OptiScaler DLSS 5 Mod on RTX 40 GPUs: Performance Gains, Multi Frame Generation and Risks

The setup: DLSS 5 outside its official lane​

MFG is officially available only on RTX 50-series GPUs, but OptiScaler makes it available on other GPUs as well, including his RTX 4070 Ti. Singh's account says the official route to DLSS 5 is an RTX 50-series card and a short list of supported games. The community has been feeding NVIDIA's neural rendering files through tools like OptiScaler and ReShade to push the tech into other games.

OptiScaler works with games that already support DLSS Super Resolution. Singh says the presets, model choices and rendering controls turned DLSS 5 from a single take-it-or-leave-it visual makeover into something he could tune.

DLSS 5 is a separate feature from upscaling and from Frame Generation. NVIDIA describes it as a real-time model that takes a game's color and motion-vector data and adds lighting and material effects. Its job is to change how the image looks, not to upscale it.

This is also not a one-off. Tom's Hardware and VideoCardz covered another community project, "DLSS Unlocked," which combines OptiScaler_DLSSNR with DLSS Enabler to bring DLSS 5 and DLSS MFG to all RTX GPUs. Reporting on that tool says it needs DirectX 12 games with native DLSS upscaling and frame generation support. It is also not NVIDIA's native RTX 50 MFG running on older silicon. DLSS Enabler redirects the frame-generation path through a modified AMD FidelityFX FSR 3.1 implementation on RTX 20 and 30 cards, and for parts of the RTX 40 implementation.

The performance tax and the trick that trims it​

Singh says that running even one neural rendering layer on top of a game more than halves his frame rate. The fork documentation backs that up. It lists the unmodified OptiScaler DLSS 5 pass at roughly 13 ms per frame on an RTX 4070.

The fork's central trick is to run the neural model before DLSS Super Resolution. The model then works on the lower-resolution render, and DLSS upscales the result afterward. The project says this means about three times fewer pixels to process. Singh says the output stays close to the full-resolution version.

The fork adds a second technique, an "edit cache." The model runs only every N frames. What it changed is stored as a per-pixel ratio, never as the image itself. That change is reprojected with motion vectors, validated against depth and color, and applied to the game's fresh frame.

The numbers, and what they don't prove​

These are the fork author's own figures, not independent benchmarks. They come from the project's release notes for version 1.2.0, released October 10, 2026. The test was The Last of Us Part I on an RTX 4070 at 2560×1440, using the built-in benchmark. It counts only frames the game renders, with frame generation excluded.

An earlier test in Control Resonant, also on an RTX 4070 at 1440p with DLSS Balanced, showed the same pattern. The off baseline was 62.2 FPS. The unmodified setup managed 34.4 FPS, running the model before the upscaler every frame managed 46.7, and running it every other frame managed 55.3.

Two caveats matter here:

• Even the fastest x8 result in the table stays below the 83.9 FPS you get with DLSS 5 off. It reduces the cost but doesn't remove it.
• The author compared the fork with two other community DLSS 5 builds. He reports a 14% lead over the fastest of them in The Last of Us Part I and 5% in Control. GamePGU's coverage repeats those claims. Neither is a broad third-party benchmark.

Image quality: the trade-off behind the speed​

The x8 preset runs the model every eighth frame, before the upscaler, at half the render resolution. According to the release notes, it switches off the anti-ghosting fingerprint, same-surface fill and light pop-in smoothing. Those are the features meant to stop stale edits from trailing behind moving objects.

The fastest mode is therefore not necessarily the steadiest. Version 1.2.1, which came out later the same day, fixed a bug in the x8 preset where the whole picture's brightness jumped slightly each time the model ran. The author says the fix costs about 2% FPS.

