Multipass neural denoising is the real contribution of this OptiScaler fork

DLSS Neural Rendering is a real-time AI denoising and image-enhancement feature NVIDIA ships as a separate model file, and this community fork of OptiScaler wires it into games that wouldn't otherwise expose it. The core idea is straightforward: intercept the game's rendering output, run the neural model over it, and hand back a cleaner, differently-lit frame. What makes this fork distinct from the upstream project is the multipass option — you can run the effect more than once per frame, each pass with its own settings, at the cost of proportionally more GPU time.

The controls are more granular than you'd expect from a mod. Strength, lighting, detail, and colour are all independently adjustable, and there are separate sliders for skin tones versus scenery — a distinction that matters because neural denoisers trained on general content often handle faces poorly. You can also choose whether the effect runs before or after the game's own upscaling step, which changes what the model actually sees and therefore what it can fix.

The hardware floor is an RTX 20-series card, but the README is honest that older cards in that range can be significantly slower. RTX 50-series owners get an additional hybrid mode, though the README warns that loading it can pause the game long enough to look like a crash. That's not a dealbreaker, but it's the kind of friction that reminds you this is experimental community software, not a polished driver feature.

A few practical constraints are worth naming. Anti-cheat-protected multiplayer games are explicitly off-limits — running this in a competitive title is asking for a ban. Using Neural Rendering before Ray Reconstruction is flagged as still experimental, meaning the interaction between two AI-driven rendering stages isn't fully characterized yet. Green noise suppression on the final composited frame is listed as a specific use case, which suggests the model's output isn't always clean on the first pass.

The project is built on top of OptiScaler and a prior Neural Rendering fork by Dagherbou, with colour processing pulled from RenoDX. It's a genuine layer of community engineering rather than a thin wrapper, and the multipass architecture in particular isn't something the upstream projects offer. Whether the performance cost pencils out depends entirely on your GPU and how much you care about the visual difference — the README doesn't promise a specific frame-time budget, and results vary by game.