OpenDLSS-NR Reimplements DLSS 5 in Vulkan, Bit-for-Bit
A third-party Vulkan port that reproduces DLSS 5's output exactly shows how much of the upscaler's behavior can be replicated outside NVIDIA's own stack, on the same Tensor Core hardware and with the model supplied by the user.
Reporting from 1 source: GIGAZINE.
A GitHub project called OpenDLSS-NR reimplements NVIDIA's DLSS 5 neural rendering network in Vulkan. It runs the same 71-block Swin/ViT network as DLSS-NR build 310.8.0 using FP8 on Tensor Cores, and the output is described as bit-for-bit identical to the original, with all 75 block boundaries matching byte-for-byte. It requires Windows, an NVIDIA Ada Lovelace or later GPU, and the user's own DLSS 5 model. The project states it is unrelated to NVIDIA.
The project re-renders frames the engine has already drawn, generating detail from injected noise, and adjusts tone, structure, and skin according to style settings. Input and output stay at the same resolution, so it is not an upscaler in the usual sense. A separate implementation runs the same network in a browser through WebGPU. On an RTX 4070 SUPER, the minimum time to generate one frame is measured while the run is held for 40 frames or more. Users must supply the DLSS 5 model themselves, and the project says it is unrelated to NVIDIA.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.