3zwr1/AMD-NR---OptiScaler Alpha0.3.0

AMDNR-v0.3.0

on GitHub

DLSS 5 Neural Rendering on AMD GPUs, built into OptiScaler: two neural runtimes (danielblnc, lmxxf), model interleave, residual composition, XeSS frame generation up to 6X, FSR Ray Regeneration. We need your support — join the Discord: https://discord.gg/AMDNR · Ko-fi: https://ko-fi.com/3zinr

Release title

AMDNRv0.3.0 — DLSS 5 Neural Rendering on AMD, now with two runtimes

Release description

DLSS 5 Neural Rendering running on AMD GPUs, built into OptiScaler so it works in any Direct3D 12
game OptiScaler already hooks (Direct3D 11 and Vulkan titles through its D3D12 bridge).

Many thanks to TheAutomatic — the DLSS 5 AMD project releases, the HIP toolchain work
(rtc_compile, the module recipes) and the asset layout are what made the second runtime in this
build possible. This release stands on his work as much as on ours.

We need your support

Join the Discord server: https://discord.gg/AMDNR — help, #bug-report, test builds,
and the place where the next build gets decided. Two runtimes, dozens of titles and a handful of
testers: every report with a log, every screenshot, every "works here" makes the next release
better for everyone. available there too if you need it (not in
these archives; the AMD path does not need it).

Read me : English | 中文 | Português

Support the project: https://ko-fi.com/3zinr

Downloads

Install

1. Download AMDNR-v0.3.0.zip and extract it. Copy everything from it into the game folder
   (next to the game's .exe).
2. Pick a runtime (both is fine; the menu lets you switch):
   • lmxxf (RX 9000 only): nothing more to do - LmxxfNrRuntime.dll and LmxxfNrRuntime.pak
     came with step 1.
   • danielblnc (RX 7000 and RX 9000): extract Runtime.zip and copy its files into the game
     folder too.
3. Rename OptiScaler.dll to dxgi.dll. If the game does not launch or the mod does not load,
   try d3d12.dll, winmm.dll, version.dll or dbghelp.dll instead — one name at a time, never
   several copies.
4. Launch the game and press INSERT for the menu. On the first launch that finds both runtimes
   it asks which one to use; the Neural tab changes it later (takes effect on the next start).

Still not working? Report it in #bug-report with the .log files from the game folder
(lmxxf_backend.log when lmxxf is selected — its lmxxf inputs: and lmxxf stats @N: lines
usually say why).

What's new in 0.3.0

• A second neural runtime: lmxxf (HIP, RDNA 4). Selectable beside danielblnc's; the menu
  says what your GPU can run. Its edit is applied one frame late and carried by the motion
  vectors, so the frame never waits for the network (about 32 ms a job at 1080p, 15 ms at the
  720 tier, on an RX 9070 XT).
• LmxxfNrRuntime.pak: the 463-file assets folder is one encrypted, authenticated file
  beside the runtime DLL; nothing is unpacked to disk. The folder layout still works.
• The network's own temporal history (DLSS 5's history path), working under Model
  interleave through a two-frame vector chain, with upstream's history guard and output-side
  smoothing.
• Real Neural passes (2 and 3 launches per frame); the history loop stays a one-pass loop,
  so passes no longer compound into ghosting.
• Edit shaper under Image look: Edit detail, Edit colour, Edge guard, Output smoothing; the
  Appearance filter runs under lmxxf too.
• After Ray Reconstruction: lmxxf now writes its result back into the RR output (new, not yet
  confirmed in a shipping title — try it with RR on and report).
• Toggle hotkey: HOME switches Neural Rendering on and off in-game (both runtimes), with an
  On / Off notice; rebind it beside the Enable checkbox or under Interface > Keybinds.
• Diagnostics for any title: lmxxf_backend.log gets an lmxxf inputs: line (formats,
  sizes, motion scale, depth, mask, exposure) and a stats line every 600 frames (exposure, fed
  brightness, the model's edit, the carried edit, keep, reactive mean, vector health).
• Self-healing: unusable motion vectors, a depth test that refuses everything, or an
  exposure that blacks out the feed are each detected from those stats and worked around for the
  session, with one line in the log saying so.
• Fixed for both runtimes: a title's one-channel reactive mask read as "everything is
  reactive" (GTA V Enhanced showed no effect at all; fixed), the NR-resolution crash (the job in
  flight is abandoned instead of consumed), the NR % slider snapping back, a Reset flag held up
  by the game, the menu's hip_ms readout, the product name (AMD-NR v0.3.0 / OptiScaler v11.0.0), the GPU support readout, Neural passes under NR resolution, Debug view 5 (the
  edit applied directly, an A/B for "it does not work").

Full notes: RELEASE-NOTES.md. Compatibility table for lmxxf: README.md.

Known

• RX 7000, RX9000 (RDNA 3, RDNA 4) runs the danielblnc runtime only; handheld APUs and RX 6000 are not
  supported by either runtime yet.
• The after-RR placement of lmxxf, Vulkan titles with lmxxf, and the TLOU / Where Winds Meet /
  RE Requiem RR reports are still waiting for logs.

Credits

• TheAutomatic — DLSS 5 AMD project — https://github.com/TheAutomatic/dlss-5-amd-project —
  many thanks: the releases, the HIP toolchain and the layout this build's lmxxf integration
  follows.
• DLSS-NR on AMD — danielblnc — https://github.com/danielblnc/DLSS-NR-on-AMD — the runtime and
  weights in Runtime.zip, unmodified.
• lmxxf / dlss5-on-amd-9070xt-porting — https://github.com/lmxxf/dlss5-on-amd-9070xt-porting —
  the HIP neural runtime (MIT) built in-tree as LmxxfNrRuntime.dll.
• Matheus / dlss-5-amd — https://github.com/MatheusGViana/dlss-5-amd-project
• OptiScaler — Overclockers — https://github.com/Overclockers/OptiScaler-Releases

Licence: GPL-3.0 (LICENSE); third-party licences in Licenses\. Not affiliated with NVIDIA, AMD
or any publisher. Use at your own risk.