English | 中文 | Português | Español | العربية | Français | Italiano | Русский | Polski

We need your support. Join the Discord server — https://discord.gg/AMDNR — for help, bug reports and test builds; every report with a log makes the next build better.

DLSS 5 Neural Rendering running on AMD GPUs, built into OptiScaler so it works in any Direct3D 12 game OptiScaler already hooks. On top of the neural pass: model interleave for a large frame-rate gain, residual composition, XeSS frame generation unlocked up to 6X (up to 10X opt-in in D3D12 games), and FSR Ray Regeneration for games that use DLSS Ray Reconstruction. Since 0.3.5, AMDNR Anywhere (preview) brings Neural Rendering to games that have no upscaler of their own, through the AMDNR Launcher, with nothing written into the game folder (see "AMDNR Anywhere").

Discord: https://discord.gg/AMDNR — support, bug reports (#bug-report), test builds.

Support the project: 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. Full credits for everyone else are at the end of this page.

New in 0.3.5: AMDNR Anywhere (preview): Neural Rendering for games that have no DLSS, XeSS or FSR 2 of their own - one PLAY ANYWHERE button in the AMDNR Launcher, nothing written into the game folder; RX 9000 (RDNA 4) in this release (see "AMDNR Anywhere"). Ray Regeneration has its own tab, right after Upscaling, and its status lines say per card and per API what runs and why not; on RX 7000 it is no longer offered by default (the game keeps its own denoiser), with an experimental opt-in that is not supported. The neural pass can run after upscaling ([DlssNr] AmdPlacement=post, Neural > Performance > Placement; the default pre is unchanged). The Frame Gen tab says why nothing generates and names the five steps (see the frame-generation FAQ). Faster on RX 7000 and Z1 Extreme-class handhelds: about 10 percent less network time, same picture bit for bit (the 0.3.5 module set in LmxxfNrRuntime.pak). lmxxf 0.37 by Kien (MIT) is on by default on RX 9000: about 20 percent less network time on an RX 9070 XT (experimental on the RX 9060 / 9060 XT; [DlssNr] AmdLmxxfL37=false turns it off). On RDNA 3 handhelds, FSR 4 (INT8) is an experimental opt-in, and custom style slots now keep the whole look. Dynamic-resolution games no longer rebuild the network at every step, the carried edit no longer vanishes when you move on a handheld with Model interleave, Uncharted: Legacy of Thieves no longer crashes on RX 9000, and many more fixes. All three files change: replace OptiScaler.dll, LmxxfNrRuntime.dll and LmxxfNrRuntime.pak together; launcher users: it updates for you. Details: CHANGELOG.md.

New in 0.3.4.2 (hotfix): the menu in Assetto Corsa: the menu key toggles once per press, clicks shorter than a frame are no longer lost, and the runtime chooser answers to 1 / 2 / Enter / Esc, has a title-bar X and closing the menu counts as Decide later. The runtime chooser no longer opens the menu by itself (one notice instead), the Ray Regeneration section of the Neural tab no longer hides - it is always there and one dim line says why it is not running - and the Wine / Proton text says that Ray Regeneration is a known issue there. AMDNR also accepts one more danielblnc runtime layout, so a newer danielblnc build can be driven without an AMDNR update. Neural Rendering is byte-identical to 0.3.4.1 except that one accepted-layout row (the neural pass, both runtimes and the pak are unchanged): coming from 0.3.4.1 or 0.3.4, replace OptiScaler.dll only; launcher users: it updates for you. Details: CHANGELOG.md.

New in 0.3.4.1 (hotfix): Ray Regeneration is less soft on Windows (0.25 sharpening when the game sends none; to turn it off: Image > Sharpness, tick Override, slider 0); no false "Upscaler failed to run!" popup in Control Resonant; on Linux / Proton the menu works (confirmed by a player, also with frame generation on) and Ray Regeneration adds no default sharpening there (see "Linux / Proton"). AMDNR Launcher 0.3.4.1, built from your Discord feedback: nine languages, search, favourites, hide, rename, CHOOSE GAME .EXE, PLAY, a full UNINSTALL and more (see "AMDNR Launcher"). Neural Rendering is unchanged from 0.3.4 (same runtime and pak): coming from 0.3.4, replace OptiScaler.dll only; launcher users: it updates for you. Details: CHANGELOG.md.

New in 0.3.4: a new menu (the Neural tab is rebuilt, every other tab follows the same look, and a Save report button zips your logs for a bug report); lmxxf is faster on RX 7000 (1440p FSR Quality: 73.3 -> 52.2 ms per network run on an RX 7800 XT, network time measured outside a game) and on RX 9070 / 9070 XT (lmxxf 0.31 kernels); lmxxf runs on handheld APUs (experimental; one tester's run, in a game: about 29 fps in Shadow of the Tomb Raider on a ROG Ally); an opt-in lmxxf Fast mode; AMDNR Screen GI, AMDNR's own screen-space GI (preview, off by default); and many fixes. Replace OptiScaler.dll, LmxxfNrRuntime.dll and LmxxfNrRuntime.pak together. Details: CHANGELOG.md.

Installation is pretty simple. On Windows, the AMDNR Launcher does all of this for you (see "AMDNR Launcher" below). By hand:

Download these files from the latest release on GitHub (https://github.com/3zwr1/AMD-NR---OptiScaler/releases; 0.3.5 is the tag Alpha0.3.5):

• AMDNR-vX.X.X.zip (for 0.3.5: AMDNR-v0.3.5.zip), with the complete lmxxf runtime.
• For the danielblnc runtime, one runtime zip: on RX 9000 and RX 7000 alike, v0.5.0-Runtime.zip (recommended) from Alpha0.3.4.2; v0.4.3-Runtime.zip (Alpha0.3.4.2 and Alpha0.3.4.1), v0.4.1-Runtime.zip and v0.4.0-Runtime.zip (Alpha0.3.4.1) are still accepted. The lmxxf runtime is in AMDNR-vX.X.X.zip and needs no runtime zip on RX 7000 and RX 9000; handheld APUs use lmxxf only. The AMDNR Launcher picks the right zip for your GPU. See "What is in the archives".

Extract the contents of both .zip files.

