Skip to content

INSTALL-DLSSNR.md

Latest commit

 

History

History
216 lines (150 loc) · 17.3 KB

INSTALL-DLSSNR.md

File metadata and controls

216 lines (150 loc) · 17.3 KB
Raw

Install Neural Rendering

This package carries the OptiScaler NR build (wilsjo2 base with the RTX 40 MFG unlock compiled in) plus a Korean menu localization and two optional helpers: an ASI loader and a Streamline runtime fetcher.

NR is experimental and disabled by default. Do not use injection mods in anti-cheat-protected multiplayer games.

Korean guide: INSTALL-KO.md. For the scoped ReShade-first XeFG finished-picture NR setup, see load order and rollback.

The optional RTX 40 MFG unlock needs an unlock-enabled build (this one) and its runtime toggle, which still defaults off.

The historical v11.1 B variant shipped the RTX 20/30 (SM75/SM86) MFG payload for Turing and Ampere cards. The new A-only candidate does not. The section below retains the historical requirements, 1..5 multiplier mapping, 6X caveat and uninstall steps.

Quick start

  1. Close the game and launcher, then back up any existing OptiScaler files and your INI.
  2. Extract the whole package into the folder that holds the real game executable (Crimson Desert: ...\Crimson Desert\bin64). Keep every folder.
  3. Install the payload - pick one method:
    • Proxy (default): run setup_windows.bat and choose a proxy name; [1] dxgi.dll unless another injector already uses it. No loader is needed.
    • ASI (use this when another injector occupies the proxy names, for example XeFG together with ReShade): run Install_AsiLoader_windows.bat and pick [1] (installs the bundled Ultimate ASI Loader as winmm.dll), then run setup_windows.bat and pick [8] OptiScaler.asi.
  4. Start the game and press Insert to open the overlay (Alt+Insert on layouts that remap Insert).
  5. Optional, Neural Rendering: place nvngx_dlssnr.dll beside the game executable, then enable NR in the overlay and start with one pass. See Requirements.
  6. Optional, frame-generation runtime: run Streamline_fetcher_windows.bat to download NVIDIA's official Streamline release into OptiScaler\streamline; do this only when the game does not already provide a suitable runtime.
  7. Optional, RTX 40 (Ada) MFG unlock: enable RTX 40 MFG unlock (restart) under frame-generation settings, save and restart. RTX 40 only, and it needs a supported DLSSG runtime.
  8. Optional, RTX 20/30 MFG unlock: the historical v11.1 B package included a third-party payload. The new A-only release candidate does not bundle it or offer a supported install path. See the release boundary below.
  9. Removal: run Remove_OptiScaler.bat (created by setup_windows.bat), and Install_AsiLoader_windows.bat -> [4] to remove the ASI loader; it restores the file it backed up.

On upgrades, replace the proxy the game loads: adding OptiScaler.dll beside an old dxgi.dll does not update it. Preserve your INI and other mods' loaders.

Requirements

  • A 64-bit game using an OptiScaler D3D12 path, a supported D3D11/Vulkan bridge, or native Vulkan NR.
  • An NVIDIA driver whose installed NGX core supports NR (feature 18).
  • The complete OptiScaler package and a separately supplied nvngx_dlssnr.dll.
  • The NVIDIA Streamline runtime used by frame generation; the bundled Streamline_fetcher_windows.bat installs it into OptiScaler\streamline.

NR runs inside OptiScaler. No NR helper DLL is required; remove the obsolete nvngx.dll_dlssnr.dll when upgrading. Keep the package's ordinary backend dependencies.

Runtime identification

These are the 310.8 variants used during development, not a verified GPU support matrix. The installed driver must accept the runtime.

Variant Intended GPUs SHA-256
Original NVIDIA-signed RTX 50 E16BCF15E16E13F527491CDF7845B2FE6521A738D8F7C9C721866A8496E1FC8E
ShortFuse compatibility RTX 20/30/40; retains RTX 50 path E67DEE209320CDAFE0E93E45675D7AA34323A53ACC57A72B2E40A181581C989A

The compatibility runtime is available in ShortFuse's RenoDX thread. Modification invalidates its NVIDIA signature: verify the hash and keep security protection enabled. Its RTX 20/30 path is substantially heavier. The RenoDX add-on itself is not needed; two NR injectors can conflict.

