Skip to content
QIO journal

Technology4 min read

DLSS 5 in 1993's DOOM: how a neural network smoothed pixels and kept the style

Enthusiast niko-zvt integrated NVIDIA's DLSS 5 frame reconstruction into the original 1993 DOOM. The id Tech 1 engine could not provide the motion vectors and scene depth a neural network needs, so they had to be added. The result: sprite graphics became smoother yet stayed recognisable. The code is open on GitHub as DOOM-DLSS-5.

DLSS 5 in 1993's DOOM: how a neural network smoothed pixels and kept the style

What DLSS 5 is and why DOOM is alien to it

DLSS reconstructs frames with a neural network: the game renders at low resolution and the network fills in detail using previous frames, object motion vectors and scene depth. Modern engines supply that data natively. DOOM from 1993 rasterises on the CPU, works with sprites and knows neither motion vectors nor a depth buffer in the usual sense.

I like the approach itself: do not redo the assets, add the missing data and let a modern tool work. We update old websites the same way: not rewriting everything, but adding what search engines and users lack.

What had to be built

The author had to adapt the transfer of three things absent from the id Tech 1 architecture: per-pixel motion vectors, scene depth parameters and render buffers in a format DLSS libraries understand. In effect it is a bridge between a thirty-year-old software renderer and a modern upscaling pipeline. The project is experimental: its goal is to test how neural upscaling copes with legacy rasterisation and sprite graphics.

A row of large cubes turning into fine grains from left to right
Large pixels are smoothed while silhouettes and contrast remain.

What the picture looks like

Pixel staircases on wall edges and sprites are smoothed and motion is more fluid, while the characteristic palette and chunky monster shapes survive: the network fills transitions rather than repainting the style. That is what separates the project from 'HD texture packs': it changes how the game is shown, not its assets.

Why it matters and where to get it

Practical value for players is small, but for developers it is an open test bench: the full DOOM-DLSS-5 source is on GitHub, so anyone can study the implementation, run benchmarks on compatible GPUs and contribute changes. Projects like this show a technology's limits better than marketing demos.

Summary

DLSS 5 in DOOM proves that neural upscaling can be bolted onto even a sprite renderer once it gets motion vectors and depth. The game's style stayed intact and the code is open to everyone.

Sources

  1. DLSS 5 smoothed the original DOOM's pixel graphics (GameGPU)
  2. NVIDIA: DLSS technology
  3. GitHub
  • DLSS
  • DOOM 1993
  • id Tech 1
  • NVIDIA
  • GitHub

Follow the journal

New pieces on websites, SEO and AI come out in the QIO journal. Follow in Google, by RSS or in Telegram to get them first.

Read next

How can we help?
Discuss a project