University of Pennsylvania & MIT · PBR Materials

Pixie: 3D Physics from Pixels

Pixie is a neural network that maps pretrained visual features to dense material fields (Young's modulus, Poisson's ratio, density) in a single forward pass, enabling real-time physics simulations from images. It achieves 1.46-4.39x better realism than existing methods while running 1000x faster, trained on PixieVerse, a large-scale synthetic dataset with automatic VLM-based annotations.