GitHub release — AMDNR v0.3.4

Title: AMDNR v0.3.4 — the new menu, faster lmxxf on RX 7000, handheld APUs (experimental), one set of controls for both runtimes, AMDNR Screen GI (preview), Save report

Tag: v0.3.4

DLSS 5 Neural Rendering on AMD, built into OptiScaler.

Join the Discord (support, bug reports, test builds): https://discord.gg/AMDNR
Ko-fi: https://ko-fi.com/3zinr

The danielblnc runtime is Daniel Blanco's work. The AMD neural runtime in the *Runtime.zip files
(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.
The lmxxf runtime is built on lmxxf by Kien (MIT). Full credits for everyone else are at the end of this page.

Highlights

• The new menu. One layout for both runtimes, danielblnc and lmxxf, and every tab redone in the same look.
  The Neural tab starts with Enable, the runtime and one status line, then Performance, Quality, Image look and,
  while a game runs it, Ray Regeneration. Rarely used tools sit in a closed row (Diagnostics / Runtime options /
  Experimental). Help opens when you hover a label. Same ini keys: your OptiScaler.ini works.
• lmxxf is faster on RX 7000 (RDNA 3): the NR size now snaps to lmxxf's cheaper network size tier by default on
  RX 7000, Radeon 8060S / 8050S and the handheld APUs. On an RX 7800 XT, 1440p with FSR Quality (a 1706x960
  render) now runs the 900 tier instead of the 1080 tier: 73.3 -> 52.2 ms per network run. A 1080p render at NR
  resolution 70% now runs the 720 tier instead of the 900 tier: 52.2 -> 34.4 ms (network time in lmxxf's probe on a tester's card, outside a game). RX 9000 is unchanged.
• Handheld APUs (experimental): Z1 Extreme / Z2 / Radeon 780M and Z2 Extreme / Radeon 890M / 880M run lmxxf
  through AMDNR's RDNA 3 backend, with two new, smaller network sizes (360p and 576p) and the model every 4th
  frame by default. Slow: first results on a ROG Ally (Z1 Extreme): 54.7 ms per network run at 360p in lmxxf's probe outside a game;
  in a game, one tester's run of Shadow of the Tomb Raider (1280x720, XeSS, Handheld preset): 62 ms on average at 360p with the model every 4th
  frame, about 29 fps with NR on. Reports wanted.
• lmxxf is faster on RX 9070 / RX 9070 XT: network time 15.35 -> 14.08 ms at 1080p (10.79 -> 9.96 ms at 900p, 6.97 -> 6.63 ms at 720p) on an RX 9070 XT,
  measured outside a game, the same network output bit for bit,
  with lmxxf's new kernels from lmxxf 0.31 (Kien, MIT) on top of AMDNR's c32w.
• Both runtimes, one set of controls: lmxxf gains Network output, Encoding, Residual edge fade and Game exposure; danielblnc gains its runtime settings (Network style, Tone curve, Black lift, Game exposure) and
  Network history with Model interleave off.
  • lmxxf now also follows Native character mask, Structure intensity and Character structure.
• Fast mode (opt-in): lmxxf runs its network one size tier lower: about 29% less network time at 1080p (RX 9070
  XT, measured outside a game), fine detail a little softer. danielblnc builds that have their own Fast mode get the
  same row.
• AMDNR Screen GI (preview, off by default): AMDNR's own screen-space bounce light and ambient occlusion, before
  NR and the upscaler, with NR on or off; about 1 ms at High for a 1080p render on an RX 9070 XT (GPU lab, not a
  game). Neural > Experimental. The inherited Screen-space GI is retired from the menu.
• Save report: one button (Neural > Diagnostics, or Advanced > Logging) zips your logs, your ini and a report
  for a bug report, with your Windows user name and PC name masked.
• Menu input fixes: with the menu open, a game that hooks the keyboard at a low level (Assetto Corsa) no longer
  blocks Alt+Tab and the Windows key for the whole desktop, and one key press toggles the menu or NR once.
• Game fixes: Shadow of the Tomb Raider no longer crashes when the upscaler starts (seen on a ROG Ally); Marvel's
  Midnight Suns no longer crashes on danielblnc; lmxxf keeps running in games that submit their frames late (The Last
  of Us Part II); no false "No clean exit" after quitting Forza Horizon 6.
• lmxxf no longer switches its colours by itself in long sessions (its stats counters were not cleared, which
  could trigger a false self-heal).

