Intel Just Changed Computer Graphics Forever!

Okay, this is insanity. An absolute miracle  research work. What I will say at first will   seem to make no sense, and at the end,  you’ll finally be… oh, I get it now,   this is fantastic. I hope. I’ll try my  best. So, these little dots might be the   future of computer graphics, movies and video  games altogether. Why, and what are these? These are Gaussian Splats, these can get  you a virtual copy of the real world even   with difficult thin structures in high  resolution and yes, real time. Much,   much faster than real time. My goodness, it is  so good and it is taking the world by storm. So…how? Well, Gaussian splatting represents  objects as countless tiny blobs, like shining   a flashlight through a cloud of dust. Then  it projects those blobs onto the screen,   focusing only where objects actually are,  skipping empty space. Absolute magic. Plus, it is fast. And it is fast because  it compresses so well, we can create a   good-looking scene, by not just storing a bunch  of detailed geometry, but instead just storing   a few of these Gaussians. This smaller, smooth  representation makes rendering very efficient. But it gets better. We can use it to  create a scene, but scientists at Intel,   AMD and New York University say, hold my paper for  a moment. Why not try this on not a scene, but on   images instead? This does not sound like it makes  a lot of sense, but bear with me for a moment. Earlier I wrote a funny little algorithm that  takes a bunch of triangles, and slowly adjusts   their positions and colors to rebuild the  Mona Lisa itself. So, in this project,   they take an input image of the curiosity  rover on Mars, compute the edges of this image,   computer graphics researchers learn this in  kindergarten, and then, initialize a few of   these Gaussian blobs based on that. Is this good?   Well, not quite. But, remember the triangles from   here. Okay, so now comes the magic. Look! Oh  my, that was beautiful. But what happened? They added some new blobs, and  started massaging them - moving them,   stretching them, and even repainting them  - until they nearly perfectly matched   the Curiosity rover image. It’s like a  swarm of tiny paint fairies all fixing   these bad spots until the picture  is perfect. Absolutely beautiful. We still don’t know what this is good for,  but goodness, it is beautiful. And it is fast.    Now hold on to your Papers Fellow Scholars  because here is how it trains over 15 seconds   and…wait a second. I hear you asking, Károly,  you accidentally put the reference image here   for the new technique. It is not training at  all. The previous technique from this year,   now that one is training. However, this is  not the case. The new one is training too!    It is just so fast, I have to slow it down  greatly to be able to show you the process of   the massaging. An incredible leap forward from  a technique that is also from this year. Wow. Now, wait wait. So, we started out from  an image, and now, we get an image. We just   got back what we started with. How does this  make any sense? What is all this good for? Dear   Fellow Scholars, this is Two Minute Papers  with Dr. Károly Zsolnai-Fehér. Dr. Carroll. Well, check this out. Oh my, yes! You get  back almost the same image, but in a file   that is 25 times smaller. In some other cases,  40 times smaller. That is absolutely incredible   compression. But wait, we are wise scholars here,  so we know that many previous techniques exist to   do that, in fact, JPEG compression has existed for  more than 30 years and is basically impossible to   beat. So now comes the moment of truth! Let’s see  the file sizes, 159 kilobytes for JPEG, 160 for   the new technique. Well, it’s not smaller than  a JPEG is it? But, wait a second…oh my goodness! It is not smaller, it is roughly the same size,   however, the quality of the new technique  is way, way better for the same size. This   is so much cleaner. And it just takes a  couple seconds to pull off. It is so fast   I had to slow it down for you to actually  see what is happening. Absolute magic. And   it’s not only a good algorithm, it’s also a  beautiful algorithm. I’m getting goosebumps.

This means razor-sharp, artifact-free images at  tiny file sizes, opening the door to instant,   beautiful graphics everywhere  - what a time to be alive! Of course, this is a good paper so it contains  a bunch of other comparisons too. Check them   out in the video description and of course,  I’ll throw in my genetic algorithm for the   Mona Lisa for free for all of you.   And bad news, from what I’ve seen,   absolutely nobody is talking about this paper.   It seems to me that you cannot hear about it   anywhere else. I feel like I am trying to save  some endangered species by talking about these   papers and if you would like to help this  video, leave a kind comment and subscribe,   that would help us greatly, and it would help  the algorithm show more of these to you too. I’d like to congratulate the authors and send  out a big shoutout to them. One of them is   Anton Kaplanyan, who is a dear friend and one  of my favorite people. Ridiculously talented.