A Two Minute Papers video (host Károly Zsolnai-Fehér) describes a new AI-driven technique for producing high-resolution smoke simulations from coarse simulations. The method learns mappings between low- and high-resolution patches using a database of simulations and two neural networks (one for density similarity, one for vorticity), applies matching high-res patches to the base simulation, and correctly handles boundary conditions. The host compares this to the older Wavelet Turbulence heuristic and praises the new results, and closes with personal remarks about the series' educational impact.
Chapters
Dear fellow scholars, this is Two Minute Papers with this work is about using AI to create super detailed smoke simulations. Typically creating a crude simulation doesn't take very long, but as we increase the resolution, the execution time and memory consumption skyrockets. In the age of AI, it only sounds logical to try to include some learning algorithms in this process. So what if we had an AI based technique that would have some sort of understanding of smoke simulations? Take our crude data and add the fine details to it. This way we could obtain a high resolution smoke simulation without waiting several days or weeks for the computation. Now, if you're a seasoned fellow scholar, you may remember an earlier work by the name Wavelet Turbulence which is one of my favorite papers of all time. So much so that it got the distinction of being showcased in the very first Two Minute Papers episode. I was a sophomore college student back then when I first seen it and was absolutely shocked by the quality of the results. That was an experience I'll never forget. It also won a Technical Oscar award and it is not an overstatement to say that this was one of the most influential works that made me realize that research is my true calling. The link to the first episode is available in the video description and if you want to see how embarrassing it is the make sure to check it out. It did something similar, but instead of using AI, it used some heuristics that describe what is the ratio and distribution of smaller and bigger vortices in a piece of fluid or smoke. Using this information it could create a somewhat similar effect, but ultimately that technique had an understanding of smoke simulations in general, but it didn't know anything about the scene that we have at hand right now. Another work that is related to this is showing a bunch of smoke simulation videos to an AI and teach it how to continue these simulations by itself. I would place this work as a middle ground solution because this work says that we should take a step back and not try to synthesize everything from scratch. Let's create a database of simulations, dice them up into tiny tiny patches, look at the same footage in low and high resolutions and learn how they relate to each other. This way we can hand the neural network some low resolution footage and it will be able to make an educated guess as to which high resolution patch should be the best match for it. When we found the right patch, we just switched the core simulation to the most fitting high resolution patch in the database. You might say that in theory creating such a Frankenstein smoke simulation sounds like a dreadful idea, but have a look at the results as they are absolutely brilliant and and as you can see, it takes a really crude base simulation and adds so many details to it. It's truly an incredible achievement. One neural network is trained to capture similarities in densities and one for vorticity. Using the two neural networks in tandem, we can take a low resolution fluid flow and synthesize the fine details on top of it in a way that is hardly believable. It also handles boundary conditions, which means that these details are correctly added even if our smoke puff hits an object. This was an issue with wavelet turbulence which had to be addressed with several follow up works. There are also comparisons against this legendary algorithm and as you can see, the new technique smokes it. However, it took nine years to do this. This is exactly nine eternities in the world of research, which is a huge testament to how powerful the original algorithm was. It is also really cool to get more and more messages where I get to know more about you fellow scholars. I was informed that the series is used in school classes in Brazil. It also used to augment college education and it is a great topic for fun family conversations over dinner. That's just absolutely fantastic. Loving the fact that the series is an inspiration for many of you. Thanks for watching and for your generous support and I'll see you next time. Ra.
