If you want to upscale images and videos without sacrificing quality, you can't go wrong with "Waifu2x-ncnn-Vulkan GUI Edition Reloaded" or "Real-ESRGAN-GUI". These tools use advanced algorithms combined with the processing power of your graphics card to deliver stunning results. But if you prefer an even more complex alternative that includes secondary parameters and additional models, then you need to download a copy of Video2x.
We live in the world of artificial intelligence... at least when it comes to upscalers. Nvidia keeps adding deep learning features to its graphics cards (DLSS, DLDSR, DLAA), and available information suggests that AMD's FSR third version has a touch of AI inside, but if you're thinking of something more "simple" like upscaling images and videos, the selection of available models is much broader. With that in mind, we come to Video2x, a project that puts half a dozen models or drivers at a click's distance.
How to Upscale Images and Videos with Video2x
The first thing that stands out about Video2x is the large number of advanced parameters in its interface. At first glance, we recognize details like the number of processes and the scale to reproduce, but if we click on the Driver Settings tab, we'll find special values for each of the six models: Waifu2x Caffe, Waifu2x Converter CPP, Waifu2x NCNN Vulkan, SRMD NCNN Vulkan, RealSR NCNN Vulkan, and Anime4K CPP.
As a starting point, I recommend doing the first tests using Waifu2x NCNN Vulkan, with the cunet submodel and TTA enabled. The TTA option increases the final quality of the frames, but multiplies the processing time by 8. In the case of RealSR NCNN Vulkan, the DF2K submodel is recommended for processing images with very little noise, but if artifacts appear, it's best to use DF2K-JPEG. RealSR is much slower than Waifu2x (even with a graphics card), so I recommend patience.
Video2x saves its results in the same source folder (unless you modify the path), and the Preserve Frames value prevents the program from deleting the cache folder. The time available has only allowed me to evaluate upscaling of images, but I am very satisfied with its results. To finish, the download from GitHub amounts to 1000 megabytes. Link below!
Official site and download: Click here