Get-FileHash .\nvngx_dlssnr.dll -Algorithm SHA256

ASI installation in detail

The ASI method needs an ASI loader; without one an OptiScaler.asi file is never loaded. The package carries the MIT-licensed Ultimate ASI Loader v9.7.4 under tools\asi-loader\.

  1. Run Install_AsiLoader_windows.bat and pick [1] to place the loader as winmm.dll (or [2] for another name the game loads).
  2. Run setup_windows.bat and pick [8] OptiScaler.asi.
  3. The helper backs up any file it replaces and can undo itself with [4].

Use this when another injector already occupies the default proxy names, or when the game runs XeFG together with ReShade.

Known game layouts

Game Executable directory Tested/configured route
Crimson Desert bin64 ASI: OptiScaler.asi via winmm.dll (Ultimate ASI Loader)
Baldur's Gate 3 bin DX11: Dx11Upscaler=dlss_12; native Vulkan: VulkanUpscaler=dlss (gameplay unverified)
Hogwarts Legacy Phoenix/Binaries/Win64 dxgi.dll
Cyberpunk 2077 bin/x64 dbghelp.dll; existing loaders may need chaining

Keep [ProcessFilter] TargetProcessName=auto for portable configurations. A different executable name intentionally disables injection, including the menu.

[DlssNr]
Enabled=true
RunBeforeSR=true
Passes=1
WorkingScale=1.0

For native Vulkan, enable NR in the INI before launch so device/swapchain support is prepared.

RTX 20/30 (SM75/SM86) MFG unlock (optional)

New A-only candidate: the sdli1995/dlssg_for_sm86 B payload, NVIDIA runtime and derived SM75 kernels are not bundled. No redistribution rights have been established for this payload, so don't copy it from an older ZIP or sideload it as a workaround. The code and settings below describe the historical v11.1 B variant, not an installation step for the new candidate. The old v11.1 B guide said "ships with this build" and "no manual sideload"; neither claim describes the A-only candidate. The A-only package has no OptiScaler\dlssg_sm86\dlssg_sm86.dll.

Requirements

  • RTX 20 or RTX 30 only (Turing SM75, Ampere SM86). On RTX 40 use the RTX 40 MFG unlock instead; the two unlockers must never be active together.
  • Driver R580 or newer recommended. Older drivers aren't refused, but R580+ is the baseline the payload's author reports.
  • Historical v11.1 B only: that ZIP carried OptiScaler\dlssg_sm86\dlssg_sm86.dll, its companion dlssg_sm86.ini and THIRD_PARTY_NOTICES.txt. These files are absent from the new A-only candidate.
  • A game whose frame generation runs through NVIDIA Streamline. The payload hands its generated frames to the game's own DLSS-G plugin, so the game's FG menu picks the multiplier.
  • The unlock is off by default, and changing it needs a restart.

Enable

In the menu, under frame-generation settings, open the RTX 20 / 30 (SM75 / SM86) MFG unlock section:

  1. Historical B only: tick Enable SM75/SM86 MFG (experimental; restart). External FG turns on with it, and the game's own FG menu then selects the multiplier. This does not install a payload in the A-only candidate.
  2. Set Max Generated Frames (1 to 5) if you want more than the default 4X.
  3. Pick a Kernel Image only if Auto misbehaves: Auto resolves the format from your card, PTX is the safe choice on RTX 20, and Cubin needs an exact physical match on Windows.
  4. Optional, RTX 30 only: Hardware Bilinear trades exact output for about 2 to 4 percent lower GPU latency.
  5. Save Settings and restart the game. The status line in that section reports the loader's state (Disabled, Ineligible, Conflict, Loaded and so on), with the reason underneath it.

