The "OpenDLSS-NR" project has emerged, aiming to reimplement NVIDIA's announced AI upscaling technology, "DLSS 5," for Vulkan.
OpenDLSS-NR released to reimplement DLSS 5 for Vulkan, achieving bit-perfect consistency with the original
This article is a translation. Read the Japanese original
OpenDLSS-NR is described as an implementation that is bit-perfectly identical to the original network. This project runs the same 71-block Swin/ViT (Vision Transformer) network used in DLSS 5 build 310.8.0 on Tensor Cores using FP8. As a result, it is reported that not only the intermediate results but also all 75 block boundaries of the final image match the original down to the byte.
This technology works by re-rendering frames already rendered by the engine and generating detail from injected noise. Depending on style settings, it adjusts tone, structure, and skin texture. Note that the input and output resolutions are the same; it is not a resolution-increasing upscaler. To use it, users must provide their own DLSS 5 models.
The required operating environment includes Windows, an NVIDIA Ada Lovelace architecture GPU or newer, and compatible drivers. The project explicitly states that it is unaffiliated with NVIDIA.
Sources
- ゲームを高解像度化するNVIDIAの「DLSS 5」をVulkanで再実装した「OpenDLSS-NR」、なんとオリジナルとビット単位で完全に一致 （GIGAZINE、2026-10-02）