PhysicsGen - Can Generative Models Learn from Images to Predict Complex Physical Relations?

This dataset comprises 300,000 pairs of images designed for the advancement of generative model applications in physical simulations. Each pair consists of an input image and its corresponding output image that represents a physical simulation. The dataset aims to facilitate research into whether generative models can effectively learn and reproduce complex physical dynamics from visual data, potentially replacing traditional differential equation-based methods with significant computational speedups. Data, baseline models and evaluation code: https://www.physics-gen.org