Or set the keys in the INI before launch:

[DLSSG]
AmpereMfgUnlock=true
AmpereMfgMaxFrames=3
AmpereMfgKernelImage=auto
AmpereMfgHardwareBilinear=false
Key Values Default Meaning
AmpereMfgUnlock true/false false Turns the 20/30 unlock on. Read once at startup.
AmpereMfgMaxFrames 1 to 5 3 Advertised maximum multiplier: 1 = 2X, 2 = 3X, 3 = 4X, 4 = 5X, 5 = 6X. The game still picks the actual multiplier from its own FG menu.
AmpereMfgKernelImage auto, PTX, Cubin auto Kernel format. Auto resolves it from the physical GPU.
AmpereMfgHardwareBilinear true/false false RTX 30 (SM86) only: approximate hardware bilinear sampling, about 2 to 4 percent lower GPU latency.

6X caveat: 5 (6X) works only when the game ships a Streamline frame-generation plugin 2.11.1 or newer. With an older plugin the payload clamps the request (its log records limit_clamped) and you get the highest multiplier that plugin supports. The game also has to ask for that many frames itself.

Every setting here is read once, at startup. Save Settings and restart the game; nothing in this section applies live.

Verification gap

No RTX 20/30 hardware was available for this build. This project's build and test host carries an RTX 4090, so the 20/30 path was verified at the loader, configuration, packaging and arming-order level only. How a real RTX 20 or RTX 30 card behaves with it is externally reported (field reports from the payload's users), not measured here. Treat the multiplier and performance behaviour on 20/30 as unverified.

Disable and remove

  1. Set [DLSSG] AmpereMfgUnlock=false (or clear the checkbox), Save Settings and restart.
  2. Turn External FG off as well if you don't want it any more. The unlock enables it, and it keeps its stored value after the unlock is disabled.
  3. Delete OptiScaler\dlssg_sm86\ to remove the module, its companion INI and its logs\ folder. Nothing else is added, and no registry key is written.
  4. Remove_OptiScaler.bat removes the whole package as usual.

Attribution

The historical v11.1 B bundled payload was sdli1995/dlssg_for_sm86 v0.3.5 (pinned commit 9621db5), shipped unmodified from https://github.com/sdli1995/dlssg_for_sm86

NVIDIA's runtime and model components and the SM75 kernel family inside that payload carry their own terms. The historical B package carried upstream notices as OptiScaler\dlssg_sm86\THIRD_PARTY_NOTICES.txt, plus a copy under Licenses\. Upstream states GPLv3 in prose but the repository carries no LICENSE file; that licensing gap is recorded here instead of being hidden. The payload is self-signed (CN=DLSSG for SM86), so SmartScreen shows an unknown publisher.

Placement and resolution

Setting Behaviour
Generate model before upscale Edit the active input before SR or RR+SR. Off runs NR afterward.
Generate before upscale, apply after upscale Upscale NR's edit separately; the game reconstructs its clean input.
Apply NR to the finished picture Apply after effects/HUD. With early generation enabled, carry the edit to presentation.

Model resolution is relative to the image at the selected stage. At 1440p input/4K output, 100% means 1440p before SR and 4K afterward. Lower percentages reduce NR work without changing the game's upscaler preset. See pipeline controls and enlargement.

Below 100% model resolution, the Enlargement setting picks how the model's answer is enlarged. Lighting + colour (Transfer=3) resizes lighting gain and colour changes separately with bilinear reconstruction, on DX12 and native Vulkan. Lighting + colour + DLSS (Transfer=4) enlarges the same fields through private DLSS SR and needs post-upscale DX12 processing, including the existing DX12 bridges; it is not available before the game upscaler or in native Vulkan processing. The default stays Matched residual (Transfer=1); Classic, Matched residual and Matched residual + DLSS keep their existing algorithms, and native/supersampled model processing bypasses the new reconstruction. The v0.8.8 release note carries the upstream validation detail.

Finished-picture NR supports D3D12, its D3D11 bridge and native Vulkan with supported SDR/HDR10/scRGB formats. It can alter HUD/menus. Early generation plus finished-picture application requires D3D12 or the D3D11 bridge; Vulkan cannot carry that private edit. Unsupported frames retain the game image. See bridges.

SR and RR share pass count, profiles and model resolution. Separate-edit placement stays manual. Keep genuine nvngx_dlssd.dll (RR) separate from nvngx_dlssnr.dll (NR). If Cyberpunk's RR option is unavailable with a d3d12.dll proxy, a reported workaround is dxgi.dll; check existing loaders first.

