# This Free AI Generates Video FASTER Than Real Life 🤯 ## Метаданные - **Канал:** Two Minute Papers - **YouTube:** https://www.youtube.com/watch?v=gT98Kq-PV8M - **Дата:** 16.09.2025 - **Длительность:** 5:49 - **Просмотры:** 61,602 ## Описание ❤️ Check out the Fully Connected Conference by Weights & Biases - https://wandb.me/fclon2025-2min 20% discount code: FCLON2025-2MIN 📝 The paper is available here: https://github.com/Lightricks/LTX-Video 📝 My paper on simulations that look almost like reality is available for free here: https://rdcu.be/cWPfD Or this is the orig. Nature Physics link with clickable citations: https://www.nature.com/articles/s41567-022-01788-5 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Benji Rabhan, B Shang, Christian Ahlin, Gordon Child, John Le, Juan Benet, Kyle Davis, Loyal Alchemist, Lukas Biewald, Michael Tedder, Owen Skarpness, Richard Sundvall, Steef, Sven Pfiffner, Taras Bobrovytsky, Thomas Krcmar, Tybie Fitzhugh, Ueli Gallizzi If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers My research: https://cg.tuwien.ac.at/~zsolnai/ X/Twitter: https://twitter.com/twominutepapers Thumbnail design: Felícia Zsolnai-Fehér - http://felicia.hu ## Содержание ### [0:00](https://www.youtube.com/watch?v=gT98Kq-PV8M) Segment 1 (00:00 - 05:00) This AI technique can do something absolutely insane. This video transformation is already good, but you will see that it is even better than it seems. The basis of this is an imagetovideo model that is free for all of us. You specify a starting image and it continues it. And four things really caught my eye here. One, it can generate plausible motion for the duck. Amazing. The legs are a bit weird, but amazing. Waving and smiling children. This is incredibly realistic. And don't forget once again, this is a model for free for all of us. No $300 monthly subscriptions or anything like that. They might also be offering some API access. I don't know. I am not affiliated with them in any way. Two, it also handles dramatic lighting changes with incredible accuracy. Three, camera movement is not a problem, which is just insane. Just think about it. You specified this image and as the camera moves, it has to imagine what the world would be like around it. I still have a hard time processing that computers can do that today. We thought send to think what a time to be alive. And don't forget about these guys. I mean, if I see something like that, I'm thinking, did I forget to pay back that loan for a new GPU? And now, look, interaction with the environment and a bit of simulation during running. And that is what really sparked my interest because you can combine this with this incredible control model. What is that? Well, your video goes in and it comes out re-imagined, but with a twist because this time that sand can be re-imagined as water. Oh my goodness, look at that. Loving this. Fantastic idea. It looks a bit AIish, but one more paper down the line and that's gone too. I am sure of that. Okay, so semantic changes. These are athletes with fencing swords, right? Not anymore. And this is Master Roshi with golf clubs. Or they can be lightsabers, too. Absolutely amazing. You can also let your imagination run wild by story nightifying yourself and your environment to make those dance moves pop a little more. Or this. This is lowkey amazing. Recreating the world from a muddy mess into a winter wonderland. I love how it makes it more believable by making it snow as well. Very subtle, very effective, and you can even become a completely different character as well. Video game characters, not a problem. Although the results are far from perfect. Or reading a research paper and feeling like a king because you finally understand it. Not a problem. Or if you don't like the lighting on the planet you just created, you can also fix that with just one prompt as well. Absolutely incredible. Wow. And hold on to your papers, fellow scholars, because it generates 5 seconds of video in 2 seconds on one H100 graphics card. Yes, that is faster than you can watch it. Faster than realtime stuff. Wow. Speechless. It's like a chef cooking a five course meal while you're still reading the menu. So why how is that even possible? Dear fellow scholars, this is two minute papers with Dr. Carola. Dr. Carol. Well, the paper reveals that it uses a 1 to 192 spatiotemporal compression variational autoenccoder with 128 latent channels. So simpler. It takes a video and squashes it down to a much smaller, more efficient version, keeping only the most important information. So, the AI has to process far fewer pieces, making it much faster. Also, it operates at a 1 to 8,000 pixels to tokens ratio. That is four times fewer tokens than typical setups. With this, the cost of attention drops a lot. So you can afford full spatiotemporal attention. It's a bit like compressing a messy eight lane grocery line into one express checkout. So really advanced paper with technical details but here in simple words anyone can understand. And finally it uses less than 2 billion parameters before distillation which is absolutely crazy. This could maybe even run on a higherend phone in your pocket. That is a very modest size and that usually comes with modest performance. But not here. Here great performance for a modest size. Loving it. Absolutely incredible ### [5:00](https://www.youtube.com/watch?v=gT98Kq-PV8M&t=300s) Segment 2 (05:00 - 05:00) work. And the even more incredible part is that this is available for free for all of us right now. So with that, don't forget to like, leave a kind comment, and subscribe. And let the experiments begin. On November 4th and 5th in London, the Weights and Biases fully connected conference returns. Yes, join Top AI engineers for 2 days of agent building and model training. It's a fantastic opportunity for networking and learning. Perfect for AI developers, researchers, and business leaders. Yes, that is you fellow scholars. And here's the best part. You can get tickets at 20% off. Just click the link in the video description and use the code to claim your discount --- *Источник: https://ekstraktznaniy.ru/video/12116*