Which files to replace

All three changed. Coming from 0.3.3.x, replace:

1. OptiScaler.dll (rename it as before, e.g. dxgi.dll or winmm.dll)
2. LmxxfNrRuntime.dll (0.3.4, Properties > Details)
3. LmxxfNrRuntime.pak (the 0.3.3.2 pak plus new entries: the handheld builds for gfx1103 and gfx1150, and lmxxf 0.31's kernels (Kien, MIT);
   the c32w kernels are unchanged)

Keep the three together. Mixed with older files:

• new OptiScaler.dll + 0.3.3.2 LmxxfNrRuntime.dll: NR runs, but lmxxf ignores the character mask and structure
  controls (one log line), has no small network sizes, and a handheld gets only the slow 720 tier;
• new LmxxfNrRuntime.dll + 0.3.3.2 pak: desktop GPUs run as before; a handheld gets "no lmxxf modules for this
  GPU";
  • RX 9070 / 9070 XT keep 0.3.3.2's speed without the new pak;
• a handheld with an older OptiScaler.dll: the new runtime refuses it ("this handheld needs OptiScaler.dll 0.3.4 or
  newer").

Your danielblnc files stay. Keep a backup of the old three. Your own OptiScaler.ini works; new keys use their
defaults, but see Settings you saved before below. If the status line says c32w=off:nofile with the new pak,
an old DLSS5-AMD\native-game-tiled-assets folder next to the game's exe is used instead of the pak: remove or
rename that folder.

New install: extract AMDNR-v0.3.4.zip into the game folder, rename OptiScaler.dll to dxgi.dll (or
winmm.dll / version.dll / OptiScaler.asi with an ASI loader), then INSERT for the menu. For danielblnc,
also extract one runtime zip from the table below into the game folder.

Downloads

Use all three pass DLLs from one zip.

Settings you saved before

Some values you saved yourself act differently in 0.3.4, and a few defaults changed:

• Preset (Quality / Balanced / Performance) now sets only NR resolution (100 / 85 / 70%) and turns Dynamic NR off;
  it no longer changes Temporal stability, Sharpening, Detail or Colour strength, so your NR style stays.
• Settings saved while on danielblnc now also act on lmxxf: AmdEncoding (sRGB / Gamma 2.2), AmdResidualFade,
  AmdNetworkOutput=true and AmdRtgiEnabled=true (Screen-space GI, which costs GPU time and VRAM).
• [DlssNr] AutoMask=false now turns danielblnc's native character mask off (0.3.3.2 kept it on).
  • On lmxxf, AutoMask=false, a Structure intensity other than 1, your own Character structure and the NR
    styles that set structure (Cinematic, Crisp, Natural, Vivid) now change the picture too, on every GPU; each change
    rebuilds lmxxf's network (about a one-second hitch).
• AmdInterleaveModelHistory=true now also keeps danielblnc's network history with Model interleave off.
• A saved AmdInterleavePacing between 0 and 1 now also paces lmxxf with Model interleave on, which lowers the frame
  rate; the default (-1, automatic) stays off on lmxxf.
• lmxxf's "Classic carry" interleave preset is now saved as AmdInterleavePreset=1 (0.3.3.2 wrote 6); an old 6 still
  reads as Classic carry on lmxxf.
• Reset appearance filter resets only the appearance filter ([AmdLook]), no longer NR resolution, passes or
  structure.
• On RX 7000, Radeon 8060S / 8050S and the handhelds the tier snap is on unless your ini says
  [DlssNr] AmdLmxxfTierSnap=false.

