OpenDLSS: A Vulkan Reimplementation of Nvidia's DLSS 5 Neural Rendering Network
OpenDLSS is a Vulkan reimplementation of NVIDIA's DLSS 5 Neural Rendering network, achieving bit-exact parity with the original model's intermediates and output.
summary is enough The project is a highly impressive technical feat of reverse engineering, but it requires external weights and is currently more of a tech demo than a usable tool.
summary
OpenDLSS reimplements the 71-block Swin/ViT network from DLSS 5 (build 310.8.0) using Vulkan and PTX kernels for high performance on Ada GPUs. Unlike traditional upscalers, this is a generative neural rendering network that re-renders existing frames to add detail and adjust style. The project includes a WebGPU port for browser-based execution and a demo using the Filament renderer, though it does not provide the model weights themselves.
key points
- The implementation is bit-exact against the original, matching all 75 block boundaries byte-for-byte.
- It uses FP8 (E4M3) activations with FP16 accumulation on tensor cores, with a total weight size of 141 MiB.
- Performance on an RTX 4070 SUPER ranges from 2.8ms at 768x768 to 29.3ms at 4K resolution.
- The network is a U-net of shifted-window transformer blocks with a global ViT at the bottom, operating as a post-process filter rather than an upscaler.
who should read it
Graphics engineers and reverse engineers interested in neural rendering and Vulkan/PTX optimization.
discussion positive
consensus
Commenters are impressed by the technical achievement of bit-exact parity but note that the high performance cost (e.g., 8ms at 1080p) makes it impractical for 60+ FPS gaming on most current hardware.
pushback
Some users express strong dislike for the AI-generated nature of the README, claiming it contains factual errors and 'bugfix storytelling'.
The technical report suggests it does not use the depth buffer: https://research.nvidia.com/labs/adlr/files/DLSS5_Report.pdf
> The inference interface uses the engine-rendered RGB image as a dense, registered observation of visible scene appearance. It provides dense, pixel-aligned evidence for object support, occlusion boundaries, composition, and local material properties; engine motion vectors separately provide temporal correspondence.
> Existing image generative models commonly rely on text embeddings, exemplar images, or spatial control fields such as depth, edges, segmentation, and …
It provides a direct quote from NVIDIA's technical report to clarify that DLSS 5 does not use the depth buffer, correcting a common misconception in the thread.
Generated by an LLM from the article and thread. Check the sources before relying on it.