GGX-GAN: a generative adversarial network for single-image material appearance editing with physical parameters

Material appearance editing methods based on generative adversarial network (or GAN) use perceptual parameters to control material appearance but are limited by the rating quality of training data. The data annotators were often confused by ambiguous perceptual parameters, leading to poor labeling or rating results. Instead, we choose physical parameters to achieve more interpretable and measurable material appearance editing. The image dataset rendered by a physically based renderer were used for GAN training and model validation, and the GGX BRDF model was chosen to control the glossiness of rendered objects. Then we embedded the BRDF model into the latent space of GAN and establishes a physical parameter control space of editing, allowing continuous material appearance tuning. Inverse rendering results of the images generated by the proposed GAN were analyzed to illustrate and evaluate its editing performance. It is shown that the proposed GAN can provide accurate, intuitive material appearance editing on both real-world photographs and rendered images.