TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
OpenDLSS is a public Vulkan reimplementation of NVIDIA’s DLSS 5 neural-rendering network, which its GitHub project says matches the original at all 75 block boundaries. It requires users to supply model weights and currently targets Windows PCs with supported NVIDIA GPUs and drivers; the project is not DLSS Super Resolution.
OpenDLSS, a project published on GitHub, offers a Vulkan reimplementation of NVIDIA’s DLSS 5 neural-rendering network and says its results match the original byte for byte at all 75 block boundaries. The release matters to graphics developers because it provides a separately runnable implementation and a browser-based WebGPU port, but it does not include the model weights: users must provide those themselves.
The project describes the network as a 71-block shifted-window transformer and global vision transformer, arranged across six pooling levels. It runs with E4M3 FP8 activations and FP16 accumulation, and the listed model weights occupy 141 MiB. OpenDLSS says parity covers intermediate block outputs as well as the final image, and includes tools to compare runs against supplied fixtures.
The project distinguishes neural rendering from upscaling. The network takes a rendered frame at its existing resolution, along with three lanes of Gaussian noise, a reprojected previous output and five conditioning values. It returns four values per pixel: an RGB residual and a temporal-blend logit. The project says the model can generate detail and alter tone, structure and skin appearance under a style setting; those descriptions reflect the project’s account of the network’s behavior.
OpenDLSS reports that its Vulkan implementation uses 241 dispatches per frame. On an RTX 4070 SUPER, the project lists minimum runtimes over 40 frames of 2.8 milliseconds at 768-by-768, 7.8 milliseconds at 1920-by-1080, 12.6 milliseconds at 2560-by-1440 and 29.3 milliseconds at 3840-by-2160. The repository cautions that GPU clock changes under sustained load can make median times a few percent higher.
A New Route to Test Neural Rendering
OpenDLSS gives researchers and graphics developers a way to examine and run a claimed replica of the network without relying solely on NVIDIA’s own implementation path. Its reported block-by-block parity is particularly relevant for debugging and reproducibility: matching intermediate results can help identify where differences arise, rather than comparing only final frames.
The browser port broadens that possibility, although it is slower and uses different hardware methods. The project reports 72 milliseconds at 512-by-512 for WebGPU, compared with 2.7 milliseconds for its faster implementation at that resolution. These are project-supplied results, not independent benchmark findings. The code also does not make DLSS 5 available to every graphics card: the main implementation depends on specific NVIDIA GPU and driver capabilities.
NVIDIA RTX 4070 SUPER graphics card
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Repository Implements
OpenDLSS describes its target as DLSS 5 Neural Rendering, which it says corresponds to build 310.8.0. The project’s documentation identifies the network as a generative neural-rendering model rather than an upscaler. It explicitly says DLSS Super Resolution is not implemented, as that is a different network.
The repository includes a C++ host application, Vulkan shader kernels and a faster PTX route, alongside a demo renderer built around Filament. A separate browser-WebGPU implementation is described as independent and bit-exact against the same captures, but without tensor cores, FP8 or cross-block fusion. The demo implements a temporal feedback path; the command-line tool instead runs single frames without history, matching the project’s reference captures.
Running the main tool requires Windows, an NVIDIA Ada-generation or newer GPU, and a driver exposing several specified Vulkan and NVIDIA extensions. A compatible compiler toolchain is also required. The source says the user supplies a model directory containing weights in the documented layout, so the repository is not a self-contained download of the network’s parameters.
““The intermediates match too, not just the final image: all 75 block boundaries, byte for byte.””
— OpenDLSS project description on GitHub
Parity Claims Need Outside Testing
The byte-for-byte match, performance figures and model description are claims made by the OpenDLSS repository in the supplied source material. It provides parity and verification tools, but the material does not include an independent audit, a third-party benchmark or confirmation from NVIDIA that the implementation reproduces its network.
It is also unclear from the supplied information how users can obtain compatible model weights, what permissions or licensing terms apply to those weights, and whether the implementation has been tested across a broad range of supported GPUs and drivers. The project’s compatibility requirements are specific, and its reported timings should not be treated as results for other systems.
Testing and Model Access
The next practical steps are for developers to build the project with compatible hardware, provide a model directory and use the included parity or verification tools against fixtures. Independent testing could clarify whether the reported exact matches and performance hold outside the project’s own setup.
For wider use, users will also need clear, lawful access to compatible weights. The repository documents the expected model-file layout, but the supplied source does not say where the weights can be obtained or announce a schedule for broader hardware support.
Key Questions
What is OpenDLSS?
OpenDLSS is a GitHub project that reimplements NVIDIA’s DLSS 5 neural-rendering network using Vulkan, with a separate WebGPU version for browsers.
Does it include the model weights?
No. The project says users must supply a model directory containing the weights in the documented format. The supplied information does not identify a source for those weights.
Is OpenDLSS an upscaler?
No. The project says the network takes and returns frames at the same resolution and is a neural-rendering network. It says DLSS Super Resolution, a different network, is not implemented.
What hardware does the Vulkan version require?
The listed requirements include Windows, an NVIDIA Ada-generation or newer GPU, and a driver exposing the Vulkan and NVIDIA extensions named in the repository. The source does not establish support for other GPU vendors.
Have the parity and speed claims been independently confirmed?
Not in the supplied source material. The project reports byte-for-byte parity and specific benchmark times, but no independent verification or third-party benchmark is provided there.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