First, copy all files from AMDNR-vX.X.X into the game's root folder — the same folder where the game's .exe is located.

Then, do the same with all files from the runtime zip (e.g. v0.5.0-Runtime on RX 9000 and on RX 7000).

Updating from an older AMDNR? Copy everything again and overwrite. In 0.3.5 three files changed together: OptiScaler.dll (replace the file you renamed, e.g. dxgi.dll, with the new one renamed the same way), LmxxfNrRuntime.dll and LmxxfNrRuntime.pak (440 MB). Do not mix them with older copies. You can keep your own OptiScaler.ini: new settings use their defaults. Your danielblnc runtime zip's files stay as they are. The AMDNR Launcher does this for you: UPDATE ALL, or REPAIR / UPDATE on a game that shows "Update available" (the launcher updates itself first).

Inside the game folder, find:

OptiScaler.dll

Rename it to:

dxgi.dll

dxgi.dll is the recommended option.

If the game doesn't launch or the mod doesn't load, try renaming OptiScaler.dll to one of these instead:

• d3d12.dll
• winmm.dll
• version.dll
• dbghelp.dll
• winhttp.dll
• wininet.dll

Test one name at a time. Do not create multiple copies of OptiScaler.dll. These are the names the mod loads under (plus OptiScaler.asi with an ASI loader); d3d11.dll is not one of them.

Resident Evil Requiem (and its demo) needs REFramework. A known requirement, not an AMDNR bug: OptiScaler relies on it to get past Capcom's anti-tamper (OptiScaler wiki). Without it the game crashes 15-60 s after launch ("An unhandled exception occurred"). 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. After a game update, expect crashes until REFramework is updated. PRAGMATA, Monster Hunter Wilds and Onimusha probably need it too (not confirmed).

HOME switches Neural Rendering on and off while playing (both runtimes; a small notice says On / Off). Rebind it beside the Enable checkbox in the Neural tab or under Interface > Keybinds.

That's it.

Launch the game normally and press:

INSERT

This will open the OptiScaler / AMDNR menu, where you can configure the mod however you like.

If the game still doesn't launch with any of the names above, please report it in the #bug-report channel on Discord.

When reporting the issue, also upload any .log files that may have been generated in the game's root folder.

These logs are very important and will help us identify the issue much faster.

The easy way: Save report. If the menu opens, click Save report (the last row of Neural > Diagnostics, or the first row of Advanced > Logging). It writes one zip, AMDNR-report-<game exe>-<date>.zip, into the game folder (on the Desktop if the game folder is read-only, else in %TEMP%), with report.txt, the logs and the ini files, and the menu shows where it went. Your Windows user name and PC name are replaced by placeholders; a name inside a game path outside C:\Users\ is not. Attach the zip in #bug-report. The AMDNR Launcher's COLLECT LOGS writes the same zip for any game and, since 0.3.5.1, adds the crash log and the newest dump after a crash.

The .exe is usually not where the shortcut points. Unreal games keep it under <Game>\Binaries\Win64\.

A second neural runtime (MIT-licensed, by lmxxf) can carry the pass instead of danielblnc's. RDNA 4 runs it natively; RDNA 3 (RX 7000, Strix Halo) runs it through AMDNR's RDNA 3 backend by 3zwr1 - slower there, see "RX 7000" below: start at NR resolution 70% or lower. Handheld APUs run it too, experimental (see "Handheld APUs" below). It needs two things next to the game:

1. LmxxfNrRuntime.dll - in this archive, beside OptiScaler.dll (it is copied with the rest).
2. LmxxfNrRuntime.pak (440 MB, included in the AMDNR zip) beside LmxxfNrRuntime.dll - lmxxf's weight files, HIP modules and HLSL in one encrypted, authenticated file. The runtime opens it in memory; nothing is unpacked to disk.

On the first launch that finds a runtime installed and no choice made, the menu asks which one to use ([DlssNr] NrBackend = daniel | lmxxf in the ini records it; Neural > Neural runtime changes it, on the next game start). The lmxxf edit is applied one frame late, carried by the motion vectors, so the frame never waits for the network (about 14.1 ms of network time at 1080p on an RX 9070 XT). Its log is lmxxf_backend.log next to the game.

Compatibility (lmxxf). The runtime sees only what DLSS sees, so what varies per title is a short list: colour format and HDR, motion vectors and their scale, depth and its direction, the reactive mask, the exposure texture, the Reset flag, and where the pass sits (before Super Resolution, or after Ray Reconstruction). Tested so far:

If a title shows no effect: lmxxf_backend.log has an lmxxf inputs: line (formats, sizes, motion scale, depth direction, mask, exposure) and an lmxxf stats @N: line every 600 frames (exposure, fed brightness, the model's edit, the carried edit, keep, reactive mean, vector length and rejected fraction). Attach the log to a report; those two lines usually say why.

Both runtimes share one Neural tab (see "The menu" below). Controls the running runtime does not have are greyed with a short tag or hidden with a count. lmxxf-only: Full network, Output smoothing (Quality > More quality options, needs Network history), Edit detail, Edit colour and Edge guard (Image look > Model strength: gain on the fine part of the model's edit, its colour against its brightness change, and a fade of the edit across depth edges) and the auto-exposure highlight cap. New on lmxxf in 0.3.4: Network output, Encoding, Residual edge fade, Game exposure, Fast mode, the interleave pacing readout, and the model's native character mask with Structure intensity and Character structure (each change rebuilds the network: about a 1 s hitch).

Full network (Neural > Performance, [DlssNr] LmxxfFullNetwork, lmxxf only) runs all 71 of the network's blocks instead of skipping 42, 43 and 46: slightly more faithful, about 0.5 ms slower at 1080p (16.6 -> 17.1 ms on an RX 9070 XT, measured in 0.3.3). Off by default.

