Predicting and Editing Perceptual Glossiness of Translucent Materials in a Single Image

The material editing of a single image leverages generative neural networks to modify pixel-by-pixel at the image level, allowing for the alteration of material properties in the target object. Current methods for perceptual glossiness editing of a single image are limited to opaque objects. This paper introduces a novel perceptual glossiness prediction and editing approach for translucent objects using generative adversarial networks. Specifically, we utilize StyleGAN and the pixel2style2pixel encoder to extract the latent code from material images. Building upon this latent code, we propose a method to predict and edit the perceptual glossiness of translucent materials. To establish the relationship between latent code and perceptual glossiness, we incorporate an autoencoder based on a multi-layer perceptron (MLP). Additionally, we create a dataset of translucent materials with perceptual glossiness labels for training. The StyleGAN, pixel2style2pixel series networks, and MLP are trained in two stages. Our experimental results, tested across different sample sets, demonstrate that the proposed network architecture effectively extracts perceptual information from material images, enabling accurate prediction and editing of the perceptual glossiness in single translucent material images.