Researchers use style transfer to upscale low-resolution smoke and fluid simulations into high-resolution, detailed versions in minutes instead of days. The technique borrows from image synthesis, breaking simulations into patches and applying textures, though keeping the results physically plausible remains tricky.
Creating absolutely stunning video footage, we can very quickly put together a core simulation and run it cheaply. However, the more turbulent motion we are trying to simulate, the more resources and time it will take if we wish to create some footage with the amount of visual quality that you see here. Well, if you think the several hour computation time for light transport algorithms was too much better, hold on to your papers because it will take not hours, but often from days to weeks to compute and to ease the computation time of such simulations. This is a technique that performs style transfer, but this time not for paintings, but for fluid and smoke simulations. How cool is that? It takes the low resolution source and detailed target footage, dices them up into small patches and borrows from image and texture synthesis techniques to create a higher resolution version of our input simulation. The challenge of this technique is that we cannot just put more swirly motion on top of our velocity fields because this piece of fluid has to obey to the laws of physics to look natural. Also, we have to make sure that there is not too much variation from patch to patch. So we have to perform some sort of smoothing on the boundaries of these patches. Our smoke plumes also have to interact with obstacles, which is anything but trivial to do. Well, have a look at the ground truth results from the high resolution simulation. This is the one that would take a long time to compute. There are clearly deviations, but given how coarse the input footage was, I'll take this any day of the week. We can now look forward to seeing even higher quality smoke and fluids in the animation movies of the near future. There was a similar technique by the name wavelet turbulence, which is one of my all time favorite papers that has been showcased in the very first two Minute Papers episode. This is what it looked like. And we are now celebrating its 10th anniversary. Imagine what a bomb this was 10 years ago. And you know what? It is still going strong. Thanks for watching and for your generous support and I'll see you next time.