Fast mode (Neural > Performance, [DlssNr] AmdLmxxfFastMode, lmxxf, opt-in, off by default) runs the network one size tier lower (1080 -> 900, 900 -> 720): about 29% less network time at 1080p (RX 9070 XT, measured outside a game), with fine detail a little softer. danielblnc builds that have their own Fast mode get a Fast mode row there too ([DlssNr] AmdDanielFastMode); the runtimes in this release's runtime zips do not have it, so the row is hidden.

lmxxf's network runs at a few fixed sizes (tiers): 720 (1280x720), 900 (1600x900) and 1080 (1920x1080), plus 576 and 360 (new, used on handhelds). A tier costs the same whatever part of it the picture fills. On RDNA 3 (RX 7000, Radeon 8060S / 8050S and the handheld APUs) lmxxf's NR size now moves onto a tier by default: down to the next smaller tier when it is nearer to it (cheaper), else grown to fill its own tier (same cost, a little more detail), never above the frame's own size.

Network time per run on an RX 7800 XT (measured by a tester with lmxxf's probe; network only, mean of 30 runs; the 900-tier time was measured at 1600x900):

• In game the gain per displayed frame is smaller: with Model interleave the network runs every 2nd frame, and the game has its own cost. Not yet measured in a game.
• The network sees a slightly smaller picture (at 1440p Quality about 6% fewer pixels per side), so fine detail can be a little softer. [DlssNr] AmdLmxxfTierSnap=false restores 0.3.3.2's sizes. RX 9000 keeps 0.3.3.2's sizes unless you set it to true.
• RX 9000: [DlssNr] AmdLmxxfTierSnap=true (off by default there) moves lmxxf's NR size to a network size at every render resolution: some sizes drop a tier (1440p FSR Quality, 1707x960 -> the 900 size: network 14.08 -> 9.96 ms per run on an RX 9070 XT, measured outside a game, a slightly softer image), others grow inside their tier (80% of a 1080p render -> 1600x900: same cost, a little more detail). Left unset, RX 9000 keeps 0.3.3.2's sizes.
• Start at NR resolution 70% or lower (the 720 tier at a 1080p render; the Performance preset is 70%). The cost beside NR resolution is priced by the tier the network runs; its hover names the tier.
• 0.3.5: about 10 percent less network time on RX 7000, same picture bit for bit: the 0.3.5 module set in LmxxfNrRuntime.pak (measured and hash-checked by a tester on an RX 7800 XT; the table above is 0.3.4's).
• 0.3.5 on RX 9000: lmxxf 0.37 by Kien (MIT) is on by default - about 20 percent less network time on an RX 9070 XT at the 1080 size (14.0 -> about 11.3 ms, measured outside a game; the picture is not bit-identical to 0.3.4.2's). On the RX 9060 / 9060 XT it is on too, together with the c32w and FastK kernels, as an experiment (not yet run on that card). Off switches: [DlssNr] AmdLmxxfL37=false, AmdLmxxfC32w=false, AmdLmxxfFastK=false.

lmxxf runs on handheld APUs with 12 or more compute units, through AMDNR's RDNA 3 backend by 3zwr1: Z1 Extreme, Z2 and Radeon 780M (gfx1103), Z2 Extreme, Radeon 890M and 880M (gfx1150). It is experimental and slow. The Neural runtime row says "experimental" after the RDNA 3 credit. First tester results (ROG Ally, Z1 Extreme): lmxxf's probe outside a game, 54.7 ms per network run at the 360p size, 110.9 ms at 576p; in a game, one tester's run (Shadow of the Tomb Raider, 1280x720 with XeSS, Handheld preset), 62 ms per network run on average at 360p with the model every 4th frame, about 29 fps with NR on. 0.3.5 takes about 10 percent off those numbers on the Z1 Extreme class (Z1 Extreme, Z2, Radeon 780M: about 50 ms at 360p, about 105 ms at 576p, same picture bit for bit, hash-checked by a tester); the Z2 Extreme / 890M / 880M modules are unchanged.

• Not supported: Z1 and Radeon 740M (4 compute units), Radeon 760M (8), Radeon 860M / 840M. danielblnc's runtime does not run on handheld APUs. RX 6000 (RDNA 2) is planned for 0.3.6; the Steam Deck and other RDNA 2 APUs are not supported.
• What it does by itself (only while your ini has no value of its own): the network runs at its smallest size, 360p (640x360), and the model runs every 4th frame (Model interleave; not saved). Neural passes stay at 1.
• Speed, honestly: For scale: an RX 7800 XT (60 compute units) needs 34.4 ms per network run at the 720 size; these chips have 12 to 16 and run at lower clocks. Expect a large frame-rate cost even at 360p with the model every 4th frame, some ghosting from the long interleave, and a softer look than on a desktop GPU. The NR cost at the end of the Neural tab's status line (and in Diagnostics) shows the real number on your device.
• Settings:
  • Sharper but slower: [DlssNr] AmdLmxxfTierCap=576 (the 1024x576 network size).
  • At a 720p or 800p render, NR resolution 100% already feeds the 360p size, so a lower NR resolution does not make it cheaper.
  • Model interleave Off is saved as [DlssNr] AmdInterleave=1 (also off), so the handheld default does not come back at the next start. To turn it off by hand, write 1, not 0.
  • Preset > Handheld sets NR resolution 100%, Dynamic NR off, the model every 4th frame, 1 Neural pass and Full network off. The button shows only on these APUs; Quality, Balanced and Performance keep the 360p network size here too (the menu says so).
• FSR 4: FSR 4 (INT8) is available as an experimental opt-in on RDNA 3 handhelds (Upscaling tab); not validated by AMD. The box FSR 4 (INT8) - Experimental on this GPU (restart) writes [FSR] Fsr4ForceModel=2; it is never switched on by itself (with Dx12Upscaler=auto the upscaler is XeSS). About 1.5-3 ms per frame on a Z1 Extreme (an estimate); with NR, FSR 4 or the 576 size, not both.
• Shadow of the Tomb Raider (and games that make their D3D12 device twice) no longer crash when the upscaler starts (fixed in 0.3.4).
• Driver: use AMD's own Adrenalin driver. lmxxf needs HIP (amdhip64_7.dll), which some handheld makers' drivers leave out; amd_bridge.log then says HIP is not available.
• Moving with Model interleave (fixed in 0.3.5): the carried edit used to vanish as soon as you moved ("the effect is gone when moving"): the network runs on frames of unequal length under Model interleave, and the carry's guard assumed equal frames. It now reads the previous frame's vectors at this frame's duration (both runtimes).
• Use the 0.3.5 files together: 0.3.5 changes all three files (OptiScaler.dll, LmxxfNrRuntime.dll and LmxxfNrRuntime.pak), so replace them together; the runtime refuses a handheld when OptiScaler.dll is older than 0.3.4 ("this handheld needs OptiScaler.dll 0.3.4 or newer").
• Testers with a handheld: ask on Discord for the handheld test kit (handheld-test.zip). Its run_probe.bat measures the network on your device and writes handheld_result.txt (your Windows user name is masked).

