Real-time Rendering of Data-driven Fluid Simulations with Style Transfer

For my Master Thesis, I was working on the problem of real-time fluid rendering and style transfer. The goal was to improve the rendering methods for a machine-learning based physics simulation technology that was developed by my supervisors. The developed methods allow rendering of water, foam, granular materials and smoke.

In the end, the resulting paper was sent to SIGGRAPH 2017 Real-time Live! and its submission can be seen in the video above.

In addition to that, I've experimented with real-time style transfer methods that could be used for procedural texturing, stylistic level design and ethereal sequences in video games. At the time of development, real-time style transfer was not yet an explored topic, and we have barely succeeded using latest generation GPUs in conjunction with fast stylization using neural networks.

Some of the produced stylized captures are featured in the videos below.

References & Further Reading

  1. N. Bartolovic. Real-time Rendering of Data-driven Fluid Simulations with Style Transfer. My Master Thesis. 2017.
  2. L. Ladicky, S. Jeong, N. Bartolovic, M. Pollefeys and M. Gross. Physicsforests: real-time fluid simulation using machine learning. ACM SIGGRAPH 2017 Real Time Live!. ACM, 2017.