GitHub release — AMDNR v0.3.3

Title: AMDNR v0.3.3 — lmxxf on RX 7000 (RDNA 3), RenoDX colour, lmxxf Full network + RAM fix, XeFG up to 10X, Vulkan fixes

Tag: v0.3.3

DLSS 5 Neural Rendering on AMD, built into OptiScaler.

Join the Discord — support, bug reports, test builds: https://discord.gg/AMDNR
(new server — the old invite still works)
Ko-fi: https://ko-fi.com/3zinr

Downloads

Install: extract the zip into the game folder, rename OptiScaler.dll to dxgi.dll (or winmm.dll /
version.dll / OptiScaler.asi with an ASI loader), INSERT for the menu. Replace LmxxfNrRuntime.dll
and LmxxfNrRuntime.pak too - they ship together with this OptiScaler.

What's new

• New Discord server: https://discord.gg/AMDNR — the menu's Discord button opens it; the old invite still works.
• danielblnc 0.3.2 is supported (new, not yet tested in a game): AMDNR now drives danielblnc's 0.3.2
  runtime as well as 0.3.1, so the 0.3.1 Runtime.zip keeps working unchanged. In 0.3.2 only the GPU
  kernels changed; the host side is the same apart from one entry point, and the existing weights file works.
• lmxxf now runs on RX 7000 (RDNA 3) and Strix Halo — our own RDNA 3 backend, by 3zwr1. Slower than on
  RX 9000: start with NR resolution at 67%.
• Colour composition: RenoDX (experimental) on both runtimes, under Neural > Image look — RenoDX's
  colour handling after the model, with a Highlight guard and skin / environment controls. Classic (the
  default) keeps your picture exactly as it was. On lmxxf the Highlight colour guard (now on by default)
  runs before the composition, so it no longer cancels Composition colour on bright pixels (new in this
  build, not yet tested in a game). Composition colour multiplies the saturation that is there, so in
  near-grey scenes such as Silent Hill 2's fog it stays subtle - try 3-4; Composition detail only scales
  the model's own change in brightness. The log now records every slider change.
• lmxxf Full network (Neural > Performance, off by default): runs all 71 of the network's blocks
  instead of skipping three - slightly more faithful, about 0.5 ms slower at 1080p (16.6 -> 17.1 ms on
  an RX 9070 XT).
• lmxxf's RAM leak is fixed, GPU-bound games included: its RAM use climbed for as long as NR ran
  (about 45 GB an hour at 60 NR fps; 14.8 GB in one Silent Hill 2 session). The runtime now also frees
  that memory while the GPU is still busy with the next frame, never waiting on it. On a pipelined probe
  (RX 9070 XT, 1080p, frames in flight like a GPU-bound game) it stays flat over 18000 frames, same
  picture, median frame time about 0.1 ms higher. lmxxf also no longer keeps about 106 MB of VRAM on
  every network rebuild at the 1080 tier. danielblnc is not affected (its runtime already frees it).
• lmxxf colours that suddenly turn grey: fixed: Highlight colour guard and Auto-exposure highlight
  cap (Neural tab) are now on by default - tested in Silent Hill 2. Turn both off to get the 0.3.2
  picture back. Also on: an exposure self-heal for a blown-out feed and a lower auto-exposure floor (not
  yet tested in a game). lmxxf only; danielblnc's equivalent is planned for 0.3.4.
• XeFG frame generation up to 10X, opt-in and paced (6X stays the default; D3D12 games only - XeFG has
  no Vulkan output). Set XeFG ceiling (restart) under FG Output (4X, 6X, 8X or 10X), restart, then pick
  the multiplier in the MFG combo. Above 6X it needs OptiScaler's own XeFG provider with Extra pacing on;
  a game's own XeSS 3 copy stays at 6X. 10X needs a 360 Hz+ display and a frame cap at refresh / 10;
  latency is high, and the provider reserves about 128 MiB more VRAM at 4K. 7X-10X is not yet confirmed
  in a game: testers, please send OptiScaler.log.
• XeFG: new "Allow dilated motion vectors" option.
• Multi-frame generation stutter fix, 3X-10X (not yet tested in a game): the ~1 fps lock after XeFG
  turns on, after FG off/on or a multiplier change, is now bounded, and the overlay no longer resets
  command allocators the GPU still uses. Testers at 4X-10X: please send OptiScaler.log.
• lmxxf on Vulkan games no longer freezes (Indiana Jones; the cause was a lazy weight upload inside
  OptiScaler's Vulkan bridge - expect a one-time ~1 s hitch when NR starts). If lmxxf is held back on a
  Vulkan title, the Neural tab's Retry lmxxf button clears it.
• danielblnc on Vulkan games (Indiana Jones): Neural passes 2-3 now run as 1 pass (each extra pass
  used to stall about 5 s, then NR stopped), the 80 ms post-submit wait is gone, and a note under the
  runtime choice explains its log lines. New opt-in [DlssNr] Amd (experimental, off by
  default, not yet tested in a game): expected to remove the ~5 s pause on the first NR frame - testers,
  set it to true and send OptiScaler.log, amd_presr.log and dlssnr_on_amd.log.
• Streamline start-up hardened, with diagnostics: OptiScaler's Streamline plugin hooks can no longer
  turn a plugin that refuses to load into slInit error 0x18, and several null / stale-pointer cases in
  slInit are guarded. Seen with NBA 2K27 on AMD, but not confirmed as its cause: OptiScaler.log now logs
  slInit returned ... and [SLINIT] lines - send it with any Streamline start-up error.
• lmxxf is faster: lmxxf 0.29's kernels and options, bit-exact, about 16.7 ms instead of 20 ms per
  network run at 1080p on an RX 9070 XT.
• FSR Ray Regeneration path-traced profile, automatic for Resident Evil Requiem and PRAGMATA: less
  grain on faces under path tracing, plus an experimental skin smoothing and A/B controls.
  Menu: Neural > Quality > Ray Regeneration.
• Highlight proxy (danielblnc, experimental, off by default): highlights are decoded from the original
  frame, so they no longer clip at 200 and the model's dimming is no longer magnified.
• RR controls now in the AMD menu, shown only while the game runs Ray Reconstruction.
• UE5 robustness fixes for lmxxf.
• Logs: the previous session's log is kept (OptiScaler.previous.<exe>.log) and a crash is flagged
  ("no clean exit recorded"); both runtimes' logs now carry VRAM, budget and process memory figures.
• Fixes: danielblnc after Ray Regeneration no longer applies the camera jitter to the finished image;
  a failed XeFG swapchain init no longer keeps its context (about 0.8-1 GB of VRAM at 4K); menu text fixes.

Everything else: CHANGELOG.md in the zip.

Known issues

• lmxxf: each NR resolution or DLSS mode change keeps about 97 MB of VRAM and as much RAM at the 1080
  tier until the game restarts (an AMD HIP driver leak; the fix is planned for 0.3.4).
• danielblnc: each new NR size above about 1 MP keeps about 75-350 MB of VRAM per Neural pass until the
  game restarts (read from the runtime's code, not yet measured in a game). The NR resolution slider now
  moves in 5% steps to keep the number of sizes down; restart the game after many changes.

Copyright

AMDNR is Copyright (c) 2026 3zwr1 (AMDNR), a GPL-3.0 fork of OptiScaler. Any copy, fork or derivative
that uses AMDNR's work must keep its notices and credit AMDNR by 3zwr1 (GPL-3.0 section 7(b); see
Licenses/AMDNR_NOTICE.txt in the zip). The upstream work below stays with its authors, under their own
licences. The source code will be published here with AMDNR 0.5.0.

Credits

• TheAutomatic — DLSS 5 AMD project — https://github.com/TheAutomatic/dlss-5-amd-project
• Dagherbou, wilsjo2 and Matheus — the DLSS-NR and pre-SR code this tree descends from (GPL-3.0)

• danielblnc — DLSS-NR on AMD — https://github.com/danielblnc/DLSS-NR-on-AMD (Runtime.zip, unmodified)

The danielblnc runtime is Daniel Blanco's work. The AMD neural runtime in Runtime.zip
(dlssnr_amd_pass1..3.dll) is DLSS-NR on AMD by Daniel Blanco (danielblnc) -
https://github.com/danielblnc/DLSS-NR-on-AMD. Copyright (c) 2026 Daniel Blanco, all rights reserved.
AMDNR ships it unmodified, with his permission; it is not AMDNR's work. Please support his project.

• lmxxf — https://github.com/lmxxf/dlss5-on-amd-9070xt-porting (the HIP runtime and kernels, MIT)
• Nukem9 — dlssg-to-fsr3 — https://github.com/Nukem9/dlssg-to-fsr3 (GPLv3, unmodified)
• RenoDX — clshortfuse — https://github.com/clshortfuse/renodx (colour composition maths, MIT)
• Coldwood1026 — XeFGUnlock (GPL-3.0), the base of the built-in XeFG multi-frame unlock and its pacing
• burak113 — the FSR Ray Regeneration preprocessor (OptiScaler branch ffx-denoise-experimental, GPL-3.0)
• OptiScaler — Overclockers — https://github.com/Overclockers/OptiScaler-Releases