← Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
Live from GTC: Supercharging Next-Gen Video with Yaron Inger of Lightricks
00:08:10
Lightricks Co-Founder and CTO Yaron Inger breaks down how LTX2 is outperforming competitors while staying fully open source.
Loading summary
Transcript12 lines
- [00:00]
Logan
Foreign.
- [00:05]
Podcast Host
Welcome to reshaping workflows with Dell Pro Precision and Nvidia, where innovation meets real world impact in high performance computing.
- [00:20]
Logan
This is Logan live from GTC 2026. This is day three and we're here at LTX at Light Tricks. So I'm with Yaron, also known if you've ever been to Starbucks with him. Known as Dan apparently, is what he said, not me. So first you're on. Really appreciate the time. Tell us, you know, all the listeners out there, kind of what your position is, what you do and then we'll get right into it.
- [00:41]
Jeroen
Okay. So hi, I'm Jeroen, co founder and CTO at Lytrix. And you know we're a company that exists for 13 years now, did a lot of apps on mobile like Facetune. In the last three years we completely transitioned to become an AI. First company started developing our own video model called LTX. The latest version is LTX2 that was released about two and a half months ago. Amazing video model, ranked number one in the world doing video and audio. So multimodal. And it's also open source which is great because everybody can just download and run it. We already have more than 5 million downloads on Hugging Face. And besides the foundational model itself, we also offer API for inference and also a product layer on top which is a one stop shop for video and image creation and storytelling called LTX Studio.
- [01:38]
Logan
So as you know. I did not know. I mean I work with Noah a little bit and I know LTX has been around for a little while. Right. And not a knock to your previous product, but I'd use the original version about a year ago and I mean it was good but nothing was great. Right. And I've installed the newer version and it's great. And I'm not a creative by any stretch of the imagination, but it definitely LTX rivals anything. I mean I'm not comparing your competition but like you know, Runway, like all the stuff you kind of expect, you guys are kind of blowing them out of the water. So I guess what really, and you don't have to share any proprietary secrets or anything like that, but really kind of gen over gen from like LTX1 to say LTX2. What has really changed? Like what have you done to really. I mean, because it went from good to freaking fantastic. So what really changed in that to make that jump?
- [02:20]
Jeroen
So basically everything we scaled everything, whether it's the number of parameters, the data and of course we also added the audio support which is amazing because who nowadays creates a video which is without an audio. And surprisingly so the audio also really affects the quality of the video because the way that we generate video is not just the video, the audio itself. And the audio generation also affects the video generation and vice versa. And so we got a ton of great feedback on features that use the model in maybe surprising ways, like audio to video, where you can maybe you have your own audio track and you want to generate a video for that, like music video for example. And then you can generate characters that sing based on that original soundtrack. And the community says that this is the best model out there for these kind of performances because the audio really drives the expression of the characters and so on. The reason going back to maybe the reason of why this is all open source. So we have various reasons. One is democratization of creativity. We think that right now, as more and more companies go closed source, it's really, really important for us and for the world, I think that to have a model that is completely open, we open source not just the weights, but also the inference code and the training code which allows you to fine tune the model. So if you have your own data, you can just use that to make a better model or a model that's customized for yourself. If you're a studio, and we have many clients that are studios that want to protect their ip, they don't want to send it over the Internet, so they want to do it on Prem. And the ability to run locally on Prem is also something that's very, very important because it means that you can run it in your own network, whether it's on data center gpu, which is worked amazing because our model is extremely efficient and very compressed. But also you can run it on consumer GPUs like we see behind us, we have a dell workstation here, T2 with RTX 6000 and it works amazingly well, which is crazy. So basically all the stack and you can also mix and match the hardware, you can run in low res, generating low resolution on consumer GPUs and then send it to the cloud for upscaling.
- [04:49]
Logan
That's fantastic. So all of that, one, I didn't know, so thank you for educating me. But two, like with all the different things and I hear from a lot of people that I have a lot of friends that are creatives. Like I don't know if you've ever heard Give him a shout out. The mid journeyman, Julian Durant, he loves like your new models, he raves about it. But for you, since you've kind of had the hand in making all this. The video generation with audio makes total sense. But what are some of your other favorite features that you've launched recently that maybe people who haven't used the model in a while should know about?
- [05:17]
Jeroen
Yeah, so we have a bunch of workflows that we are also working as we, as we speak on, on many more. I'll give some examples. But it's really important first to emphasize that in creative workflows you have many, many ways to interact with the model. It's a bit different than LLMs where you just, you know, you prompt to an LLM, you get an output back and so everyone in the community can just download the model and interface with it in any way possible. I think one really great workflow is called Retake, which allows you to take an existing clip, whether it was generated by AI or a real one it and just modify part of that clip. And basically, for example, let's say you have a character talking, you can just replace part of it. So think about use cases for ad tech, for example, where you want to personalize ads. So you shot an ad already, but you want to say something a bit different or you want the actor to act a bit differently with different emotion, or rotated the head a bit to the right instead to the left, you can just select the part that you want to change, then write a prompt. And the model knows how to fill in the blanks really amazingly well and also preserve the audio and the voice signature of the speaker with zero training. So that's a really, really cool feature. And we have many control features like that, many video to video features where you can take human acting and transform it to a different video. Use depth maps to generate videos from. As we speak, we're going to roll more and more features through the API. But the community also drives a lot of these new ideas, which is again the great benefits of having an open source model.
- [06:55]
Logan
Agreed. So with that, I mean I've used LTX2 kind of through Comfy, right? Both on my GB10 and then my Dell Promax T2 with a 6000 Blackwell. So for let's kind of wrap with if somebody's interested in checking out LTX2, learning more about Litrix ComfyUI but where else can they go, like website, where should they download? What would be your recommendation?
- [07:16]
Jeroen
It would be best to go to LTX IO. From there go and download the model. All the links are there. Tagging face GitHub. We have online documentation for open source and for the API.
- [07:29]
Logan
All right, well, you're really Appreciate the Time. Also known as Dan. Really Appreciate the Time. So great model. Check it out. One of my personal favorites. So this is Logan at LTX Elytrix, and we'll see you on the next one.
- [08:02]
Podcast Host
This podcast was produced in partnership with Amaze Media Labs SA.