What changed

The menu

• One look on every tab, drawn from the approved mock: new theme, text tabs with a red underline, a one-line
  footer (Menu Scale, Save Settings, Close), help on hovering a label (no more (?) markers), Menu Scale scales
  everything. The window title reads AMDNR v0.3.4 - <exe> - <game>; the startup splash says "AMDNR - Insert for
  menu".
• Header: AMDNR (hover for the version and build), Discord, GitHub (it opens AMDNR's README, which replaces the
  old Guide button) and the copyright, then the credits row ("Neural runtimes: DLSS-NR on AMD by Daniel Blanco
  (danielblnc) - lmxxf by Kien (MIT)"; click Daniel's name for his GitHub, hover for the full credits).
• Neural tab, top: Enable with its key, the Neural runtime combo with one state word (running / restart the game to
  switch / not installed / not for this GPU), the running runtime's credit as a link, and one status line (state,
  size, NR and model rate, and the NR cost in ms). At most one orange line below it, with a fix button
  where there is one (Retry lmxxf, Switch to danielblnc after an lmxxf error, Open Upscaling for a D3D11 / Vulkan game
  on the wrong upscaler). Preset is three buttons; NR style reads the style from your sliders on both runtimes;
  Style slots (Store / Apply / Clear) sit in a small popup.
• Performance and Quality: NR resolution shows its cost on the same row (on lmxxf priced by the network size tier
  it runs; hover for the tier);
  Interleave preset appears under Model interleave while it is on; one Network history checkbox for both runtimes
  under Quality > More quality options, which also holds Output smoothing, Stability mode, Residual temporal, Residual
  edge fade and Still-surface steadiness. A row a runtime does not have is greyed with the reason, or hidden with a
  count.
• Image look: Colour composition, Detail and Colour strength, then Model strength, Exposure and highlights, the
  Appearance filter (moved here from the tools row) and Skin and environment.
• Ray Regeneration has its own section (Neural > Ray Regeneration, was under Quality) while a game runs FSR Ray
  Regeneration, with the new sliders listed below.
• Tools row (closed at start): Diagnostics (Network output, Debug view, the RR debug view, the ghost and
  self-tuning readouts, the NR cost in ms, a GPU line, an Edit shaper A/B combo for danielblnc, and Save report), Runtime options
  (Encoding, Every-frame NR, NR slots, Highlight proxy) and Experimental (AMDNR Screen-space GI, a preview; its settings show while it is on; the inherited Screen-space GI is
  retired from the menu and shows, tagged "retired", only while your ini has it on).
• Other tabs: Upscaling starts with the upscaler and one status line and holds Render resolution; Image holds
  Sharpness, Textures, Init Flags and the Magnifier; Frame Gen starts with FG Input / Output; Interface has the FPS
  overlay and Keybinds (each key is a button); "Components: N of 7 active" sits in the header of every tab, under the credits (the component chips left every
  tab's header); Advanced starts with Active Quirks, then Display (V-Sync combo), Compatibility and Logging (Save
  report and the build stamp).
  "DLSS w/Dx12" is no longer listed on AMD / Intel cards.
• On an AMD or Intel GPU without runtime files the Neural tab shows only the "No neural runtime found" card (and Ray
  Regeneration while it runs), not the NVIDIA layout.
• The first-launch runtime chooser lists both runtimes with their credits and whether they run on this GPU. Before the
  first frame the Neural tab and the Home notice ("Neural Rendering: On (lmxxf)") name the runtime that will run.
• Also new: a Handheld preset (handheld APUs only), the exact runtime versions in the Neural runtime combo (e.g.
  "lmxxf 0.3.4"), "default" / "custom" tags on the folded sections, a closed Live row under the status line, Fast
  mode in Performance, and a line when a game submits its frames late.

