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 GitHub project that says it reimplements NVIDIA’s DLSS 5 neural rendering network in Vulkan, matching the original at all 75 block boundaries when supplied with model weights. It also offers an independent WebGPU port; the project is not DLSS Super Resolution, and access to model weights is separate from the code.
The GitHub project OpenDLSS publishes a Vulkan reimplementation of NVIDIA’s DLSS 5 neural rendering network, which its developer says matches the original byte for byte at all 75 block boundaries. The project also includes a separate WebGPU version designed to run in a browser, but users must supply the model weights, and the code is not NVIDIA’s DLSS Super Resolution upscaler.
According to the project’s GitHub description, the network uses 71 shifted-window transformer blocks across six pooling levels, with a global vision transformer at the bottom. It uses E4M3 FP8 activations with FP16 accumulation and has 141 MiB of weights. OpenDLSS says its implementation reproduces intermediate outputs as well as the final result, with comparisons covering 75 block boundaries. These are claims made by the project; the supplied material does not include an independent evaluation.
The network takes a rendered frame and additional inputs, including three lanes of Gaussian noise, a reprojected previous output and five conditioning values. It returns four channels per pixel: an RGB residual and a temporal-blend logit. The project describes this as generative neural rendering: it re-renders an engine’s existing frame, rather than increasing its resolution. Its demo implements a temporal feedback path, while the command-line tool processes single frames without history.
OpenDLSS lists performance figures from an RTX 4070 SUPER, measured as minimum network time over 40 frames: 2.8 milliseconds at 768×768, 7.8 ms at 1920×1080, 12.6 ms at 2560×1440 and 29.3 ms at 3840×2160. The project says the GPU alternated between clock states during sustained load, making medians a few percent higher. The WebGPU port is reported at 72 ms at 512×512, compared with 2.7 ms for the Vulkan route at that resolution. The source does not provide an independent benchmark or broader hardware comparisons.
What the Vulkan Port Makes Possible
The release gives developers and researchers a way to inspect and run a claimed reproduction of a neural rendering pipeline outside NVIDIA’s own implementation, provided they have compatible hardware, drivers and weights. The project exposes its network graph, arithmetic reference code and GPU kernels, which could help technical users examine how the model’s stages fit together. The reported intermediate-output matching, if independently reproduced, would offer a more detailed test than comparing final images alone.
The browser port broadens the implementation options, but its reported speed is much lower than the Vulkan version and it does not use tensor cores or FP8. Hardware requirements also limit access: the Vulkan build calls for Windows and an NVIDIA Ada-generation or newer GPU with specified extensions. These constraints mean the project is not a general replacement for NVIDIA’s supported features, nor evidence that the network runs equivalently across GPUs.
NVIDIA DLSS 5 neural rendering GPU
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
DLSS 5 Is Not Super Resolution
The project’s description draws a distinction between DLSS 5 neural rendering and DLSS-SR, which it says is a different network and is not implemented. OpenDLSS operates at the input frame’s resolution and describes its role as adjusting or generating image detail under a style setting, rather than upscaling an image. That distinction matters because the DLSS name is often associated with resolution upscaling, while this project targets a separate rendering process.
The Vulkan implementation has a reference route in GLSL and a faster route using generated PTX kernels. The project also patches Filament for its demo renderer, including motion-vector and Vulkan interoperability support. Its WebGPU port is presented as an independent implementation that matches the same captures, but uses neither FP8 nor tensor cores. OpenDLSS says the weights must be supplied in a model directory; the source material does not say the project distributes them.
“A Vulkan reimplementation of NVIDIA’s DLSS 5 Neural Rendering network, bit-exact against the original.”
— OpenDLSS project description on GitHub
What Parity Claims Do Not Establish
The supplied source is the project’s own GitHub description. It does not include independent testing of bit-exact parity, image quality or performance, nor does it provide the model weights. The exact provenance and availability of weights, and whether all users can obtain compatible files, are not established by the material provided.
Results may also depend on the GPU, driver, model files and execution path. The stated Vulkan requirements identify a narrow compatible setup, while the reported timing figures cover one GPU. The project says its command-line tool runs without temporal history, so those single-frame tests do not describe every part of the demo’s feedback behavior. It is also unclear whether NVIDIA has commented on or endorsed the reimplementation.
Testing on Supported Systems
OpenDLSS provides build, benchmark, profile and parity commands for users who meet its listed Windows and NVIDIA hardware requirements and have the model directory. The next useful checks would be independent parity runs against the stated fixtures, along with benchmark results on other supported GPUs and drivers. Until those are available, the project’s accuracy and performance figures should be read as developer-reported results.
The project description does not state a release schedule or announce a forthcoming milestone. Further clarity would come from documentation or testing that addresses model-weight access, compatibility beyond the listed setup, and whether the reported frame timings hold across systems.
Key Questions
What is OpenDLSS?
OpenDLSS is a GitHub project offering a Vulkan implementation of NVIDIA’s DLSS 5 neural rendering network and a separate WebGPU port. Its bit-exact matching claims come from the project itself.
Does it upscale images?
No. OpenDLSS says the network takes and returns frames at the same resolution. The project says DLSS-SR, a different network, is not implemented.
Does OpenDLSS include the model weights?
The project says users must provide the weights in a model directory. The supplied description does not establish whether or where those weights can be obtained.
What hardware does the Vulkan version require?
The listed requirements include Windows, an NVIDIA Ada-generation or newer GPU, and a driver exposing several specified Vulkan and NVIDIA extensions. The source does not claim support for other platforms or GPU vendors.
Are its speed and accuracy independently verified?
Not in the source material provided. The parity and benchmark figures are developer-reported; independent tests are needed to confirm how they hold up on other systems.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
