Quote:
Originally Posted by nevcairiel
Wut? Most Neural Networks rely on GPUs to run them at any speed, because GPUs are actually pretty good at that, because it can be paralized really well. Not to mention even newer stuff with specialized hardware just for that task, ie. Tensor Cores.
We had NNEDI3 in madVR for a while, a Neural Network based scaler. But it was removed because it was ultimately decided that the added complexity in madVR was not worth it (since it was the only component to use OpenCL IIRC), and other algorithms could flat out replace it at higher speeds and quality.
Neural Networks can be really powerful, and who knows if some day one might return to madVR, but the real challenge with Neural Networks is training them. You need absolutely massive compute power to train a Neural Network. All those fancy NVIDIA demos you see for image processing, those are trained on super computers, large clusters with hundreds if not thousands of NVIDIA GPUs to train the network. Of course no mere mortal has that sort of resources to train a network.  |
It was mentioned a while ago that NGU is based on neural networks. madshi posted at some point that the training was done by comparing downscaled images to the original. Maybe he has an evil lair with a bunch of super computers/GPUs to do this type of training?
If you look at screenshots, you can see that NGU picks up details like eyelashes that are missed by Lanczos and super-xbr. NGU very high pushes the GPU so hard that hundreds of thousands of repeated calculations must be taking place. NGU Anti-Alias is also a very similar upscaler to NNEDI3 and most seem to prefer it.
Personally, I think all images should eventually be scaled by neural networks when Tensor Core technology is scaled down. If, like me, you are frequently comparing Kodi VideoPlayer with Lanczos3 to NGU Sharp very high, they aren't the same image. Lanczos creates some artifacts, but it also smears a lot of valid detail. All of those eyelashes add up when the image is large.
With that said, I don't think NGU Sharp was trained well with poor material. I have seen some content look like an oil painting that kind of made me drunk while watching it.