Ray tracing is getting faster and cleaner. A new denoising method preserves sharp reflections while cutting noise, displacement mapping achieves 11× speedups with a third of the memory, and a real-time demo runs 2 million rays per second in under 20 MB.
Questions this video answers
More than 2 million rays per second.
Less than 20 megabytes of memory.
Up to 11 times faster.
It required three times less memory than previous techniques.
A SIGGRAPH Asia paper (single-author, named Shinji in the narration).
Chapters
Feast your eyes upon these light transport simulation algorithms. Ray tracers if you will. These enable us to simulate the path of millions of light rays bouncing in a scene and create absolutely beautiful photorealistic images. Hmm. But there is a huge problem. What is the problem? Dear fellow scholars, this is Two Minute Papers with Dr. Caro Zsolnai. Now look at the previous technique. Here is the problem. Noise. My goodness, so much noise. This would take quite a bit of more time to get rid of. And now have a look at the new technique. Oh, these images are so much cleaner. Are they usable? Well, we have noise filtering algorithms for light transport simulations. Are they any good? Well, let's have a look together. Let's plug in the previous technique into a noise filtering algorithm and the input has these walking light blobs. Are they gone now? Hmm, unfortunately not. This is supposed to remove noise and not fix these blobs. So let's see what happens with the new technique. And oh my God, I already see that we have less noise. And after filtering. Bingo. Love it. Not perfect, but so much better. So let's have a closer look at what just happened. Here is a comparison with path tracing. This is one of the first algorithms you learn on your light transport class. It can barely do anything on this notoriously difficult scene. MMIs. This is marginal multiple importance sampling. Now this is so much better. However, with this we also get lots of noise. This is the restir technique that you just saw with the blobs. This is from 2022 and gives us less noise. That is good. However, look, the specular highlights on the teapot are different. They do not match the true simulation. It does not give us the image we are looking for. And now look at the new technique. Oh yes, less noise. And it is correct at the same time. This was a particularly good case for the new technique. But even on a selection of other scenes it performs really well. Better than its predecessors. A solid step forward. Loving it. And we are not done yet. Not even close. I have a bunch of incredible new papers on ray tracing for you to marvel at. When creating games and virtual worlds, we often find ourselves having 3D geometry that is not too detailed, but we would like to make it look a bit more interesting. Slapping a texture on it kinda works, but it is still very flat and not very exciting. Now we would like this geometry to stay simple and quick to render, easy to store, but at the same time to appear to be a lot more detailed. Look so much better. This technique is called displacement mapping and it can make a really simple piece of geometry come alive in your games and virtual worlds. However, it is not new. Not new at all. It goes back 30 years, possibly even longer. So, what is interesting here in this new paper? Well, two things. One, hold on to your papers, fellow scholars, because it can do all this up to 11 times faster than previous techniques. Hmm. Now is the point where I immediately think that potentially by leaning more heavily on memory, so maybe faster, but at the cost of using more memory. However, I was very surprised to find out that this was not only not the case, but it also requires three times less memory. That is incredible. Two, it takes a possibly fractional level of detail parameter. That is very cool. You see, level of detail helps you save resources by rendering a coarse version of a piece of geometry and and add details to it only as we get closer. This is when we would see them, after all. However, this often happens in steps where you can often see from detail level 0 to 1 or 1 to 2. This object suddenly looks like this object. There is a pop. However, look, with a fractional level of detail parameter, we can make it pop a little less and these transitions will look more continuous. Note that other techniques can do this too, but this is one more tool in our toolkit. So cool. I could also not resist showing you this real time technique that can simulate a bit of light transport. For instance, it can do indirect illumination or color bleeding. This is the bleeding effect of light as it bounces around multiple diffuse surfaces. And. And it is a beautiful phenomenon in nature, but unfortunately less so in many virtual worlds because it is so slow and expensive to compute, but not here. This technique can pull off limited light transport, but in real time. And the implementation of it is available on Shadertoy, where you can try it right now and even look at the code or modify it. And we are still not done yet. This new technique can create this gorgeous material by shooting more than 2 million rays at it per second. And it uses less than 20 megabytes of memory. That is almost nothing for a geometry of this detail. This is a Siggraph Asia paper, which is perhaps the Olympic gold medal of computer graphics research written by a single author. Bravo, Shinji and I would also like to show you this. These gorgeous light transport simulations are done with Lux core render. This is a free, fully open source renderer with many advanced algorithms. I used to contribute to this project, wrote some code in it, and enjoyed every minute of it. It is an amazing system still used by many, many people, and they need your help. If you smart fellow scholars could contribute, that would be absolutely amazing. I have a link for you in the video description. By the way, if you wish to learn how to understand and write these amazing light simulation programs, I have a completely free master level course on it for you where we build up such a system together from absolutely nothing. It is completely free from for all of you fellow scholars. No strings attached. If you watch it, you will see the world differently. Have fun. If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best prices in the world for GPU cloud compute. No commitments or negotiation required. Just sign up and launch an instance and hold on to your papers. Because with the Lambda GPU cloud you can now get on demand H100 instances for just 199 per hour. Yes, 199. And they are one of the first cloud providers to offer publicly available on demand H100 access. Did I mention they also offer persistent storage? So join researchers at organizations like Apple, MIT and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today.
