Modder Gets DLSS 5 Running in a Web Browser Through WebGPU

Nvidia’s DLSS 5 neural rendering technology has been made to run directly inside a web browser through an unofficial WebGPU implementation, allowing the technology to operate without its normal Nvidia specific software stack.

The experimental project was created by a developer known as MAAN. The browser demo can run on macOS and hardware from vendors other than Nvidia because its neural network has been reimplemented using WebGPU compute shaders rather than relying on Nvidia’s normal DLSS interfaces and driver requirements.

The result is technically impressive, although it is far too slow for conventional gaming in its current form.

DLSS 5 browser demo uses a 147 MB neural model

The implementation consists of roughly 147 MB of model weights and a compressed JavaScript runtime measuring around 1 MB. The neural model reportedly uses weights extracted from a leaked DLSS 5 library, although it is unclear whether those weights differ from Nvidia’s official release.

The interactive demo includes several prepared 3D scenes and allows you to rotate models in the browser. You can also upload your own compatible 3D assets through file selection or drag and drop.

The normal 3D viewer can remain responsive, but applying the DLSS 5 neural rendering pass is considerably slower.

Testing on an RTX 40 series desktop showed that each processed image could require roughly one or two seconds. That makes the technology unsuitable for a game targeting 30, 60 or more frames per second, where processing would need to happen in milliseconds rather than seconds.

WebGPU removes the usual Nvidia hardware dependency

Nvidia normally provides DLSS through its NGX framework or Streamline SDK for applications using graphics APIs such as DirectX and Vulkan.

WebGPU is not listed as an officially supported route.

The browser project instead reconstructs the neural network using general WebGPU compute functionality. Because WebGPU is available across different hardware platforms, the implementation does not inherently require an RTX graphics card or Nvidia’s proprietary driver path.

That explains why the demonstration can operate on macOS and potentially other WebGPU compatible graphics hardware.

It does not mean Nvidia has officially made DLSS 5 hardware independent. This remains an unofficial reimplementation rather than a supported DLSS release.

The browser version may be more useful outside gaming

At one or two seconds for each neural rendering pass, this version of DLSS 5 has little immediate value for real time games.

There are other applications where the processing delay could matter much less.

Architectural visualization, 3D model previews, product design tools and browser based graphics applications often work with individual images or relatively static scenes. In those situations, waiting a second or two for a higher quality processed image could be more practical than trying to maintain a real time frame rate.

The project also demonstrates how neural graphics techniques can potentially be separated from the environment for which they were originally designed.

DLSS 5 officially debuted with NBA 2K27 and was initially associated primarily with RTX 50 series hardware. Modders have since experimented with running the technology on older GeForce cards and other GPUs. Nvidia has also said official RTX 40 series support is planned.

The WebGPU experiment pushes that work further by putting the neural renderer inside an ordinary browser tab. Performance remains the major limitation, but the project shows that at least part of the DLSS 5 neural rendering pipeline can be reconstructed using broadly available GPU compute technology rather than being permanently tied to one graphics architecture.