NVIDIA's DLSS 5 has already attracted considerable attention from the modding community, and developer MAAN has now released OpenDLSS-NR, an open-source Vulkan reimplementation of the technology's Neural Rendering network. The project recreates the original 71-block neural network found in DLSS-NR runtime/DLL version 310.8.0, with the developer claiming bit-exact results against NVIDIA's implementation, including intermediate outputs rather than just the final rendered image.

Unlike DLSS Super Resolution, which reconstructs higher-resolution images from lower-resolution inputs, DLSS 5 Neural Rendering operates on an already rendered frame, enhancing its appearance with AI-generated detail and adjustments to lighting, materials, and facial features. OpenDLSS-NR reproduces this process using a U-Net architecture combining shifted-window transformers with a global Vision Transformer (ViT), FP8 activations, FP16 accumulation, and approximately 141 MiB of model weights.

The implementation also includes optimized PTX kernels for NVIDIA's Tensor Cores, alongside GLSL compute shaders. MAAN has shared preliminary performance results obtained with an RTX 4070 SUPER, revealing the following execution times for the complete network:

These figures represent the minimum execution times recorded across 40 frames, rather than average frame times. They nevertheless demonstrate that the reimplementation can execute on Ada Lovelace hardware, whereas NVIDIA officially supports DLSS 5 on GeForce RTX 50 Series GPUs, with planned official support for RTX 40 Series GPUs in the future.

Perhaps even more interestingly, OpenDLSS-NR includes an independent WebGPU implementation that reproduces the same neural network without relying on NVIDIA Tensor Cores or native FP8 support. This potentially opens the door to running the technology on other GPU architectures and platforms, although performance remains substantially lower. MAAN reports an execution time of 72 ms at 512 × 512 with the WebGPU implementation, compared to approximately 2.7 ms with the optimized native implementation.

There are, however, some important limitations. The repository does not distribute NVIDIA's proprietary model weights, which users must supply separately. Furthermore, OpenDLSS-NR is currently a standalone implementation rather than a ready-to-install mod for existing games. It includes a demonstration renderer based on Google's Filament engine, complete with motion vectors and temporal feedback, but does not implement DLSS Super Resolution.

Nevertheless, the project represents an interesting development for the modding community, particularly as enthusiasts continue experimenting with bringing NVIDIA's latest neural rendering technologies to older RTX GPUs and alternative hardware.

OpenDLSS-NR is available on GitHub under the MIT license, with additional information and demonstrations available through MAAN's original X post.

Follow Wccftech on Google to get more of our news coverage in your feeds.