The Super Resolution Demo: A Live Showcase of Advanced Technology

We have two demos to showcase today, and we're starting with the super resolution demo that's been generating a lot of buzz. We're running a gaming content through our QSRNet, a deep learning solution that's been trained to upscale this kind of content in real-time. What's impressive is that we've managed to compress this process into a small model that can run fast enough to produce stunning results.

As you can see on your screen, the demo is running smoothly, and we're able to pause it at any time to highlight the differences between the original image and the upscaling result. For example, take a look at the left side of the image – you'll notice that the textures aren't as sharp as they are on the right side. This is where our super resolution technology shines, taking a 540p input and producing a stunning 4K image.

One of the most impressive aspects of this demo is its speed. We're running it at an incredible 142 frames per second, even with a cap on the frame rate to ensure that we can see the effects in real-time. This level of performance is made possible by our advanced neural network architecture, which allows us to process complex images quickly and efficiently.

Now, let's take a closer look at what's happening behind the scenes. We're using a technique called integer four precision, which enables us to run the demo on a lower-powered device while still achieving stunning results. This is where our technology truly shines – by leveraging advanced algorithms and computational techniques, we can bring high-quality images to devices that wouldn't normally be able to handle them.

But what's even more exciting is where this technology is headed. We're already seeing applications of super resolution in various fields, from PC gaming to mobile devices. And with the recent advancements in AI research, we're poised to take this technology to new heights. In fact, our team has been working on integrating these techniques into edge devices, which will enable us to bring high-quality images to a wider range of users.

To give you a better idea of what's possible, let's compare our demo to some other super resolution technologies available in the market. While Snapdragon GSR from Qualcomm is also impressive, our approach uses a different algorithmic framework that allows for even faster processing times and higher quality results. This is especially notable when compared to algorithms used by companies like Nvidia for their DLSS technology.

Finally, it's worth noting that we've been pushing the boundaries of what's possible with super resolution even further by experimenting with lower precision formats, such as four-bit integers. While these smaller values may not be immediately noticeable in our demo, they do offer significant power savings without sacrificing too much image quality. This is a crucial aspect of any technology aimed at mobile devices – reducing power consumption while maintaining performance is essential for widespread adoption.

In conclusion, our super resolution demo showcases the incredible advancements that have been made in this field. By leveraging deep learning and advanced algorithms, we've created a technology that can upscale images in real-time with stunning results. Whether it's PC gaming or mobile devices, we're confident that our technology will continue to push the boundaries of what's possible in image processing.

