OpenDLSS-NR Ships a Vulkan DLSS 5 Network Under MIT, Minus the Weights That Make It Run

A developer who goes by MAAN published OpenDLSS-NR on Sept. 21, 2026, a Vulkan reimplementation of the neural network behind NVIDIA's DLSS 5 Neural Rendering, under the MIT licence. The repository shows 691 stars and 59 forks, and it contains no NVIDIA weights, which means it cannot render a frame until the user supplies them.

The code is the part the repository can legally and technically hand over. The model data is not, and the README says so in two sentences that matter more than the benchmark table.

What the repository contains

Per the README, the project reimplements a 71-block network of shifted-window (Swin) and Vision Transformer (ViT) layers, matching what it calls DLSS-NR build 310.8.0. Wccftech's report adds the layout: a U-Net with six pooling levels, FP8 activations with FP16 accumulation, and about 141 MiB of weights.

It is not an upscaler. The README describes a network that "re-renders the frame the engine already drew, generating detail from injected noise and adjusting tone, structure and skin under a style setting." Input and output share a resolution.

The maintainer's headline claim is parity: "The intermediates match too, not just the final image: all 75 block boundaries, byte for byte." That is the developer's own test, and The Terminal found no independent run of it.

The weights are the open question

The README says the project "contains no NVIDIA software, weights, headers, or instructions for obtaining them." It adds that "No rights under any NVIDIA intellectual property are granted or implied by this repository or its license, and you are responsible for the licenses that apply to whatever model data you use with it."

That split resembles the one in VoiceStudio, an AGPL application whose default voice model is non-commercial. In both, the permissive code licence says nothing about the file that does the work. Here the file is NVIDIA's, and the project does not claim otherwise.

One inference follows from the parity claim. To compare 75 intermediate buffers byte for byte against the original, the developer must have run the original network, so MAAN had a copy of the weights. The README does not say how, and The Terminal draws no conclusion about it.

What it needs and how fast it runs

The listed requirements are Windows, an NVIDIA Ada-generation GPU or newer, Visual Studio 2022 or later with the C++ toolset, Python 3, Node.js and specific Vulkan extensions. Wccftech lists the developer's timings on a GeForce RTX 4070 SUPER, a card from the Ada generation (the photograph above shows a standard RTX 4070, not the SUPER):

Resolution	Time per frame
768 x 768	2.8 ms
1920 x 1080	7.8 ms
2560 x 1440	12.6 ms
3840 x 2160	29.3 ms

1. 768 x 7682.8 ms
2. 1920 x 10807.8 ms
3. 2560 x 144012.6 ms
4. 3840 x 216029.3 ms

Source: OpenDLSS-NR benchmarks as reported by Wccftech, accessed 2026-10-01

Those are the developer's numbers. A 60-frames-per-second game has 16.7 ms for the whole frame, so the 1440p figure of 12.6 ms leaves about 4 ms for everything else, and 4K at 29.3 ms works out to roughly 34 frames per second for this pass alone. That is The Terminal's arithmetic, not a measured game result.

The project is also not game-ready. Wccftech says it is a standalone implementation with a demonstration renderer rather than an installable mod. A WebGPU port runs in a browser at 72 ms for 512 x 512, per the same report.

NVIDIA's position and the other reimplementation

Wccftech reports that NVIDIA officially supports DLSS 5 on GeForce RTX 50 Series cards, with RTX 40 Series support planned for later. That is the draw: the Ada-class hardware OpenDLSS-NR targets is the hardware NVIDIA has not yet enabled. The Terminal found no NVIDIA statement on the repository.

A second project, dlss-nr-on-intel, describes an independent reimplementation of the inference pass running on an Intel Xe2 integrated GPU through a Vulkan layer, "Code only. No NVIDIA binaries, no weights." The Terminal saw only its search listing, not the repository.

The repository's issue tracker showed no open issues when read. PageIndex is a reminder that star counts and a maker's own benchmark are separate evidence.