OpenDLSS-NR, a Vulkan-based re-implementation of NVIDIA's DLSS 5 for high-resolution gaming, boasts an astonishingly perfect match to the original down to the bit level.

A project called ' OpenDLSS-NR ' has emerged that reimplements NVIDIA's AI-powered upscaling technology '

The results of upscaling with OpenDLSS-NR are as follows.

OpenDLSS-NR is a Vulkan reimplementation of DLSS 5, described as 'a perfect match to the original down to the bit.' It runs the same 71-block Swin/ViT network as DLSS-NR build 310.8.0 on Tensor Cores using FP8 . Intermediate results match, and not only the final image, but all 75 block boundaries match down to the byte. OpenDLSS-NR re-renders frames already drawn by the engine, generating detail from injected noise. It adjusts tone, structure, and skin according to the style settings. The input and output are at the same resolution and it is not an upscaler. An implementation that runs the same network in a browser on WebGPU also exists.

Please note that users will need to provide their own DLSS 5 model.

The following are the results for the minimum time it took to generate one frame with an RTX 4070 SUPER. This is the minimum value when the game was paused for 40 frames or more.

This requires Windows, an NVIDIA Ada Lovelace or later GPU, and a compatible driver to run. It is explicitly stated that this is 'unrelated to NVIDIA,' so caution is advised when using it.