The following work explores and compares the differences between rendering fire and smoke simulations in high resolution vs rendering these same simulations in low resolution and using deep learning based neural networks to up-scale the output via super-resolution. Several simulations are created at different levels of detail, both with and without post-processing noise added to them. The simulations are then rendered in both high and low resolutions, the lower of which is used for the super-resolution step. The results are then compared in terms of quality and time cost, to determine whether such a computationally expensive task can be improved with deep learning methods. The evaluation shows that using low resolution inputs does not create comparable results to classic high resolution renders, however using a high resolution render of a lower detail simulation creates similar results to high resolution renders of more detailed simulations.

Top