Dr. Károly Zsolnai-Fehér reviews a new cloth simulation that combines fast mesh-based physics with yarn-level detail in real time. The hybrid method runs mesh on CPU and yarn detail on GPU, handling garments with up to 42 million vertices in milliseconds. It captures holey geometry and tightening yarns that simpler mesh-only approaches miss.
Questions this video answers
December 11, 2021.
Dr. Károly Zsolnai-Fehér.
Georg Spell (a PhD student).
61 views.
The mesh (coarse) simulation runs on the processor (CPU); the yarn-level part runs on the graphics card (GPU).
Chapters
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. This is not some hot new Two Minute Papers merchandise. This is what a new class simulation paper is capable of doing today. So good. And today we are also going to see how to put the holes back into holy geometry and do it quickly. But wait a second. That is impossible. We know that yarn based cloth simulations take forever to compute. Here is an example. We showcased this previous work approximately 60 episodes ago. And curiously, get this. Some of the simulations here were nearly 60 times faster than the yarn based reference simulation. That is amazing. However, we noted that even though this technique was a great leap forward, of course it wasn't perfect. There was a price to be paid for this amazing speed, which was. Look. The pulling effects on individual yarns were neglected. That amazing holey geometry is lost. I noted that. I'd love to get that back. You will see in a moment that maybe my wish comes true today. Hmm. So, quick rundown. We either go for a mesh based cloth simulation. These are really fast, but no holy geometry. Or we choose the yarn level simulations. These are the real deal. However, they take forever, unfortunately. Still, there seems to be no way out here. Now let's have a look at this new technique. It promises to try to marry these two concepts. Or in other words, use a fast mesh simulator and add the yarn level cloth details to on top of it. Now, of course, that is easier said than done. So let's see how this new method can deal with this challenging problem. So this is the fast mesh simulation to start with. And now hold on to your papers and let's add those yarns. Oh my. That is beautiful. I absolutely love how at different points you can even see through these garments. Beautiful. So how does it do all this magic? Well, look at this naive method. This neglects the proper tension between the yarns. And look at how beautifully they tighten as we add the new method on top of it. This technique can do it properly. And not only that, but it also simulates how the garment interacts with the virtual body in a simulated world. Once again, note how the previous naive method neglects the tightening of the yarns. So I am super excited. Now let's see how long we have to wait for this. Are we talking fast mesh simulation timings or slow yarn simulation timings? My goodness, look at that. The mesh part runs on your processor while the yarn part of the simulation is implemented on the graphics card. And whoa. The whole thing runs in the order of milliseconds easily in real time. Even if we have a super detailed garment with tens of millions of vertices. In the simulation, the timings are in the order of tens of milliseconds at worst. That's also super fast. Wow. So yes, part of this runs on our graphics card and in real time. Here we have nearly a million vertices. Not a problem here. Nearly 2 million vertices. 2 million points if you will. And it runs like a dream. And it only starts to slow down when we go all the way to a super detailed piece of garment with 42 million vertices. And here there is so much detail that finally, not the simulation technique is the bottleneck, but the light simulation algorithm is. Look, there are so many high frequency details that we would need higher resolution videos or more advanced anti aliasing techniques to resolve all this detail. All this means that the simulation technique did its job really well. So finally, yarn level simulations in real time. What a time to be alive. Huge congratulations to the entire team and the first author, Georg Spell, who is still a PhD student, making important contributions like this. And get this, Georg's short presentation was seen by. Oh my 61 people. Views are not everything, not even close. But once again, if we don't talk about this work here, I am worried that almost no one will. And this is why two minute papers exist. Subscribe if you wish to see more of these miracle papers, we have some great ones coming up. This video has been supported by Weights and Biases. They have an amazing podcast by the name Gradient Descent where they interview machine learning experts who discuss how they use learning based algorithms to solve real world problems. They've discussed biology, teaching robots, machine learning in outer space, and a whole lot more. Perfect for a fellow scholar with an open mind. Make sure to visit them through WNB MEGD or just click the link in the video description. Our thanks to WeightsAndBiases for their long standing support and for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.