What it is. Neural Rendering for games that have no DLSS, XeSS or FSR 2 of their own - and nothing is written into the game folder. AMDNR runs inside a window-capture host: the host captures the game's window, scales it to your screen with FSR 3, and Neural Rendering runs on the captured picture; our menu draws inside the host (your menu key, INSERT by default) with its own Anywhere tab. The host is Magpie by Blinue, experimental fork by SAOG0721 (GPL-3.0), fetched from the author, not redistributed by AMDNR: the AMDNR Launcher downloads it from its author's release (467 MB, once).

How to use it. In the AMDNR Launcher a game with no upscaler shows PLAY ANYWHERE instead of INSTALL. Press it: the launcher fetches the host (the first time), starts the game, and the host captures its window. Run the game windowed or borderless, not exclusive fullscreen, and press your menu key for the AMDNR menu inside the host. A game with an upscaler of its own keeps the normal INSTALL route: Anywhere is for the games that have none.

The host settings live in the menu, in the Anywhere tab's Host settings, not in the launcher: the game window size (720p / 900p / 1080p - advice on what to set in the game; the host captures whatever window the game opens), the stage list (V1: one FSR 3 pass to the screen; V2: FSR 3 at 1x, then a fill pass), the NR tier (Auto / 720 / 900 / 1080), VRR, frame pacing and the host frame rate (Default = your display's refresh rate, at most 60; Auto = the rate the network sustained in the last session; 30 to 120; Display refresh = no cap). They apply at the next PLAY ANYWHERE, and the launcher shows a read-only summary beside the button. In the host's own OptiScaler.ini they are [DlssNr] AnywhereWindow, AnywhereEffect, AnywhereNrTier, AnywhereVrr, AnywherePacing and AnywhereHostFps. Cap the game too, with its own limiter, at 60 to 90 fps: the host can only show frames the game drew, and every host frame runs the network once.

What the host cannot give the network. A captured window has no depth, motion vectors, jitter or exposure of its own; the host estimates the motion. So the Anywhere tab names what is estimated; the Neural tab's rows that cannot act there are hidden or refused with a reason (Ray Regeneration, Screen GI, and Model interleave - it would spend the carried edit on estimated motion); and a status line on the Anywhere page says when the network is over the host's frame budget and what to lower. A 1920x1080 or smaller game window is the pixel-exact case: a larger window is brought down to the network's ceiling first and upscaled back, and the page says so, with the share of the screen's pixels the network saw. The NR resolution row shows the network's real size inside the host.

Status: preview. RX 9000 (RDNA 4) only in this release; RX 7000 follows once tested there. A light flicker or judder in fast motion can remain (lower the NR tier to 720, cap the game at 60-90 fps, leave Model interleave off - the host refuses it). The capture host is fetched from its author's GitHub release, not from ours. Reports: the Save report zip from the menu inside the host (its title names the scaled game), or the launcher's COLLECT LOGS.

AMDNR runs under Proton and Wine as an OptiScaler build. Neural Rendering does not run on Linux (Windows only): both NR runtimes need the Windows AMD driver's HIP, which Proton and Wine do not provide. With NR switched on under Proton, NR does not run, and since 0.3.5 the Neural tab and the report say so (instead of "Idle"). That is expected, not a crash or a broken install. The AMDNR Launcher is a Windows program that can run under Proton (experimental, not tested by us yet, see below); installing by hand works without it.

What works: the FSR upscalers (FSR 3.1, and FSR 4 on GPUs and drivers that support it), the menu (INSERT) and Save report. A player confirmed on Steam Proton (RX 9070 XT, vkd3d-proton, Resident Evil Requiem) that the game boots, the menu opens and takes the mouse, and Save report works, also with frame generation on.

Ray Regeneration and frame generation:

• Known issue: Ray Regeneration can show pink / magenta patches on Proton; the default sharpening is off there now, but if you still see them use plain FSR (FSR 4 on RX 9000) and send a Save report. On Proton, AMDNR adds no sharpening after Ray Regeneration when the game sends none (on Windows it adds 0.25): Image > Sharpness shows "RR default 0 (off on Proton)", and Override still sets your own value.
• Frame generation now turns on without crashing, but fps counters count the generated frames too: under a 60 fps cap or 60 Hz V-Sync that is 30 real frames, which looks like 30. Leave it off on Proton for now ([FrameGen] FGOutput=nofg), or use it only when the game reaches about 60 fps without it, on a screen faster than 60 Hz.
• Vulkan titles (RTX Remix games, id Tech 8), on Proton as on Windows: Ray Reconstruction and AMDNR's own frame generation are answered "not supported" by design (the denoiser is D3D12, and the ray-tracing buffers stay on the Vulkan device); since 0.3.5 the Ray Regeneration and Frame Gen tabs say so instead of asking you to turn them on in a game that greys them. DLSS super resolution works through the bridge.
• Since 0.3.5, [Spoofing] Dxgi=true under Proton (the FSR 4 upgrade path) no longer faults at start.

Requirements: a current Proton or Wine (tested: Proton 11, which is Wine 11), with the game on vkd3d-proton (D3D12) or DXVK (D3D11), which is Proton's default. Older versions are not tested.

The AMDNR Launcher on Linux (experimental, not tested by us yet). The launcher is the same Windows program, AMDNR-Launcher.exe (self-contained: no .NET or other runtime to install). Under Wine / Proton it detects Wine and shows a notice with what to do. It also looks for your Linux Steam library through Wine's Z: drive (~/.steam/steam and ~/.local/share/Steam, and the library folders in libraryfolders.vdf). We have not tested this ourselves yet: if you try it, tell us on Discord whether it works. To try it:

1. In Steam, add AMDNR-Launcher.exe as a non-Steam game (Games > Add a Non-Steam Game to My Library).
2. In its Properties > Compatibility, force a Proton version (Proton Experimental), then start it from Steam.
3. If the game is not in LIBRARY, press ADD and choose the game's folder (your Linux folders are on the Z: drive); if the launcher picks the wrong .exe, use CHOOSE GAME .EXE.
4. Select the game and press INSTALL.
5. In the game's Properties > General > Launch Options, enter WINEDLLOVERRIDES="dxgi=n,b" %command% (if the launcher used another DLL name for the game, put that name in place of dxgi), then continue with steps 5 and 6 of the install by hand below (start the game from Steam; PLAY in the launcher is off under Wine / Proton).