lmxxf

• Network size tiers on RX 7000: see Highlights. The rule: a size nearer the next smaller tier is lowered to it
  (cheaper, slightly softer); a size that fills 90% or more of its tier is grown to fill it (same cost, a bit sharper);
  never above the frame's own size. On by default on RDNA 3 GPUs only; [DlssNr] AmdLmxxfTierSnap=true / false
  forces it on or off. The RX 7000 and Radeon 8060S advice is now "start with NR resolution at 70% or lower".
• Two smaller network sizes, 360p (640x360) and 576p (1024x576). Handhelds use them by default.
  [DlssNr] AmdLmxxfTierCap = 0 (auto), 360, 576, 720, 900 or 1080 caps the network size on any GPU (read at game
  start, not saved).
• New on lmxxf: Network output (the plain network answer); Encoding sRGB / Gamma 2.2 (the game's colour is decoded
  to linear light for the network; upscaler path only); Residual edge fade (acts while the network runs at a size other than the frame's); Game exposure (Auto / Use / Ignore the
  title's exposure texture); the interleave pacing readout; the NR cost in ms.
• Native character mask, Structure intensity and Character structure act on lmxxf (Image look > Model
  strength), and so do the NR styles that set structure. Each change rebuilds the network (about a second). At their
  defaults lmxxf gets exactly 0.3.3.2's input.
• Stats fix: lmxxf's stats window now clears all 12 counters, so long sessions no longer build up a false
  "self-heal" that switched lmxxf's colours mid-game.
• LmxxfNrRuntime.dll 0.3.4: the small network sizes, the handheld builds and checks, and the character mask
  controls. At default settings its output on RX 9000 equals 0.3.3.2's bit for bit (RX 9070 XT reference hashes, c32w
  on and off).
• Faster on RX 9070 / RX 9070 XT with lmxxf 0.31's kernels (Kien, MIT: the ViT projection (lmxxf031-vit-wide-deep), the C512 QKV and mix kernels (lmxxf031-c512-m32-mh, lmxxf031-c512-m32-deep) and one-wave-per-head attention (lmxxf031-c64-wave2)) on top of
  AMDNR's c32w: 15.35 -> 14.08 ms at 1080p, 10.79 -> 9.96 ms at 900p and 6.97 -> 6.63 ms at 720p, network time on an RX 9070 XT
  (outside a game), the same output bit for bit. LmxxfNrRuntime.dll checks their SHA-256 and falls
  back to the previous kernels otherwise. RX 9060 XT keeps the previous kernels for now. The runtime no longer reads
  lmxxf's own DLSS5_VIT_* environment switches.
• Fast mode (Neural > Performance, [DlssNr] AmdLmxxfFastMode, off by default): the network runs one size tier
  lower (1080 -> 900, 900 -> 720): about 29% less network time at 1080p, fine detail a little softer. Off, lmxxf gets
  exactly the input it got before.
• The Last of Us Part II (games that submit a frame's work late): a job whose command list arrives one frame late
  is kept and its answer used a frame later, warped into place, instead of dropping it and restarting the network
  history ([DlssNr] AmdLateSubmitGrace=false = 0.3.3.2). After an error that stops lmxxf for the session it no
  longer calls the runtime every frame. The lmxxf runtime also asks Streamline for the real device (Midnight Suns).
• RenoDX colour composition stays off with Encoding sRGB / Gamma 2.2 and with Network output, as on danielblnc.

