OptiScaler-DLSSNR-PreSR-Multipass: Neural Rendering as a Game Mod

What this mod actually changes in a running game

NVIDIA's Neural Rendering is normally something a game developer integrates. This project takes the opposite route: it injects the effect into an existing DirectX 12 title through the OptiScaler modding layer. The README describes the result plainly as "a game mod that uses NVIDIA AI to change lighting, detail and colour," with controls for how strong the effect is and what it costs in frame time.

The audience is narrow by design. You need an NVIDIA RTX 20, 30, 40 or 50-series GPU and a supported 64-bit game. The README warns that older cards "can be much slower," which is a polite way of saying the effect is not free on Turing. This is for people who already run DLSS in a game, are comfortable editing files beside an executable, and want to see what a neural pass does to the image before and after upscaling. It is not a general-purpose graphics improvement, and the repository itself labels it "an experimental community version of OptiScaler" where "results and game support vary."

Pass order, skin versus scenery, and the multipass design

The interesting engineering here is the ordering. The mod can run the Neural Rendering effect before or after the game's upscaling, and the README notes that applying it to the finished picture "help[s] with green noise" and works with frame generation on or off in native DirectX 12 games, including HDR. That post-upscale placement is the more forgiving option: the effect sees a stable, final-resolution image rather than the low-resolution buffer that DLSS is about to consume.

The multipass feature is where the name comes from. You can apply the effect more than once with different settings each time, which lets you separate, for example, a subtle skin treatment from a heavier scenery treatment. The README states that "extra passes cost more performance," and that is the central trade-off of the whole project. There is no free lunch in the architecture: each pass is another neural evaluation over the frame.

A resolution control exists as the escape valve. Lowering the model resolution reduces the performance cost, at the price of detail in the effect itself. Combined with the separate skin and scenery settings, this gives you four dials that interact: pass count, model resolution, per-pass strength, and placement relative to upscaling. The README does not document a recommended starting combination beyond "start with one pass."

Installing it on Windows and getting one pass running

The install is manual file placement, not a package manager. The README's steps assume you have closed the game and backed up existing mod files. Extract the release ZIP beside the game's executable, then add the NVIDIA model file to the same folder. The model file is not bundled: you download nvngx_dlssnr.dll separately, and the choice depends on your GPU. The setup guide at INSTALL-DLSSNR.md#choose-the-correct-runtime explains which runtime to pick and how to check it.

Once the files are in place, the batch script handles configuration:

setup_windows.bat

When it asks, choose NVIDIA. Then launch the game, select DLSS in the game's own settings, and press Insert to open the OptiScaler overlay. Enable Neural Rendering there and start with a single pass, as the README advises. The overlay is where the strength, lighting, detail and colour controls live.

If something goes wrong, the README asks for an issue report containing your game, GPU, settings and the OptiScaler.log file. That log is the only diagnostic surface documented. There is also a setup_linux.sh in the repository root, but the README's install section is Windows-only and does not describe a Linux flow, so treat that script as undocumented from the README's perspective.

Two operational notes from the README matter before you start. Hybrid mode is for RTX 50 GPUs, its files are included, and loading may pause the game in a way that "look[s] like a freeze" while you wait. And anti-cheat-protected multiplayer games should be avoided outright.

Flicker, Ray Reconstruction, and the cases where this is the wrong tool

The README is unusually direct about failure modes: Neural Rendering "costs performance and can cause flicker or other visual problems." Flicker is not a cosmetic footnote. A temporal effect layered on top of DLSS or frame generation can fight the game's own accumulation, and the README singles out the pre-Ray-Reconstruction placement as "still experimental." If your game uses Ray Reconstruction, the ordering question becomes much harder to answer empirically.

The clearest wrong-tool case is multiplayer. The README says to avoid anti-cheat-protected games, and that is a hard boundary rather than a tuning preference. Beyond that, anyone on an AMD or Intel GPU is outside the supported set entirely, because the NVIDIA model file is a required component. And a game that does not expose DLSS as an upscaler has no documented entry point here, since the install flow depends on selecting DLSS first.

There is also a maintenance cost that the README acknowledges indirectly. It is a pre-release stream: v0.8.4 is tagged "NR memory fix and pass unlock (pre-release)," v0.8.3 is "NR loader compatibility and recovery," and v0.8.1-nr-direct-runtime is a separate "Direct NR runtime, no helper DLL" variant. Three releases in three days, one of them a recovery fix, tells you the loader path is still moving. Pinning a known-good build and keeping a copy is more sensible than tracking latest.

How it differs from stock OptiScaler and from RenoDX

Stock OptiScaler is an upscaler replacement layer: it lets a game that ships DLSS use FSR or XeSS instead, and it is widely used for exactly that. This fork keeps that foundation and bolts Neural Rendering on top of it. The credits list the lineage explicitly: it is built on OptiScaler and on Dagherbou's Neural Rendering fork, with colour processing from RenoDX. So the practical difference from upstream OptiScaler is that you get neural lighting and colour passes, plus the multipass and pre/post-upscale ordering, in exchange for a pre-release code path and an extra model file you must source yourself.

Against RenoDX, the split is philosophical. RenoDX targets colour and tone mapping through shader-level processing, and this project borrows that colour work rather than replacing it. The Neural Rendering passes are a different mechanism: a model evaluated over the frame, with a resolution knob and a real frame-time cost, rather than a shader rewrite. If your goal is purely to fix a washed-out or clipped image, a colour-focused tool is the smaller intervention. If you want the neural effect itself, and you accept the flicker risk, this is the fork that exposes it.

Licence, upgrade cost, and what the repository does not tell you

The project is GPL-3.0. That is a copyleft licence, and it applies to the mod's own code; the NVIDIA model file you download separately is not part of this repository and carries its own terms, which the README does not restate. If you plan to redistribute a bundled package rather than point users at the releases page, the interaction between GPL-3.0 and the NVIDIA runtime is something to check with someone qualified. Nothing here is legal advice.

Upgrade cost is real but bounded. Because installation is file placement beside the executable, upgrading means replacing the extracted files, and the README's first step is to back up existing mod files. The pre-release cadence means a new build can change loader behaviour; v0.8.3's description mentions "compatibility and recovery," which suggests users have needed a way back. Keep the previous ZIP.

What the README does not cover is equally worth noting. There is no documented rollback procedure beyond deleting files, no list of confirmed-working games, no benchmark table, and no guidance on which model resolution to start from. The setup guide is the place those details would live, and it is referenced rather than reproduced. Changelog.md, Config.md and Features.md exist in the repository root for readers who want the fuller picture.

Editorial conclusion

Adopt it if you have an RTX 20-series or newer card, a single-player DirectX 12 game, and patience for flicker and per-game tuning; the repository's own README calls the results variable and the before-Ray-Reconstruction path experimental. Skip it for anti-cheat-protected multiplayer games, for non-NVIDIA GPUs, and for any title where you are not willing to restore a backup. Verify first that you have the correct nvngx_dlssnr.dll for your GPU, that the game runs DLSS, and that you can roll back by deleting the extracted files from the game folder.

Frequently asked questions

Official sources

1. Issues
2. License: GPL-3.0
3. README
4. Releases
5. wilsjo2/OptiScaler-DLSSNR-PreSR-Multipass on GitHub