Install by hand (without the launcher):

1. Download AMDNR-vX.X.X.zip from the releases page (for 0.3.5: AMDNR-v0.3.5.zip). The danielblnc runtime zips (v0.5.0-Runtime.zip and the others) are only used by Neural Rendering, so you do not need them on Linux (copying one does no harm).

2. Extract the zip and copy everything into the game folder, next to the game's .exe.

3. Rename OptiScaler.dll to dxgi.dll.

4. In Steam, open the game's Properties > General > Launch Options and enter:

   WINEDLLOVERRIDES="dxgi=n,b" %command%
   

   This tells Wine to load the dxgi.dll in the game folder instead of its own; without it AMDNR does not load. If you used another name (for example winmm.dll or version.dll), put that name in place of dxgi, e.g. WINEDLLOVERRIDES="winmm=n,b" %command%. Lutris, Heroic and Bottles: add the same override (dxgi = native,builtin) in the runner's DLL overrides or environment settings.

5. In OptiScaler.ini, set [FrameGen] FGOutput=nofg (frame generation off, see above).

6. Start the game and press INSERT to open the menu. Set up the upscaler there.

The menu. In 0.3.4, with frame generation on, the menu could open without taking mouse or keyboard input, or not open at all. 0.3.4.1 attaches the menu to the game window; a player confirmed on Proton that the menu opens and takes the mouse, also with frame generation on. If it still happens on your setup, AMDNR shows the warning "Menu window lost". Then, in OptiScaler.ini, set [FrameGen] FGOutput=nofg; if the menu still does not respond, also set [Menu] OverlayMenu=false (the classic menu, which does not depend on the overlay window).

HDR. AMDNR does not switch HDR on under Proton. HDR comes from your Proton and desktop setup: a Proton build with HDR support and a session that can show HDR (for example gamescope, or a Wayland desktop with HDR on). If HDR works in the game without AMDNR, it keeps working with AMDNR; if the game's HDR option is greyed out, the fix is in your Proton or desktop setup.

Reporting a Linux problem: use the menu's Save report button (keep [Log] LogToFile=true, the default, so the report has this session's log); the report shows whether the game ran under Wine/Proton, vkd3d-proton or DXVK. Please add your distribution, GPU, Mesa version and Proton version.

• Windows 10 or 11 (64-bit) for Neural Rendering. The AMDNR Launcher is a Windows program that can run under Proton (experimental, not tested by us yet). Under Linux / Proton AMDNR runs as an OptiScaler build without NR (see "Linux / Proton").

• An AMD GPU with AMD Software: Adrenalin Edition 26.9.1 or newer. The neural runtime uses HIP through the driver; no HIP SDK and no developer mode are needed. Which chips:

  • RX 9000 (RDNA 4): both runtimes.
  • RX 7000 (RDNA 3, desktop and mobile): both runtimes - lmxxf through AMDNR's RDNA 3 backend, slower than on RDNA 4 (the network size tier is on by default, see above).
  • Strix Halo (Radeon 8060S / 8050S): lmxxf.
  • Handheld APUs with 12+ compute units (Z1 Extreme / Z2 / 780M, Z2 Extreme / 890M / 880M): lmxxf, experimental and slow. Z1 (4 CU), 760M / 740M and 860M / 840M: not supported.
  • RX 6000 (RDNA 2): not supported yet, planned for 0.3.6. Steam Deck and RDNA 2 APUs: not supported (for Neural Rendering; for the upscalers under Proton see "Linux / Proton").

  The Neural tab says what your GPU can run (hover the runtime entries, or the GPU line in Diagnostics).

• AMDNR Anywhere (preview): Windows, an RX 9000 (RDNA 4) card in this release, and the AMDNR Launcher, which fetches the capture host; the game runs windowed or borderless. See "AMDNR Anywhere".

• A Direct3D 12, Direct3D 11 or Vulkan game. The AMD neural path itself is D3D12; D3D11 and Vulkan titles reach it through OptiScaler's D3D12 bridge, which means the upscaler must be one of the "w/Dx12" backends (ffx_12). Leave Dx11Upscaler / VulkanUpscaler on auto and this build picks it for you when neural rendering is on. With Neural Rendering on, the Upscaling list names them "... w/Dx12 - Neural".

• About 2 GB of spare VRAM at 1080p-class render resolutions.

AMDNR-vX.X.X.zip

The danielblnc runtime zips (DLSS-NR on AMD by Daniel Blanco, unmodified, with his permission; use one)

Which one: on RX 9000 and RX 7000 alike, v0.5.0-Runtime.zip (recommended) from Alpha0.3.4.2; v0.4.3-Runtime.zip (Alpha0.3.4.2 and Alpha0.3.4.1), v0.4.1-Runtime.zip and v0.4.0-Runtime.zip (Alpha0.3.4.1) are still accepted. The lmxxf runtime needs no runtime zip on RX 7000 and RX 9000; handheld APUs use lmxxf only. The AMDNR Launcher offers 0.5.0 (recommended), 0.4.3, 0.4.1 and 0.4.0, and picks it for you.

danielblnc 0.5.0 is the recommended danielblnc runtime since 0.3.5 (it runs on 0.3.4.2 too). 0.3.5 also passes your own Async (or the older Inline) key from dlssnr_on_amd.ini through to it instead of forcing the same-frame mode; with neither key a default install is unchanged. A danielblnc build newer than 0.5.0 is not driven by this release.

Each one holds:

Press INSERT. Every tab has the same look: text tabs, a header row with Discord and GitHub (it opens this page), a credits row (Daniel Blanco's name opens his GitHub page), the Components line (how many of OptiScaler's seven components are active; click it for the list), and a footer with Menu Scale, Save Settings and Close. Help opens when you hover a control's label.

