A patent highlighting AMD’s upcoming alternative to DLSS, FidelityFX Super Resolution, or simply Super Resolution has surfaced, giving us some insight into how the company plans on tackling NVIDIA’s dominance in ray-traced titles. At its core, it appears the Super Resolution is using the fundamentals of convolution neural networks trained to upsample images, but using a larger dataset (feature maps) to take advantage of a non-linear upscaling network that retains the color accuracy and detail better than traditional methods:
As you can see in the above images, AMD highlights the problems faced by traditional AI-based upscaling algorithms, namely blurring, distortions, and hallucination of details which was very evident in DLSS 1.0. However, with DLSS 2.0, NVIDIA has largely gotten rid of these flaws, thanks to the use of temporal and spatial data to make sure that the network doesn’t invent or lose details in the process:
