Skip to content

Releases

v0.8.93 (fork) - neural-rendering settings re-read while running, and host fixes

Choose a tag to compare

@jlrouzies-fr jlrouzies-fr released this 12 Nov 17:00

Fork release: v0.8.92 (fork) plus six commits. Everything else is wilsjo2's work; see their releases for what v0.8.x does.

New: [DlssNr] is re-read from OptiScaler.ini while the game runs

OptiScaler looks at the time stamp of OptiScaler.ini every 200 ms. When it has changed and then stood still for one look, the neural-rendering keys are read again and applied, with no restart: Enabled, Preset, Style, Intensity, LocalStructure, LocalTone, SkinStructure, AutoMask, UnlockPasses, Passes, WorkingScale, TransferStrength, ColourStrength, MaxRatio, WhitePointScale, WhitePointSource, WhitePointTrim, AutoExposureTrim, AutoExposureHighlightProtection, SkinProtection, SkinDetail, SkinColour, EnvironmentDetail, EnvironmentColour.

This is what lets another tool drive the neural pass. DLSS5-Feeder's in-game menu writes the ini and recognises a build that does this by the text DLSS-NR live settings in the DLL. The log says DLSS-NR live settings: read from OptiScaler.ini while running: <key>=<value> for each key that changed. OptiScaler's own menu is unchanged and still applies its changes at once.

Only those keys are re-read. Every other section of the ini is read at start, as before.

Fixes since v0.8.92

  • Neural pass follows the command list's device, not the output texture's. A wrapper layer (ReShade's proxies) can give the output texture one identity for the game's device and the command list another; using the wrong one faulted inside the model (access violation in D3D12Core, or a hang with RunBeforeSR).
  • Finished picture with ReShade loaded through [Plugins] LoadReshade. The Present queue was another wrapper of the game's device, so every Present returned early and nothing was composed. On a device mismatch with the same adapter, the work is submitted on the DLSS-side queue.
  • WinUI 3 hosts: [FrameGen] WrapCompositionSwapChain=false leaves composition swap chains unwrapped (a wrapper handed to a SwapChainPanel crashed Elgato Studio), and [Menu] InputHooks=false runs the menu on polled input with no Win32 detours. Both default to the old behaviour. Neither key is in the shipped OptiScaler.ini: add the line under its section.
  • SkipD3D11FeatureLevelElevation quirk (Metro / Luma), and no conflict with a dummy D3D12 device probe.
  • In-picture menu no longer rebuilt every other frame while the host is in the background. That leaked a descriptor heap 30 times a second until the device refused new ones and the neural pass went off for the session. The menu now only ever attaches to a window of its own process, and a neural pass whose construction failed is tried again every 600 frames.

What was and was not run

  • Live settings: run on Direct3D 12 through DLSS5-Feeder's lab (a 32-bit game with the 64-bit helper, and a 64-bit game with OptiScaler in its own process): on/off, style, intensity, passes, working scale and the composition keys all took effect while running. The native Vulkan path makes the same call but was not run.
  • The command-list device fix was verified in Kameo (through the feeder's helper). The WinUI 3 options are what Elgato Studio runs with.
  • The in-picture menu fix builds but its behaviour in Elgato Studio was not re-checked after the change.
  • Not run: RTX40 MFG. No RTX40-MFG package is provided for this release.

NVIDIA model / FG runtime DLLs are not bundled. Preserve game-specific INI settings when updating.

Download

OptiScaler-NR-v0.8.93.zip, SHA-256 8003065D88813B342049B869E8C894F8771C58F3D1A630DDF42FABD5C9D2F5FC. Extract everything into the game folder, as for any OptiScaler build; DLSS5-Feeder's installer does it for you.

v0.8.92 (fork) - matched guides and motion scale for the NR model

Choose a tag to compare

@jlrouzies-fr jlrouzies-fr released this 24 Sep 16:22

Fork release: wilsjo2's v0.8.91 plus fixes to what the neural model is given, offered upstream as pull request #109. Everything else is wilsjo2's work; see their releases for what v0.8.x does.

Updated 2026-09-26 (third upload): the second upload (2026-09-24) fixed motion below 100% but made 100% flicker in games that upscale with low-resolution vectors (Onimusha). This upload corrects that and ports both fixes to the native Vulkan path. If you downloaded v0.8.92 before, download it again; the SHA-256 below is the current zip.

Fix 1: flicker and slow settling at every Model resolution below 100%

Below 100% the model was given a colour at the working size and depth and motion vectors at the frame's size. Depth and motion are now resampled to the working size (the point resample the DLSS-enlargement path already used for its private upscaler) and handed to the model as a full zero-origin region. Peripheral spatial compression is untouched; it packs its own guides. [DlssNr] MatchGuides=true is the default; false restores the old behaviour.

Fix 2: motion vectors at the wrong length in games that upscale

The model reads its motion-vector scale in pixels of the motion texture it is handed, and the game's scale turns its vectors into pixels of the size they are measured in: the render size for low-resolution vectors. The model path converted with working / frame, which below 100% shrank every vector in a game upscaling with low-resolution vectors. Onimusha: Way of the Sword at DLSS Performance, 4K, 70%: the model got 1344 where the motion was 2688. The scale is now texture the model reads / size the vectors are measured in (the matched working-size guides below 100%, the game's own region at 100%), the DLSS-enlargement path's formula. Games fed at their output size (DLAA, DLSS5-Feeder) were never affected. [DlssNr] RenderMotionScale=true is the default; false restores the old conversion.

The 2026-09-24 upload measured against the working size in both cases, which at 100% handed the model a scale for a frame-size texture while it read the game's render-size one -- twice too large in a 2x upscale -- and 100% flickered where it had always been steady. Fixed here.

Fix 3: the same two fixes on the native Vulkan path

The Vulkan model path had the shape the D3D12 path had before either fix (full-size guides for a smaller colour below 100%, and a working / frame scale). Both fixes are ported behind the same two switches, with the same shader. Compile-only: I have no Vulkan NR game to test on; reports welcome. OptiScaler.log prints DLSS-NR Vulkan model motion scale ... with the texture the scale is for.

OptiScaler.log prints once what the model got: DLSS-NR guides matched to the working size: ... and DLSS-NR model motion scale ....

Verified

  • Fable Anniversary (4K, through DLSS5-Feeder's 32-bit helper): 70% is as steady as 100% with fix 1, and flickers without it.
  • Onimusha: Way of the Sword (native D3D12 DLSS, Performance, 4K, 2 passes, Finished Picture, Transfer=2): 70% very flickery with fix 1 alone, steady with fix 2; after this upload 100% and 85% steady.
  • RTX 5090, driver 617.14. Not run: Vulkan (see fix 3), RTX40 MFG.

Details: docs/RELEASE-v0.8.92.md. Commits: 4a83ccf (fix 1), 4e3b3d7 + 75e2cd5 (fix 2), 57aa64e (Vulkan). Source branch: fix/reduced-scale-guides-0.8.91.

Package

Same layout as v0.8.91 with OptiScaler.dll replaced. Extract the full archive to the game folder (or host64\ for a 32-bit game through DLSS5-Feeder) and rename OptiScaler.dll to winmm.dll as usual. NVIDIA model/FG runtime DLLs are not bundled. No RTX40-MFG variant in this release.

SHA-256 of OptiScaler-NR-v0.8.92.zip: 9605352AF2378EEB5EF96AEDF69C648B36152EEA40A388CB5413CA6A408BF728