Deep Learning Super Sampling

Deep Learning Super Sampling, usually shortened to DLSS, is a family of real-time image enhancement and upscaling techniques developed by Nvidia for video games. The idea is to run most of the graphics pipeline at a lower resolution, which is cheaper, and then use a trained neural network to infer a higher resolution image that approximates the detail of a native render. That trade lets a player choose higher graphical settings or higher frame rates for a given output resolution. The first generation arrived with the GeForce RTX 20 series in 2018 and was mainly a spatial upscaler built from convolutional auto-encoder networks, trained separately for each game. DLSS 2, shipped in 2020, became a temporal upsampling design that draws on motion vectors, depth and exposure data from earlier frames and no longer needs per-game training. Later releases added Frame Generation, Multi Frame Generation and Ray Reconstruction, and the same network can be run at full resolution as an anti-aliasing mode that Nvidia brands DLAA.