FluidsFormer

Abstract

In this work, we introduce FluidsFormer, a Transformer-based approach for fluid interpolation and edition within a continuous-time framework. By combining the capabilities of a Physics-Informed Transformer Token (PITT) architecture and a residual neural network (RNN), we analytically predict the physical properties of the fluid state. Our network architecture enables us to interpolate substep frames between simulated keyframes, enhancing the temporal smoothness and sharpness of physics-based animations. We demonstrate promising results for smoke interpolation and conduct initial experiments on liquids.