The Neural tab, top to bottom:

• Enable Neural Rendering and its key (the button, e.g. Home: click it, then press another key to rebind).
• Neural runtime (danielblnc / lmxxf, with the exact version of your files, e.g. lmxxf 0.3.4) with one state word: running, restart the game to switch, not installed, not for this GPU, or stopped. Under it the running runtime's credit and one status line, e.g. Running - 1920x1080 at 100% - NR 62/s - model 62/s - 15.3 ms (the last number is the NR cost), and a closed Live row with more detail. When something needs you, one orange line follows, with a button when there is a fix (Retry lmxxf, Switch to danielblnc, Open Upscaling). In the default state there is none.
• Preset: Quality / Balanced / Performance set NR resolution to 100 / 85 / 70% and turn Dynamic NR off; nothing else. On handheld APUs a fourth button, Handheld (see "Handheld APUs"). NR style, and Style slots (Store / Apply / Clear).
• Performance: Placement (before / after upscaling, new in 0.3.5; see "Settings worth knowing"), NR resolution (%) with its cost, Neural passes, Full network, Fast mode, Dynamic NR resolution, Model interleave (Interleave preset and the pacing line appear under it while it is on).
• Quality: Residual strength, Residual limit, Temporal stability, Sharpening (CAS), and More quality options (Network history - one checkbox for both runtimes -, Output smoothing, Stability mode, Residual temporal, Residual edge fade, Still-surface steadiness).
• Image look: Colour composition, Detail and Colour strength, and three folds: Model strength (Tone and Structure intensity, Character structure, Edit detail / colour, Edge guard, Native character mask, and danielblnc's Network style, Tone curve and Black lift), Exposure and highlights (Auto-exposure, its highlight cap, Highlight colour guard, Game exposure) and Appearance filter (its off / on word after the name). A dim "default" or "custom" after a fold's name says whether you changed something inside it.
• Ray Regeneration: since 0.3.5 a pointer. While Ray Regeneration runs in the title, its controls are on their own Ray Regeneration tab (right after Upscaling) and this line offers an Open Ray Regeneration button; while it is not running, the line says why (the game has not turned Ray Reconstruction on, the driver refused the denoiser on this card, Ray Regeneration gave this title up and why, or when it last ran).
• The tools row, closed at start: Diagnostics (Network output, Debug view, Edit shaper A/B, NR cost, the ghost and self-tuning readouts, the GPU line, Save report; the RR debug view is on the Ray Regeneration tab since 0.3.5), Runtime options (Encoding, Every-frame NR, NR slots, Highlight proxy) and Experimental (AMDNR Screen-space GI, a preview).

A control the running runtime does not have is greyed with a short tag (e.g. "not in lmxxf yet") or hidden with a count ("3 danielblnc-only options hidden"); switching runtime moves no other row.

The other tabs: Upscaling starts with the upscaler, one status line and Render resolution (the former Upscale Ratio Override and Output Scaling); on a non-NVIDIA card "DLSS w/Dx12" is no longer listed. Ray Regeneration (new in 0.3.5) follows Upscaling while Ray Regeneration runs in the title: the status lines (per card and per API), the Denoiser backend row (Automatic / Off - Off tells the game Ray Reconstruction is unsupported, so it keeps its own denoiser; after a restart), the controls, More Ray Regeneration options, and its own Diagnostics block with the RR debug view and the noise number (grain in and out, still-camera flicker); it is never drawn inside AMDNR Anywhere. Image holds Sharpness, Textures, Init Flags and the Magnifier. Frame Gen opens with FG Input and FG Output and, since 0.3.5, one line that names the step still missing before anything generates. Interface has the FPS overlay and Keybinds (one button per key). Advanced starts with Active Quirks, then Display (V-Sync), Compatibility and Logging. Inside AMDNR Anywhere the menu shows an Anywhere tab (the capture and network line, the host settings) in place of the Frame Gen and Advanced tabs. The settings, keys and what Save Settings writes are unchanged, except where CHANGELOG.md says so.

Open the Neural tab. The defaults are the most recent tested arrangement, so the useful first move is to change one thing at a time.

