Case study

DLSS 5 Neural Rendering: reconstructing a runtime model from DLL analysis

DLSS 5 Neural Rendering reconstructs parts of NVIDIA's model from static analysis of nvngx_dlssnr.dll 310.8.0.0 shipped with NBA 2K27. It documents a 153-record weight container, a 71-block Swin/ViT structure and numerical references for selected operators. The author reports byte-for-byte agreement for one FFN operator across 24 comparisons.

Targets: Binary
Category: Independent reconstruction cases
Relationship: Independent reconstruction project
Sources reviewed: 2026-11-09

Copy as Markdown View JSON Cite

Requirements & limits

Prerequisites: The readers require the user's own copy of nvngx_dlssnr.dll 310.8.0.0 and its extracted WEIGHTS_HT resource.
Prerequisites: Parsing, references and tests use PyTorch 2.14.0+cpu; native comparison requires a Blackwell GPU.
Limits: A full forward pass and most operator comparisons remain unfinished.
License: No licence stated on the project page.

Cite

BibTeX

@misc{dlss5-nr-reverse-engineering-case,
  title = {{DLSS 5 Neural Rendering: reconstructing a runtime model from DLL analysis}},
  howpublished = {\url{https://huggingface.co/inarikami/dlss5-nr-reverse-engineering}},
  url = {https://huggingface.co/inarikami/dlss5-nr-reverse-engineering},
  urldate = {2026-11-09},
  note = {Awesome REA entry: \url{https://awesomerea.org/resources/dlss5-nr-reverse-engineering-case/}}
}

Plain text

DLSS 5 Neural Rendering: reconstructing a runtime model from DLL analysis. https://huggingface.co/inarikami/dlss5-nr-reverse-engineering (accessed 2026-11-09). Awesome REA entry: https://awesomerea.org/resources/dlss5-nr-reverse-engineering-case/

Useful? Star Awesome REA on GitHub2.4k

About this entry

The description was reviewed against primary sources.

Report an outdated entry Suggest a translation fix

Primary sources

https://huggingface.co/inarikami/dlss5-nr-reverse-engineering/blob/2f3db2562d18f750169c85ed947dc15bffca5a3b/README.md