OpenDLSS-NR Reimplements Nvidia's DLSS 5 Network in Vulkan, but You Supply the Weights

A developer has published OpenDLSS-NR, an MIT-licensed Vulkan reimplementation of the neural rendering network inside Nvidia's DLSS 5, and claims its output matches Nvidia's byte for byte. The repository ships no model weights, so nobody can reproduce that claim without first obtaining Nvidia's.

Wccftech reported the project on Sept. 23, saying it found no independent verification of the bit-exact claim and no response from Nvidia. The repository later reached Hacker News with 265 points.

What the OpenDLSS-NR repository contains

The OpenDLSS-NR README describes a 71-block Swin and ViT transformer network running on FP8 tensor cores. It re-renders a finished frame to add detail, tone and structure. It is not an upscaler, and the README says DLSS-SR, the separate super-resolution network, is "not implemented".

The licence line reads "MIT for everything in this repository". The disclaimer reads "This project is not affiliated with, endorsed by, or supported by NVIDIA." The code carries no Nvidia headers or software.

The weights problem

The README says "You supply the weights", as a model directory with a manifest.json describing the 71 blocks. The aloshdenny/open-dlss copy adds that nothing in the repository produces that directory and gives no instructions for obtaining one. Wccftech puts the weights at about 141 MiB.

That matters for the headline claim. The aloshdenny README says its parity and verify commands compare "all 75 block boundaries, byte for byte" against captured fixtures from the original. Those fixtures are the author's. A reader without the weights can build the code and cannot check the result.

Two repositories carry the same project. One sits under SilyNoMeta, whose page says it was forked from maanHimself/OpenDLSS-NR, and the other under aloshdenny. Wccftech credits a developer called MAAN. On the day the pages were fetched they showed 0 and 5 stars. Hacker News attention is not adoption.

Performance, as the author reports it

The README gives minimum frame times over 40 frames on an RTX 4070 SUPER. These are the author's figures, not independent measurements, and the benchmark tool runs single frames with no temporal history.

1. 768x7682.8 ms
2. 1920x10807.8 ms
3. 2560x144012.6 ms
4. 3840x216029.3 ms

Source: OpenDLSS-NR README, accessed 2026-10-02

At 4K that is 29.3 ms, roughly 34 frames per second if nothing else ran on the card. The repository is a standalone implementation, not a mod that plugs into an existing game.

Which GPUs it runs on

GPU family	Path	Reported cost
Ada and Blackwell (RTX 40 and 50)	Native FP8 cooperative matrices	2.8 ms at 768x768
Ampere (sm86)	FP16 tensor cores, experimental, not bit-exact	about 14 ms at 512x512
Ampere, software path	E4M3 emulated in GLSL	about 0.65 s at 512x512
Any browser	WebGPU port, no tensor cores	72 ms at 512x512

The WebGPU port is the part that matters outside Nvidia's ecosystem. The author says it matches the Vulkan output without tensor cores, at 72 ms against roughly 2.7 ms natively.

What is not known

The README does not say where the weights come from. Whether a network reimplementation that needs the original's weights sits comfortably with the MIT licence is a question for a lawyer, and the README leaves model-data licences to the user.

Nvidia has not commented in any source fetched for this story. The verification question has a precedent on this site: Bez, a browser engine generated from specs and tests, also sells a reimplementation whose correctness depends on the test material. For Nvidia's wider position, see our report on the chip lease-back. The next signal is whether independent testers publish parity results using weights they obtained themselves.