color image shadow generation method, PaintingLight, to compare and analyze the outcomes on this dataset.