OpenDLSS: A Vulkan Reimplementation of Nvidia's DLSS 5 Neural Rendering Network(github.com)
267 points by sagacity 14 days ago | 122 comments
tl;dr: OpenDLSS is a Vulkan reimplementation of NVIDIA's DLSS 5 Neural Rendering network (a 71-block Swin/ViT model running FP8 on tensor cores), bit-exact against the original down to all 75 block boundaries. It includes a WebGPU port that achieves the same exact outputs in-browser without tensor cores or FP8, demonstrating that the determinism lives in the spec rather than the hardware. Users must supply their own weights; the repo ships no NVIDIA code, weights, or instructions for obtaining them, and benchmarks on an RTX 4070 SUPER show 7.8 ms at 1080p.
HN Discussion:
  • ↑Impressed by the technical achievement of bit-exact reproduction across hardware
  • ↓Questions the practical utility given weights aren't provided and may be game-specific
  • ~Concerned about performance cost at 1080p being too high for real-world use
  • •Curious about technical details like inputs (z-buffer) and training data
  • ↓Skeptical due to LLM-generated README suggesting lack of author effort