Relightable Neural Human Assets from Multi-view Gradient Illuminations

Figure 1. We present UltraStage, a new dataset containing more than 2000 human assets captured under multi-view and multi-illumination settings. The high-quality images allow us to extract detailed normal, albedo, and material maps, as well as reconstruct fine geometry (left). We further propose a neural processing pipeline to interpret each capture into a neural human asset, which enables various applications like photo-realistic relighting (middle) and exquisite novel view synthesis (right). Our assets faithfully model human details, e.g., the delicate cloth wrinkles or the vivid classical fan textures.