• NR resolution — the main quality/cost lever. Below 100% the model works on a smaller picture and only its correction is carried back up to the full-resolution frame, so the frame keeps its own detail. Above 100% cost grows with the square (150% is 2.25x). The slider moves in 5% steps: each new NR size can keep VRAM until the game restarts, so restart the game after many changes. The cost beside it reads 1.00x at 100%; on lmxxf it is the price of the network size tier it runs (its hover names the tier). The Preset buttons set it to 100 / 85 / 70%.
• Placement (Neural > Performance, [DlssNr] AmdPlacement = pre | post, new in 0.3.5, both runtimes) — where the neural pass runs. pre (the default, and what every earlier build did) edits the render-resolution picture the upscaler is about to read. post edits the upscaler's finished display-resolution picture instead: sharper, because the upscaler no longer re-filters the edit, and more expensive - the network runs at display size up to its 1920x1080 ceiling, so a 1080p screen pays the top tier at any NR resolution, and a 1440p or 4K screen gets a 1080-line edit lifted back - a little less forgiving in motion, and the HUD is included if the game composites it before upscaling. Refused (back to pre, one line in amd_bridge.log) inside AMDNR Anywhere, in final image mode, and once Ray Regeneration has run in the title. The NR resolution row shows both sizes while post runs.
• Residual strength — how much of the model's edit is applied; above 1 it amplifies. This is the control that changes the picture most.
• Residual limit — a ceiling on how far one pixel may move. Blotchy patches: lower it.
• Model interleave — runs the model every second frame for a large frame-rate gain. The skipped frames are filled by the Interleave preset; Edit accumulation (preset 10, both runtimes) is the default: every frame is that frame's own picture plus the model's carried correction, so no picture is held over. Guided fill v2 (preset 6, danielblnc) and Classic carry (lmxxf) are the older fills. Pacing of the two frame types is automatic on danielblnc and off on lmxxf ([DlssNr] AmdInterleavePacing 0..1 paces both, at a frame-rate cost); a dim line under the preset shows the measurement. Adaptive interleave is switched off in this build.
• Neural passes — 2 and 3 stack the model, with diminishing returns. Under lmxxf the network's history stays its first pass; the extra passes are spatial refinement only. danielblnc runs 1 pass on Vulkan titles (a note under the slider says so).
• Colour composition (Neural > Image look, both runtimes) — Classic (default) is the picture you had before. RenoDX (experimental) runs RenoDX's colour composition after the model, as the NVIDIA path does: Composition detail and colour, a two-sided Highlight guard (2x by default) that bounds the model's answer against the original, and optional skin / environment controls. On a display-referred (SDR) frame, with Network output, or with Encoding sRGB / Gamma 2.2 it falls back to Classic on both runtimes; the menu's note then offers a button that turns the blocker off. NR styles and presets leave it alone.
• Native character mask (Image look > Model strength, [DlssNr] AutoMask, on by default) — the model's own treatment of faces and skin. Unticking it now acts on both runtimes (on lmxxf it rebuilds the network: about a 1 s hitch); on lmxxf, Structure intensity and Character structure now act too.
• Frame generation is off in a fresh ini, and it takes five steps: FG Input and FG Output on the Frame Gen tab (e.g. "DLSSG via Streamline" in a game with DLSS frame generation, and XeFG), Save Settings, a full restart of the game, the game's own frame generation switched on, then Active ticked under Frame Generation. Since 0.3.5 the tab and the log name the step that is missing. The whole list, with what to switch off in the game, is in the FAQ below ("Frame generation: no fps gain?").
• XeFG multi-frame generation — 3X to 6X is built in and on by default (XeFG\UnlockMFG), for OptiScaler's copy and the game's own. Delete XeFGUnlock.asi from OptiScaler\plugins if you still have it: two copies of the same patch crash the game. Up to 10X is opt-in (D3D12 games only): set XeFG ceiling (restart) under FG Output in the Frame Gen tab (4X, 6X default, 8X or 10X; [XeFG] MaxInterpolatedFrames), 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 at most. 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.
• FSR Ray Regeneration — RX 9000 (RDNA 4); on RX 7000 (RDNA 3) only as an experimental opt-in, not supported (see below); only in games that use DLSS Ray Reconstruction (Cyberpunk 2077, Alan Wake 2), with the game running DLSS (spoofing on), ray tracing and Ray Reconstruction enabled in its own settings. Neural Rendering then runs after it, on its output, which costs more: lower the NR resolution if the frame rate drops. Since 0.3.5 its controls are on their own Ray Regeneration tab, right after Upscaling, drawn while Ray Regeneration runs in the title; the Neural tab points at it and, while Ray Regeneration is not running, keeps the dim line that says why (the game has not turned Ray Reconstruction on, the driver refused the denoiser on this card, Ray Regeneration gave this title up and why, or when it last ran). The tab's Denoiser backend row ([FSR-RR] RrBackend = auto | off) can tell the game Ray Reconstruction is unsupported, so it keeps its own denoiser (at the next start of the game). On a Vulkan title Ray Reconstruction is "not supported" by design (the denoiser is D3D12) and the tab says so. The path-traced profile (less grain on faces under path tracing) is opt-in since 0.3.3.1: tick it there to try it in Resident Evil Requiem or PRAGMATA. The same tab has the bias mask strength and skin smoothing (experimental, for games that publish an SSS guide; off by default, but on by default in Resident Evil Requiem since 0.3.3.2); the temporal tuning sliders are under More Ray Regeneration options, and the RR debug view and the noise number (grain in and out, still-camera flicker) are in the tab's own Diagnostics block. On RX 7000 (RDNA 3) Ray Regeneration is not supported and, since 0.3.5, not offered by default: AMD's denoiser has no provider for RDNA 3, so the game keeps its own denoiser. The Upscaling tab's box Experimental: Ray Regeneration on this card (restart) (tagged "experimental - not supported") is for testing only: ticked, the denoiser refuses to start and the game gets FSR without a denoiser, which can look noisier than the game's own. AMDNR's own denoiser for RX 7000 is planned. RX 6000 and older get it only with [FSR-RR] FfxDenoiserAllowPreRdna4=true (Upscaling tab: Offer FSR Ray Regeneration on this GPU (restart)). Sharpening after RR (0.3.4.1): when the game sends no sharpness, AMDNR sharpens by 0.25 after RR on Windows (0 on Linux / Proton); to turn it off: Image > Sharpness, tick Override, slider 0. Since 0.3.4.2 that number is its own ini key, [Sharpness] RrDefaultSharpness (same 0.25 default): set 0.15, 0.10 or 0 there without touching Override, and a value your ini kept under [Sharpness] Sharpness while Override is off is tagged in the menu as waiting.
• AMDNR Screen GI (preview, new in 0.3.4, off by default; Neural > Experimental, or [AmdGi] Enabled=true) — AMDNR's own screen-space bounce light and ambient occlusion, from the game's depth, before NR and the upscaler; works with NR on or off; about 1 ms at High for a 1080p render on an RX 9070 XT (measured outside a game). Screen-space: light from off-screen is missing. See CHANGELOG.md.
• Save report (Neural > Diagnostics, or Advanced > Logging) — one zip with every log and the ini files for a bug report; see "If it doesn't work" above.

OptiScaler.log appears in the game folder. Attach it in #bug-report, and say which game and which GPU; Save report (Neural > Diagnostics, or Advanced > Logging) zips it with everything else. The AMD backend also writes amd_presr.log and amd_bridge.log, which are the useful ones when the neural pass specifically misbehaves. The last three sessions' logs are kept as OptiScaler.previous.<exe>.log (newest), OptiScaler.previous-1.<exe>.log and OptiScaler.previous-2.<exe>.log ([Log] KeepPreviousLogs, 1 keeps only one as before). After a crash, attach them too: the new log then says "no clean exit recorded" (since 0.3.4 no longer after a normal quit).