danielblnc

• Runtime settings in the menu (Image look > Model strength and Exposure and highlights): Network style (Auto /
  Style 0-2), Tone curve (Auto / Reinhard / ACES), Black lift, Game exposure (Auto / Use / Ignore). Auto writes
  nothing, as in 0.3.3.2. They work with the public danielblnc 0.3.3 and 0.4.0 builds; with other builds they are
  greyed. A change restarts the model history.
• Network history with Model interleave off (AmdInterleaveModelHistory=true).
• Native character mask off works ([DlssNr] AutoMask=false), with the Native character mask row in Model
  strength.
• Highlight colour guard (Image look > Exposure and highlights > Highlight colour guard, or
  [DlssNr] AmdDanielHighlightGuard=true; off by default): over-exposed areas (sky, lamps, lit fog) keep the game's
  colour instead of going grey. The first use compiles a shader (a short hitch).
• The style change away from 100% NR resolution: two new edit shaper modes to compare, F1 and F2, off by default
  until testers pick one: Diagnostics > "Edit shaper (A/B, not saved)", or [DlssNr] AmdEditShaper=true with
  AmdEditShaperLimit=1 or 2 (AmdEditShaperScope, AmdEditShaperCarryCap for more A/B).
• NR cost in ms on DX12 titles (measured only while the menu shows it).
• Fast mode for danielblnc builds that have one ([DlssNr] AmdDanielFastMode; auto = the runtime's own mode): the
  row shows only with such a build. The runtimes in this release's runtime zips do not have it.
• Marvel's Midnight Suns: danielblnc no longer crashes where the game's D3D12 device is a Streamline proxy; it runs
  on the device behind it ([DlssNr] AmdStreamlineDeviceFix=false = 0.3.3.2).
• Games that submit late (The Last of Us Part II): amd_presr.log and the Neural tab say how often NR ran; the fix
  for danielblnc is planned for 0.3.5, so use lmxxf there.

Ray Regeneration (FSR-RR)

• New sliders (Neural > Ray Regeneration > More Ray Regeneration options): Stability bias,
  Cross-bilateral normal strength, Gaussian kernel relaxation, Radiance clip, Max radiance; and Profile temporal
  values under the Path-traced profile ([FSR-RR] FfxDenoiserPathTracedTemporal).
• Bias mask: Resident Evil Requiem publishes no DLSS bias mask, so Bias mask strength does nothing there; the
  slider is greyed "no bias mask in this game".
• Unreal titles whose DLSS plugin sends empty camera matrices (Satisfactory): after the fallback to FSR, Ray
  Regeneration stays off for the game session, and the Neural and Upscaling tabs say so and why (turn Ray
  Reconstruction off in the game). In other titles, re-picking FSR Ray Regeneration after a fallback clears the "off"
  state.
• The game textures Ray Regeneration kept are released when it turns off.
• RX 7000 / RDNA 3: FSR Ray Regeneration is offered by default again (0.3.3.2: RDNA 4 only). If the driver
  refuses it, a notice says "FSR Ray Regeneration could not start" and the game gets FSR without the denoiser.
  [FSR-RR] FfxDenoiserAllowPreRdna4=false keeps it to RDNA 4; RX 6000 and older get it only with true (Upscaling
  tab: "Offer FSR Ray Regeneration on this GPU (restart)").
• Fixed: in Streamline titles that load sl.interposer before OptiScaler, the Ray Reconstruction option could
  be missing on RDNA 4; the RR hooks now attach before the GPU is known and decide per call.
• Diagnostics for titles without camera data: once per RR handle, a small probe (about 0.8 MB, a few dispatches) on
  frames whose camera is missing. Two experimental ini keys, off by default: FfxDenoiserAlbedoFp16,
  FfxDenoiserNgxDirectSLConstants.
• The path-traced profile is retired (not recommended): every Resident Evil Requiem report found it worse than
  the normal route. Its row shows only while it is on, tagged "retired", and the log warns once.
• Much less log noise: repeated RR input errors log 5 times, then every 1000th; the RR dispatch snapshot is logged on
  resets 1-3 and every 20th; the version check and the denoiser provider line no longer raise false warnings.

