# NVIDIA’s New AI: Generating 3D Models! ## Метаданные - **Канал:** Two Minute Papers - **YouTube:** https://www.youtube.com/watch?v=shy51E-MU8Y - **Дата:** 03.12.2022 - **Длительность:** 7:09 - **Просмотры:** 373,557 - **Источник:** https://ekstraktznaniy.ru/video/13369 ## Описание ❤️ Check out Runway and try it for free here: https://runwayml.com/papers/ Use the code TWOMINUTE at checkout to get 10% off! 📝 The paper "GET3D: A Generative Model of High Quality 3D Textured Shapes Learned from Images" is available here: https://nv-tlabs.github.io/GET3D/ 📝 Our paper "Gaussian Material Synthesis" is available here: https://users.cg.tuwien.ac.at/zsolnai/gfx/gaussian-material-synthesis/ Andrew Price’s legendary donut tutorial: https://www.youtube.com/watch?v=nIoXOplUvAw&list=PLjEaoINr3zgFX8ZsChQVQsuDSjEqdWMAD&index=1 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Edward Unthank, Eric Martel, Geronimo Moralez, Gordon Child, Jace O'Brien, Jack Lukic, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Kyle Davis, Lorin Atzberger, Lukas Biewald, Luke Dominique Warner, Matthew Allen Fisher, Matthew Valle, M ## Транскрипт ### Intro [] Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Today we are going to see how this new AI can  create entire virtual worlds out of thin air,   and to do all that, we don’t even need to  be a professional artist. This can do so   much! But how much exactly? Well, let’s give  it a hard time and see what it can do through   4 really cool examples, and then, it gets  even better, I’ll tell you in a moment why! ### 1 - Great meshes [0:31] One, as you see here, it can generate a  great variety of different object types,   cars, animals, chairs, you name it. And,  would you look at that! I am very happy,   because I am already seeing two really nice  features here. Do you see them too? One is   that these are textured objects, most previous  techniques just give us the geometry, with this   one, we get nicely textured objects, that is super  nice, but, look! These are also regular 3D meshes,   which means a data structure that can be used  by most 3D modeling systems right away as it is.    That is not true for all previous techniques,  for instance, some of them even give us point   clouds instead. And textures aside, if you  compare the quality of the results to these   previous methods, there is a night and day  difference. This is so much better. So cool! ### 2 - Variant generation [1:32] Now, two, the fact that it can generate a  great variety of objects is all well and good,   but what if we kind of like this object, find it  to be almost exactly what we like, but we would   wish to change it a little. Can we do that? Well, that’s tough. For that, we would need   a latent space interpolation technique that is  akin to this earlier font generation method,   or like this material generator from our  earlier paper where we can walk around on   this 2D plane and fine-tune a material that we  already like. Doing the same with these objects   would be fantastic, but let’s not kid ourselves.   That is significantly harder in this context. So,   can it do it? And the answer is, look at that!   Yes it can! And what is even more impressive   is that nearly all of the intermediate results  make sense. That is the hardest part. Loving it! And, if it stopped there, that would  already be amazing, but it doesn’t stop   there. There is so much more! Three, get  this, it even understands the difference ### 3 - Textures vs geometry [2:40] between geometry and texture. This means  that whenever we modify the geometry,   the lamps still have the appropriate texture  and colors, it also understands that as we   create an SUV from a smaller car, the doors  and windows now have different placements.    That is an excellent showcase of an AI technique  that actually understands what it is generating. ### 4 - Text-guided generation [3:11] And now, here comes my favorite, they also  promise text-guided shape generation. Does   this ring a bell? Oh yes, we can prompt this  technique like we can prompt a text to image AI   like DALL-E 2. That would be amazing, but, have a  look. This is a previous technique trying to give ### VS previous technique [3:28] different animals tiger stripes, or actually, I am  not even sure what this is trying to do, or, make   these dogs a bit fluffier. Those are noble goals,  but unfortunately, this previous technique could   not pull this off. Not even close. And now, hold  on to your papers and let’s have a look at the new   technique’s results. Oh my goodness! There is no  contest here. Tiger stripes, fluffy dogs, you name   it. With this new method, we can create a tiger  or a panda out of any other animal, or ask for a   brick house, or burn it down by just writing these  two words and pressing enter. This is incredible   - we just write the text, the AI does all the hard  work, and we take credit for it. How cool is that? And wait a second, we noted that this was the  best of what a previous method could do. How   long ago is this paper from? And this is  where I fell off the chair when reading   this new paper. This previous method is  not even a year old. It was published   approximately 11 months ago. And this kind  of progress in just a year? I am stunned. And wait, I promised to tell you how it gets even  better. Well, it does all these amazing things, ### It gets even better! [4:50] and not only that, but normally,  with previous handcrafted techniques,   if we wanted to do these 4 things, we would have  needed 4 different algorithms for that. Years ago,   if we wanted to conquer Chess and Starcraft and  Dota 2, these are three different applications,   and thus, need 3 different handcrafted  algorithms to solve. However, this is   not the case here. Because this new technique  can do all of them at the same time. This is   not 4 tasks, and 4 techniques, this is just one  technique. That is super important. Why? Well,   there are many definitions for artificial  intelligence, but most of them contain   some sort or generality. A good AI technique  should be able to learn from a large dataset,   and obtain general knowledge from it. And  this is an excellent testament of that. And, just imagine what this technique will be able   to do just a couple more papers down  the line. What a time to be alive! Thanks for watching and for your generous  support, and I'll see you next time!