OpenDLSS-NR: NVIDIA DLSS 5 Neural Network Reimplemented in Vulkan
Original: OpenDLSS: A Vulkan Reimplementation of Nvidia's DLSS 5 Neural Rendering Network
Why This Matters
A verified open reimplementation of DLSS 5's neural core could enable non-NVIDIA platforms to run the same rendering pipeline.
Developer maanHimself has published OpenDLSS-NR on GitHub, a Vulkan reimplementation of NVIDIA's DLSS 5 Neural Rendering network that matches the original bit-for-bit. The project replicates all 71 transformer blocks, FP8 precision, and 141 MiB weight layout, with a separate WebGPU browser port included.
OpenDLSS-NR is an open-source Vulkan reimplementation of NVIDIA's DLSS 5 Neural Rendering (NR) network, built to be bit-exact against NVIDIA's official build 310.8.0. The implementation reproduces all 71 blocks of the Swin/ViT U-Net architecture across six pooling levels, using FP8 (E4M3) activations with FP16 accumulation — and the author claims all 75 intermediate block boundaries match byte-for-byte, not just the final output image.
The network itself is a generative neural renderer, not a traditional upscaler. It re-renders an already-drawn frame, generating detail from injected noise and adjusting tone, structure, and skin rendering based on a style setting. Input and output share the same resolution. Weights (141 MiB) must be supplied separately by the user.
A secondary, independent WebGPU port in ports/browser-webgpu/ runs the same network in a browser without tensor cores or FP8 hardware. The project has attracted around 700 GitHub stars and 60 forks shortly after publication, signaling sharp interest from the developer and modding communities.