Bug reports and logs

• Save report (Neural > Diagnostics, or Advanced > Logging): AMDNR-report-<exe>-<date>.zip in the game folder (on the Desktop if the folder is
  read-only, else in %TEMP%) with report.txt, the logs (OptiScaler.log, the previous logs, amd_bridge.log,
  amd_presr.log, lmxxf_backend.log, dlssnr_on_amd.log) and the ini files. Your Windows user name and PC name
  are masked.
• Three previous logs are kept per game ([Log] KeepPreviousLogs=3; 1 = 0.3.3.2).
• Optional crash log: [Log] CrashHandler=true (and CrashDump=true for a minidump) writes amdnr_crash.log, which
  names the module of a crash; Save report collects it. Off by default.
• A normal quit no longer leads to a "No clean exit recorded" warning at the next start.
• Repeated frame-generation resource-create errors (58,000 lines in one RDR2 log) log 5 times per resource, then
  every 1000th; amd_bridge.log writes one "AMD settings:" line when you change an NR setting.

Stability and input

• Menu input (Assetto Corsa and other games with low-level input hooks): with the menu open, the game's hook is
  skipped and the key passes on, instead of being swallowed for the whole desktop
  ([Hotfix] MenuLowLevelHookPassThrough=false restores 0.3.3.2). A second menu or NR key press within 400 ms is
  ignored ([Hotfix] MenuToggleDebounceMs=0 restores 0.3.3.2).
• OptiScaler no longer loads into danielblnc's installer (under any name) or into crash reporters and browser helper
  processes, so a crash reporter cannot empty OptiScaler.log.
• Frame generation no longer reads a freed depth / velocity copy after a failed re-create.
• [Anisotropy] settings saved by the menu are read back correctly.
• Help tooltips and the splash no longer garble text that contains % or ##.
• [Hotfix] WarnMissingREFramework=false hides the REFramework warning.
• Shadow of the Tomb Raider (and games that make their D3D12 device twice) no longer crash when the upscaler
  starts: OptiScaler now uses the device of the game's command list.
• Radeon 780M / 760M / 740M (ROG Ally, Legion Go): FSR 4 INT8 is no longer switched on by itself; with
  Dx12Upscaler=auto the upscaler is XeSS ([FSR] Fsr4ForceModel=2 still forces FSR 4).
• An NR resolution change alone no longer pauses NR for 300 ms. Forza Horizon 6: no false "No clean exit" after a
  normal quit. Wine / Proton: the Neural tab says NR cannot run there (HIP is Windows-only) while RR and FSR work.
• The menu's input-health log lines no longer fire on healthy sessions.
• OptiScaler.dll reports AMD-NR v0.3.4.

Everything else: CHANGELOG.md in the zip.

Handheld APUs (experimental)

• Runs lmxxf (danielblnc's runtime has no code for these chips) on:
  • Z1 Extreme, Z2, Radeon 780M (gfx1103, 12 compute units);
  • Z2 Extreme, Radeon 890M, Radeon 880M (gfx1150).
• Defaults on these APUs (only while your ini has no value of its own): the 360p network size (640x360), the model
  every 4th frame (Model interleave), one Neural pass. [DlssNr] AmdLmxxfTierCap=576 gives 576p (1024x576): sharper,
  more cost. Model interleave Off is saved as [DlssNr] AmdInterleave=1 (also off), so it stays off.
• Needs all three new files (see Which files to replace).
• Not supported: Z1 / Radeon 740M (4 compute units), Radeon 760M (8), Radeon 860M / 840M, and RDNA 2 (RX 6000,
  Steam Deck). RX 6000 is planned for AMDNR 0.4.0.
• Speed: first tester result (ROG Ally, Z1 Extreme, lmxxf's probe outside a game): 54.7 ms per network run at
  360p, 110.9 ms at 576p, 168.0 ms at 720p (an RX 7800 XT needs 34.4 ms at 720p). In a game (one tester's run): Shadow of the Tomb Raider (1280x720, XeSS, Handheld preset), 62 ms per network run on average at 360p
  with the model every 4th frame, about 29 fps with NR on. Not yet measured on a Z2 Extreme / 890M. Expect a large drop in frame rate. The Neural tab's status line and Diagnostics > NR cost show
  the real network time on your device.
• If the Neural tab says HIP is missing, install AMD's Adrenalin driver.
• The Neural tab has a Handheld preset: NR resolution 100%, Dynamic NR off, the model every 4th frame, 1 Neural
  pass, Full network off. On these chips the network stays at 360p whatever the Preset, and the menu says so.
• FSR 4 INT8 is no longer switched on by itself on these chips (XeSS with Dx12Upscaler=auto), and Shadow of the
  Tomb Raider no longer crashes when the upscaler starts.
• Please report: Save report zip, your device, its power mode and driver: https://discord.gg/AMDNR (#bug-report).