The fork also offers the knobs Singh praises:

• Model styles: Default, Natural and Cinematic. The author describes Natural as adding more micro-detail and darkening a little, and Cinematic as smoothing the image. Those are the author's descriptions, not an objective ranking.
• Multi-pass: one to three model passes.
• Strong preset: a two-pass look that works after the upscaler and with Ray Reconstruction.
• Live comparison: F6 cycles between optimized, vanilla and off, and shows the real rendered FPS.
• Built-in benchmark: it runs the modes back to back and writes an HTML report.

Singh says each style comes with models A, B and C. I couldn't corroborate that detail in the project documentation I reviewed, so treat it as his description of his own setup.

His broader point is subjective: adjustable neural rendering can look good where the default makeover bothered him. Whether it looks better is a matter of taste and of the game, scene and settings. He is explicit that his concerns about DLSS 5 overriding artistic intent haven't gone away.

The MFG piece is a separate project​

Singh credits OptiScaler with unlocking MFG on his RTX 4070 Ti. A separate community repository, forked from Dagherbou's OptiScaler_DLSSNR, documents how that works. NVIDIA gates nvngx_dlssg.dll to RTX 50 for anything above 2x. With the option enabled, OptiScaler rewrites two Blackwell architecture checks in the mapped image only, so the game's own Frame Generation can offer 3x, 4x and 6x on an RTX 40.

The author reports testing on an RTX 4070 Ti in Crimson Desert. The game's menu exposed 3x, 4x and 6x, and 6x delivered about 240 fps from about 60 real frames. That figure belongs to that fork and that game. It is not evidence for the DLSS 5 benchmark numbers above.

Frame generation also changes what a high FPS number means. A displayed 6x output from a 60 FPS base still carries the input latency of that 60 FPS base, and higher multipliers can add artifacts. The DLSS 5 fork warns that an in-game counter can mislead: it multiplies the game's rendered frames and then stops at the refresh rate. Its author recommends pressing F6 to see the real rendered number.

One known issue is documented. If a game crashes with a GPU timeout while a ReShade add-on forces multi-frame generation on an RTX 40-series card ("MFG Unlock"), try disabling that add-on first.

How installation works, and what to watch for​

The Skynizz DLSS 5 fork documents a simple Windows install:

1. Download the versioned Windows ZIP from the project's Releases page. At the time of writing, that is OptiScaler-DLSS5-1.2.1-win-x64.zip.
2. Unzip it and run INSTALL.bat.
3. Drop the game folder, the one containing the executable, onto the installer.
4. Launch the game, enable DLSS, press Insert to open the OptiScaler menu, then use the DLSS 5 tab.

The installer chooses a free filename rather than overwriting an existing file, according to the project. UNINSTALL.bat reverses the changes. Other limits apply:

• No NVIDIA model included: the fork does not distribute nvngx_dlssnr.dll. You must supply it yourself. The project says the installer searches the game folder and %LOCALAPPDATA%\RHI\DLSS-NR.
• DirectX 12 and DLSS Super Resolution only: GamePGU's coverage says the fork works only in DirectX 12 projects that support DLSS Super Resolution.
• Unofficial status: the project says it is not affiliated with NVIDIA and relies on an undocumented driver feature.
• Avoid multiplayer: the MFG project warns against using it in multiplayer games because it modifies NVIDIA code inside the game process. Anti-cheat systems may object, though that specific risk is my inference.

Sensible practice, from general modding experience rather than the project docs, is to stick to single-player games and back up your game folder. Compare presets with the built-in benchmark or the F6 live comparison. Remove the mod if the game misbehaves.

Bottom line​

The evidence supports a narrow conclusion. In the author's tests on an RTX 4070, running the model before the upscaler, caching its edits and lowering its cadence sharply reduce the measured DLSS 5 cost. Singh's experience on an RTX 4070 Ti is consistent with that. Visual quality, game compatibility, artifacts and any MFG benefit remain dependent on configuration and title. Nobody has independently replicated these numbers at scale. If you are curious and comfortable with DLL-level tinkering, it's an interesting experiment. It isn't a supported path to a free GPU upgrade.