"WEBVTTKind: captionsLanguage: enwhy don't we why don't you guys want to light up a couple of demos and uh let poor Joseph breathe a little bit here might be a good idea um so let's uh Joseph we have two demos we have a stable diffusion live on a handset and we have uh an INT for uh super resolution demo I think why don't we um why don't we start because we have we probably have a fair number of gamers in our audience why don't we start with the super resolution uh in Ford demo if you can if you can pull that off and I think we can get it running folks this is live it's not Memorex I think we can get this running and live so that Joseph can actually talk a little bit over it and tell us what we're seeing yeah yeah okay so what we have is um gaming content that's running through uh our supervisor position network we call it uh qsrnet and what it is a it is a deep learning Solution that's been trained to up sample this kind of uh content and it's something that can run very fast and we you know we've it's an example of um focusing on the domain having a smaller model being able to run it fast and and this is also even further quantized in four now I'm actually running it um with a cap on the frame per second just so that you can see uh things you know this is like the natural speed of movement so you can kind of see things that are going on and you know I can pause it and you can see you know the left side can be more Jagged than the right side it can be more aliasing going on with the background um catch something here textures aren't quite as sharp left yeah and then I think earlier there was a flame and and uh you know the Flames themselves also weren't rendered as accurately so this is something of an example where yeah if we if we have a certain content we can make it run really well now what what we can even do further when we move to uh integer Network and even in force is we can see how fast we can run it and and this is where we have this other mode where it looks like it's fast forward but it's actually more about running it as fast as possible you know as the processor will allow us and so here we have um the uh frame per second on the top right corner is is uh almost like 140. assistant and that's even though we're zooming in on a crop the you know the left says like a 137 by 137 crop this output is is doing the whole rendering so we are getting a 4K image coming out um and in the problem with 540p input yeah from a 540p input so the process is going through all the work to do it and it's still getting at 142 frames per second uh meanwhile uh we're zooming in so you can you can take a look and see the differences between the left and the right so this we we've seen we saw uh something from Qualcomm earlier uh I guess I'm gonna say about a month ago I think um called Snapdragon GSR which is different this isn't an algorithmic uh implementation of super resolution um so very much similar to uh let's say dlss from Nvidia or what have you for you know PC gaming um this is similar in its algorithmic uh processing uh and you know integer four Precision is what is what you're actually showcasing here right um yeah the the like the algorithm behind it might be a little different uh than some of the other ones from other companies yeah um but I would say that yeah even relative to what you you mentioned at Snapdragon Summit this is more of like AI research demo showing that you can run a neural network based um super resolution algorithm and you know get it up to this resolution and this speed because I think at the time when we kind of started this project it wasn't it was kind of in question and we wanted to demonstrate that yeah these techniques are really really amazing but they're also something that we can you know potentially bring to the edge one day um from you know the research and the papers and so forth so this is where yeah so that's that's where the the point of this is um but yeah and then we take it a step further within four like even last year um we did even show this with say natural image and in eight but then we you know pushed it even further and went down to forbid what what the the difference um not just in speed but like for going down to four bit more the benefit was that we it's hard to see it now but we had a power measurement too I'm getting um like a better power drill um when running these because because this is super resolution just running every frame and it's running it you know if I run it at the max rate that's allowable it's running every you know um it's running 142 frames per second so you can imagine that if we can spend a lot less power on that that's a big dealwhy don't we why don't you guys want to light up a couple of demos and uh let poor Joseph breathe a little bit here might be a good idea um so let's uh Joseph we have two demos we have a stable diffusion live on a handset and we have uh an INT for uh super resolution demo I think why don't we um why don't we start because we have we probably have a fair number of gamers in our audience why don't we start with the super resolution uh in Ford demo if you can if you can pull that off and I think we can get it running folks this is live it's not Memorex I think we can get this running and live so that Joseph can actually talk a little bit over it and tell us what we're seeing yeah yeah okay so what we have is um gaming content that's running through uh our supervisor position network we call it uh qsrnet and what it is a it is a deep learning Solution that's been trained to up sample this kind of uh content and it's something that can run very fast and we you know we've it's an example of um focusing on the domain having a smaller model being able to run it fast and and this is also even further quantized in four now I'm actually running it um with a cap on the frame per second just so that you can see uh things you know this is like the natural speed of movement so you can kind of see things that are going on and you know I can pause it and you can see you know the left side can be more Jagged than the right side it can be more aliasing going on with the background um catch something here textures aren't quite as sharp left yeah and then I think earlier there was a flame and and uh you know the Flames themselves also weren't rendered as accurately so this is something of an example where yeah if we if we have a certain content we can make it run really well now what what we can even do further when we move to uh integer Network and even in force is we can see how fast we can run it and and this is where we have this other mode where it looks like it's fast forward but it's actually more about running it as fast as possible you know as the processor will allow us and so here we have um the uh frame per second on the top right corner is is uh almost like 140. assistant and that's even though we're zooming in on a crop the you know the left says like a 137 by 137 crop this output is is doing the whole rendering so we are getting a 4K image coming out um and in the problem with 540p input yeah from a 540p input so the process is going through all the work to do it and it's still getting at 142 frames per second uh meanwhile uh we're zooming in so you can you can take a look and see the differences between the left and the right so this we we've seen we saw uh something from Qualcomm earlier uh I guess I'm gonna say about a month ago I think um called Snapdragon GSR which is different this isn't an algorithmic uh implementation of super resolution um so very much similar to uh let's say dlss from Nvidia or what have you for you know PC gaming um this is similar in its algorithmic uh processing uh and you know integer four Precision is what is what you're actually showcasing here right um yeah the the like the algorithm behind it might be a little different uh than some of the other ones from other companies yeah um but I would say that yeah even relative to what you you mentioned at Snapdragon Summit this is more of like AI research demo showing that you can run a neural network based um super resolution algorithm and you know get it up to this resolution and this speed because I think at the time when we kind of started this project it wasn't it was kind of in question and we wanted to demonstrate that yeah these techniques are really really amazing but they're also something that we can you know potentially bring to the edge one day um from you know the research and the papers and so forth so this is where yeah so that's that's where the the point of this is um but yeah and then we take it a step further within four like even last year um we did even show this with say natural image and in eight but then we you know pushed it even further and went down to forbid what what the the difference um not just in speed but like for going down to four bit more the benefit was that we it's hard to see it now but we had a power measurement too I'm getting um like a better power drill um when running these because because this is super resolution just running every frame and it's running it you know if I run it at the max rate that's allowable it's running every you know um it's running 142 frames per second so you can imagine that if we can spend a lot less power on that that's a big deal\n"