Resident Evil Requiem needs REFramework

A known requirement, not an AMDNR bug. OptiScaler relies on REFramework to get past Capcom's anti-tamper
(OptiScaler wiki). Without it, Resident
Evil Requiem (and its demo) crashes 15-60 s after launch. Put dinput8.dll from REFramework.zip in the latest
nightly (https://github.com/praydog/REFramework-nightly/releases) next to dxgi.dll, and change REFramework's menu
key (e.g. to Delete): it is also Insert. AMDNR warns when it is missing.

Known limits

• The RX 7000 numbers above are lmxxf's network time per tier on an RX 7800 XT, measured outside a game (with
  0.3.3.2's runtime, which runs the same RDNA 3 kernels as 0.3.4), not a game's frame rate.
• The tier snap lowers some sizes (for example 1440p FSR Quality at 100%), so the picture can be slightly softer;
  [DlssNr] AmdLmxxfTierSnap=false restores 0.3.3.2's sizing.
• Handhelds are experimental: measured on one ROG Ally (lmxxf's probe and Shadow of the Tomb Raider), not yet on a
  Z2 Extreme / 890M; the 360p and 576p sizes are reached on desktop GPUs only with [DlssNr] AmdLmxxfTierCap.
• danielblnc: leaving 100% NR resolution can still change the style at default settings; compare the F1 / F2 shaper
  modes and tell us which one matches.
• danielblnc's runtime settings need the public 0.3.3 or 0.4.0 runtime.
• AMDNR Screen GI is a preview, off by default: screen-space (light from off-screen is missing), D3D12 games with an
  upscaler only, run so far only in AMDNR's GPU lab (the bounce into shadow is weak, still scenes can shimmer).
• RenoDX composition is refused with Encoding sRGB / Gamma 2.2 and with Network output, on both runtimes.
• danielblnc: ticking Network history mid-game can freeze the picture for about 4 s (fix planned for 0.3.5).
• Ray Regeneration stays off for the session in Unreal titles whose DLSS plugin sends empty camera matrices.
• Save report masks the user folder and PC name; a personal name elsewhere in a game path is not masked. Look at the
  zip before you post it.
• The c32w speed-up runs on RX 9070 / RX 9070 XT only; RX 9060 XT keeps the old kernels unless you set
  AMDNR_C32W=1 (untested).
• danielblnc is slow on RX 7000 (about 27 ms per frame even at 320x180 with 0.4.0 on an RX 7900 XTX): lower the
  render resolution or NR resolution.
• An old DLSS5-AMD\native-game-tiled-assets folder next to the game's exe overrides LmxxfNrRuntime.pak: remove
  or rename it.

Not yet tested in a game (built and checked on the desk: unit tests, reference hashes on an RX 9070 XT, menu
screenshots).

• lmxxf on a Z2 Extreme / 890M / 880M, and the 576p size in a game;
• the tier snap default in a game on RX 7000;
• Ray Regeneration by default on RX 7000 in this build;
• lmxxf's Network output, Encoding, Residual edge fade and Fast mode, and AMDNR Screen GI;
• the character mask and structure controls on lmxxf;
• lmxxf 0.31's kernels in a game;
• danielblnc's runtime settings and the F1 / F2 shaper modes;
• the menu input fix in Assetto Corsa, the Ray Regeneration session latch in Satisfactory, and the crash log;
• the Midnight Suns and The Last of Us Part II fixes.

Where a change has an ini switch to turn it off, it is named above. Reports with logs help:
https://discord.gg/AMDNR (#bug-report).