Turning off Apply model hides the edit while NR still runs; the master switch stops NR. GPU timings measure elapsed work and can overlap other work; compare total frame time for performance. A private upscaler adds cost beyond the NR model timer.

NR under XeFG frame generation

XeFG placement is its own route. The config-only path below needs no new build and keeps the deferred carrier working while frame generation is on. Frame generation stays ON throughout. The finished-picture handoff under XeFG is covered by the bridges note: this build ships the owned application-frame handoff, so the finished-picture path is available under XeFG. In-game verification is still pending.

Setting Value Where
Enlargement Matched residual (Transfer=1) Prepare NR input
Model resolution 75% (WorkingScale=0.75) Prepare NR input
Generate before upscale, apply after upscale ON NR pipeline
Apply NR to the finished picture OFF for the config-only path; the new handoff ships in this build, so the finished-picture path is available under XeFG NR pipeline
Private NR upscaler DLSS (PrivateUpscaler=0) NR pipeline
Apply model ON NR pipeline
Compare / Debug view / skin-mask preview OFF Inspect NR
HdrTransfer auto (false) INI

Transfer=1 drops the DLSS-carried enlargement (composed by the shader instead); WorkingScale=100% is the alternative that keeps Transfer=2 at ~+78% model pixels.

With [Log] LogToFile=true and LogLevel=2, the placement reports itself:

  • Deferred carrier: DLSS-NR deferred upscale: running: <dimensions> contribution -> private DLSS RR -> <dimensions>; applied after SR
  • Finished-picture handoff: DLSS-NR finished picture: <N> frames, <W>x<H>, OptiScaler FG true, same producer queue true, game-frame handoff true
  • Cold start on the XeFG path: NR_XEFG_ROUTE streamline_registered=0 source=app_proxy queue_source=xefg_application format=<DXGI_FORMAT> colorspace=<DXGI_COLOR_SPACE>
  • After the original proxy Present: NR_XEFG_PRESENT generation=<G> frame=<F> enabled=1 framegen_result=0 frames_presented=6

Three caveats:

  • Keep frame generation on: no FG-OFF workaround exists.
  • The bundled XeFG SDK supports HDR10 (R10G10B10A2_UNORM) only: no FP16/scRGB.
  • frames_presented counts submitted pictures, not verified scanout.

These placement paths are not verified in-game on this machine: it has no RTX 20/30 hardware and runs no XeFG session. The owner runs the in-game acceptance. For a scoped ReShade-first setup, see load order and byte-exact rollback; this R2 candidate has not been game-verified.

Files in this package

File Purpose
OptiScaler.dll The payload; setup_windows.bat renames it to the proxy or ASI name you pick.
OptiScaler.ini Settings. Keep it when upgrading.
setup_windows.bat, setup_linux.sh Installer and uninstaller generator.
Install_AsiLoader_windows.bat Installs or removes the bundled Ultimate ASI Loader (MIT).
Streamline_fetcher_windows.bat Downloads NVIDIA's official Streamline runtime when the game lacks it.
tools\ Scripts used by the two helpers; tools\asi-loader\ carries the loader and its license.
docs\, Licenses\ Detailed notes and third-party licenses.

Troubleshooting

Enable [Log] LogToFile=true and LogLevel=2, then enter a rendered scene.

  • No log/menu: check executable directory, loaded proxy, process filter, quarantine and loader conflicts. For ASI installs confirm the loader is present (Install_AsiLoader_windows.bat -> [3]). Try Alt+Insert for alternate keyboard layouts.
  • Menu opens but ignores input: try [Hotfix] ManualInputPolling=true and disable conflicting overlays.
  • Model initialization fails: check runtime hash and driver support; include the exact log error in a report.
  • Unexpected model size: inspect active/target/model dimensions and fallback messages; padded input explains supported rectangles.
  • XeFG together with ReShade: the proxy method can collide with other injectors. The ASI method alone doesn't establish load order. Follow the scoped ReShade-first procedure, or stop if its preconditions don't hold.

Follow upstream OptiFG guidance for frame generation. NR success alone does not establish FG compatibility. See tested games and remaining issues.