TL;DR
Get the latest gadgets delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
OpenDLSS is a public Vulkan reimplementation of NVIDIA’s DLSS 5 neural rendering network, with its authors reporting byte-for-byte agreement at all 75 block boundaries against reference captures. The project requires users to supply model weights and currently targets Windows PCs with recent NVIDIA GPUs and specific driver features; it does not implement DLSS Super Resolution.
A GitHub project called OpenDLSS presents a Vulkan reimplementation of NVIDIA’s DLSS 5 neural rendering network, with its authors reporting that the implementation matches reference outputs byte for byte across all 75 network block boundaries. The code could let developers study or run the network outside NVIDIA’s original implementation, but it requires users to provide the model weights and currently depends on recent NVIDIA GPU and driver support.
The project describes the network as a 71-block shifted-window transformer with a global vision transformer at its deepest level, arranged across six pooling levels. Its account says the model has 141 MiB of weights and uses E4M3 FP8 activations with FP16 accumulation on the Vulkan path. OpenDLSS describes its parity claim as applying not only to the final image but also to the intermediate outputs at every block boundary, compared with reference captures.
OpenDLSS also includes a separate browser WebGPU implementation. The project says this version produces the same captured bytes without tensor cores or FP8, though it is slower: the supplied comparison gives 72 milliseconds at 512-by-512 pixels for the browser port against 2.7 milliseconds for the Vulkan implementation at that resolution. Those are project-reported measurements, not independent benchmarks.
The project’s performance table reports minimum whole-network times on an RTX 4070 SUPER, measured over 40 frames: 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 authors say sustained GPU load can cause clock changes, making median times a few percent higher. The command-line tool processes single frames; temporal feedback is implemented in the separate demo, according to the project documentation.
A New Route to DLSS 5 Research
OpenDLSS matters because it offers an alternate implementation of a proprietary neural rendering system, potentially giving graphics researchers and developers a way to inspect its computation and test it in a Vulkan-based environment. If its parity results hold up under independent testing, the reported block-level matches would make it easier to compare implementations and investigate where rendering differences arise.
The project’s scope is narrower than the DLSS name may suggest. Its authors describe DLSS 5 here as same-resolution generative rendering: the network takes an already rendered frame, additional inputs and conditioning values, then produces an RGB residual and a temporal-blend logit. The project explicitly says it does not implement DLSS Super Resolution, a different network. The release therefore does not amount to a replacement for NVIDIA’s broader DLSS suite or a general upscaler.
Practical access is also limited by the stated requirements. The Vulkan build needs Windows, an NVIDIA Ada-or-newer GPU, and a driver exposing several specified Vulkan and NVIDIA extensions. Users must separately obtain and supply a model directory in the required format. The source material does not establish that the weights are distributed with OpenDLSS or that users have authorization to obtain them from any particular source.
NVIDIA DLSS 5 neural rendering GPU
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Reimplementation Covers
According to the project’s description, the input is a rendered frame represented as a low-dynamic-range proxy, three lanes of Gaussian noise, a reprojected previous-frame output and five conditioning scalars. The network returns four floating-point channels per pixel: an RGB residual and a value used to control temporal blending. The demo incorporates that history input into a feedback loop, while the command-line tool is documented as running without history.
The repository separates a reference Vulkan route, faster NVIDIA-specific PTX kernels, a Filament-based rendering demo and the browser port. The project says the Vulkan implementation uses cooperative-matrix operations and other GPU-specific techniques; the WebGPU version instead prioritizes matching the reference specification without the same hardware acceleration. Its reported speed figures should be read in that implementation-specific context.
NVIDIA is cited by the repository as describing the model in a report titled “DLSS 5: Generative Neural Rendering.” OpenDLSS identifies its target as DLSS-NR build 310.8.0. The supplied material does not include an independent technical review, a response from NVIDIA, or a separate verification of the repository’s claims.
““all 75 block boundaries, byte for byte””
— OpenDLSS project description on GitHub
As an affiliate, we earn on qualifying purchases.
Parity and Model Access Remain Open
The byte-exact results and performance figures are claims reported by the project; the supplied source material does not include independent reproductions or a third-party benchmark. It also does not clarify how users should obtain model weights, what distribution or licensing terms apply to those weights, or whether NVIDIA endorses or supports the reimplementation.
Compatibility may depend on the exact GPU, driver and extension support listed by the repository. The available material does not establish performance across other supported cards, operating systems or driver versions. It also does not provide independent evidence about visual quality across different scenes, style settings or longer temporal sequences.
As an affiliate, we earn on qualifying purchases.
Testing Depends on Users and Weights
The repository provides build, verification, profiling and demo instructions for users who meet its system requirements and have a correctly formatted model directory. Its parity command compares results with fixtures, while a separate verification tool can inspect kernel behavior. Independent testing across compatible systems would help establish whether the reported exactness and performance figures reproduce beyond the project’s own setup.
No future release date, NVIDIA response or formal validation milestone is given in the source material. For now, the next practical steps are for technically equipped users to examine the code, confirm how they may lawfully access weights, and report reproducible results or limitations. Whether the project will broaden hardware support or extend beyond the stated DLSS 5 network remains unknown.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is OpenDLSS?
It is a GitHub project offering Vulkan and browser WebGPU implementations of the DLSS 5 neural rendering network described by the repository.
Does OpenDLSS upscale images?
No. The project describes the network as producing an output at the same resolution as its input and says it does not implement DLSS Super Resolution.
Does OpenDLSS include the model weights?
The project instructions say users must supply a model directory containing the weights. The source material does not specify where users can obtain them or what terms govern their use.
What hardware does the Vulkan version require?
The listed requirements include Windows, an NVIDIA Ada-or-newer GPU and a driver exposing the Vulkan and NVIDIA extensions named by the project.
Have the parity and speed claims been independently verified?
Not in the supplied source material. The byte-for-byte parity and performance numbers are reported by the project; independent confirmation is not provided.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