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.

The AMDNR menu - its layout, design, texts and the code AMDNR added for it - is Copyright (c) 2026 3zwr1 (AMDNR). It is part of this GPL-3.0 fork, with these additional terms (GPL-3.0 section 7): (b) anyone who reuses any part of it must keep this copyright line and credit AMDNR by 3zwr1 visibly, in the menu and the README; (c) you may not present it, or a modified copy, as your own work; modified versions must be clearly marked as changed; (e) no rights are granted to the AMDNR name or logo; other projects may not use them.

Credits

• Daniel Blanco (danielblnc) — DLSS-NR on AMD — https://github.com/danielblnc/DLSS-NR-on-AMD (the runtime in
  the *Runtime.zip files, unmodified, used with his permission)
• lmxxf (Kien) — https://github.com/lmxxf/dlss5-on-amd-9070xt-porting (the HIP runtime, its network and
  kernels, MIT)
• lmxxf 0.31 kernels — Kien (lmxxf), MIT: the ViT projection (lmxxf031-vit-wide-deep), the C512 QKV and mix kernels (lmxxf031-c512-m32-mh, lmxxf031-c512-m32-deep) and one-wave-per-head attention (lmxxf031-c64-wave2) in LmxxfNrRuntime.pak are lmxxf's, built
  by AMDNR from lmxxf's sources and build recipe; AMDNR's part is loading, SHA-256 pins, per-GPU gating and fallbacks
• TheAutomatic — the lmxxf runtime API that LmxxfNrRuntime.cpp, LmxxfNrApi.h and LmxxfProductionOptions.h
  derive from: portions contributed to lmxxf by TheAutomatic (MIT)
• TheAutomatic — DLSS 5 AMD project — https://github.com/TheAutomatic/dlss-5-amd-project (many thanks: the lmxxf
  integration builds on this work)
• AMDNR's RDNA 3 backend — 3zwr1 (AMDNR): lmxxf's network on RX 7000, Radeon 8060S / 8050S and, new, the handheld
  builds (gfx1103, gfx1150); the network size tier policy and the 360p / 576p sizes (Copyright (c) 2026 3zwr1
  (AMDNR))
• c32w kernels — 3zwr1 (AMDNR): AMDNR's own one-wave C32 kernels for lmxxf's network on RDNA 4 (Copyright (c)
  2026 3zwr1 (AMDNR)); they run lmxxf's network (MIT). Acknowledgement: AMD's public RDNA 4 WMMA documentation
  (AMD GPUOpen, the ROCm matrix instruction calculator), used for ideas only.
• Dagherbou, wilsjo2 and Matheus — the DLSS-NR and pre-SR code this tree descends from (GPL-3.0):
  https://github.com/Dagherbou/OptiScaler_DLSSNR, https://github.com/wilsjo2/OptiScaler-DLSSNR-PreSR-Multipass,
  https://github.com/MatheusGViana/dlss-5-amd-project
• Screen-space GI (the inherited effect; retired from the menu in 0.3.4) — an effect AMDNR inherited from the
  OptiScaler AMD pre-SR lineage above; the credit is its original authors'
• AMDNR Screen GI (0.3.4 preview) — AMDNR's own, Copyright (c) 2026 3zwr1 (AMDNR), written from published papers
  (Therrien, Levesque and Gilet 2023; Jimenez et al. 2016; Schied et al. 2017; the full list is in CHANGELOG.md)
• hhkbble — the CubeScale gamut handling
• 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 and contributors — https://github.com/Overclockers/OptiScaler-Releases

Not shipped by AMDNR, but needed for some features: REFramework by praydog (Resident Evil Requiem) and
DLSS Enabler by Artur Graniszewski (the Enabler frame-generation options).

None of this is endorsed by, affiliated with, or supported by NVIDIA, AMD, or any game publisher. NVIDIA's
nvngx_dlssnr.dll is not in these archives. Use it at your own risk.