NR frames 0/s, and the Neural tab or amd_presr.log says the pass DLL is a build this AMDNR does not drive? Your dlssnr_amd_pass1..3.dll are a danielblnc build this AMDNR does not know (a 0.2.16 set was seen in the wild), or one of the three is missing. Since 0.3.3.2 the Neural tab names the file and its version and says what to do. Use the recommended runtime, all three pass DLLs from the same zip: on RX 9000 and RX 7000 alike, v0.5.0-Runtime.zip from Alpha0.3.4.2 (149,550,553 bytes, SHA256 starting 7a49ab0e); v0.4.3-Runtime.zip (Alpha0.3.4.2 and Alpha0.3.4.1; 116,484,918 bytes, SHA256 starting 07dd7774), v0.4.1-Runtime.zip and v0.4.0-Runtime.zip (Alpha0.3.4.1) are still accepted (the dlssnr_amd_pass1.dll in v0.4.1-Runtime.zip is 9,916,928 bytes, SHA256 starting 823063eb; in v0.4.0-Runtime.zip: 10,027,008 bytes, d62be3d8). Supported builds: 0.2.17, 0.3.0, 0.3.1, 0.3.2, 0.3.3, 0.4.0 and the runtime zips named above, up to 0.5.0. Do not install danielblnc's own setup or its dxgi.dll / version.dll / winhttp.dll next to AMDNR: AMDNR already runs his runtime. A danielblnc build newer than 0.5.0 is not driven by this release: the Neural tab names the file and its version and says so. With 0.5.0, 0.4.3, 0.4.1 and 0.4.0 the danielblnc-only settings (Network style, Tone curve, Black lift, Game exposure, Fast mode) work, and since 0.3.5 your own Async key in dlssnr_on_amd.ini reaches the runtime (see "What is in the archives").

lmxxf does nothing, or stops at once, on a PC with integrated graphics? Fixed in 0.3.3.2. On a Ryzen desktop with its integrated graphics on, a laptop with an AMD APU and a Radeon, or a PC with two AMD GPUs, the game's GPU is often not HIP device 0. lmxxf then failed on its first frame (hipErrorInvalidHandle (400), then "session is poisoned" in lmxxf_backend.log) and stayed off. Replace both OptiScaler.dll (the file you renamed, e.g. dxgi.dll) and LmxxfNrRuntime.dll with the files of 0.3.3.2 or newer. Not yet tested on such a PC: if lmxxf still stops, the Neural tab now says why; send lmxxf_backend.log and amd_bridge.log (it lists the HIP devices).

A PC with integrated graphics and a Radeon (a Ryzen desktop with its integrated GPU on, or a laptop): NR never starts, and the Neural tab or the GPU line names the integrated GPU? AMDNR's neural pass runs on the GPU the game draws with. If Windows started the game on the integrated GPU, NR does not run on your Radeon at all. Set the game to the discrete GPU: Windows Settings > System > Display > Graphics, add the game's .exe, Options, High performance; then restart the game and check the GPU line in Neural > Diagnostics, which names the adapter NR runs on (OptiScaler.log has an AMD neural: NR runs on ... line when that adapter is not the primary GPU). Seen in Starfield on a Ryzen desktop. Since 0.3.5 the Neural tab's line says which of three cases it is - no runtime yet, NR running on the integrated GPU (an APU a runtime accepts, such as a Radeon 780M beside a Radeon card: NR runs there, far slower than on the card), or a runtime on another adapter - and amd_bridge.log lists every adapter once.

lmxxf's status line says c32w=off:nofile on an RX 9070 / 9070 XT? An old DLSS5-AMD\native-game-tiled-assets folder next to the game's .exe (left from an earlier lmxxf setup) is used instead of LmxxfNrRuntime.pak. It has no c32w kernels, so lmxxf runs at the old speed. Remove or rename the DLSS5-AMD folder: the pak holds everything lmxxf needs. A LmxxfNrRuntime.pak older than 0.3.3.2 shows the same status; replace it with the one in this release. fk=fff- in the same line means the same (an old pak or a loose folder): lmxxf still runs, at the old speed.

danielblnc: the NR style still changes when NR resolution leaves 100%? Still open in 0.3.4 to 0.3.5, and the default is unchanged. At 100% Residual strength 0.99 gives 99% of 1.00 (fixed in 0.3.3.2); away from 100% (also Dynamic NR steps and the Balanced / Performance presets) strength, limit and edge fade still act on the whole result, so the look can change. 0.3.4 adds an A/B to find the right fix: Neural > Diagnostics > Edit shaper (A/B, not saved) with Literal, F1 and F2, plus Only below 100% and Carry cap (danielblnc only; Save Settings does not keep it; the ini keys are [DlssNr] AmdEditShaper, AmdEditShaperLimit, AmdEditShaperScope and AmdEditShaperCarryCap). If one of them makes 85% look like 100% in your game, tell us on Discord with screenshots. lmxxf is not affected.

The menu opened and closed twice per key press, or keyboard and mouse stopped working on the whole desktop while the menu was open (Assetto Corsa)? Fixed in 0.3.4: a second menu or NR key press within 400 ms is ignored ([Hotfix] MenuToggleDebounceMs, 0 = the old behaviour), and while the menu is open a game's low-level keyboard or mouse hook is skipped but the key still reaches Windows ([Hotfix] MenuLowLevelHookPassThrough=false = the old behaviour). Not yet confirmed in Assetto Corsa: send the report zip if it still happens.

The menu opened by itself on the runtime chooser, clicks did nothing, or the menu key hid the menu only while it was held (Assetto Corsa)? Fixed in 0.3.4.2: the menu key toggles once per physical press (a key message that arrives late is ignored), clicks and menu keys shorter than a frame are replayed, the runtime chooser answers to 1 / 2 / Enter / Esc and its title-bar X, and closing the menu counts as Decide later; the chooser no longer opens the menu by itself. Not yet confirmed by the Assetto Corsa player: send the report zip if it still happens. To get 0.3.4.1's menu key and clicks back: add DiagInputHooksSkip=presslatch,clickreplay under [Hotfix] in your OptiScaler.ini yourself (no new key; the shipped ini only describes the line in a comment). 0.3.5 adds two things for Assetto Corsa: AMDNR steps aside from a foreign nvngx.dll in the game folder and guards D3D11On12 device creation, and a DX11 game that never presents through D3D12 gets a bootstrap D3D12 queue for the neural pass ([DlssNr] AmdBootstrapQueue, auto). Not yet confirmed by an Assetto Corsa player.

Uncharted: Legacy of Thieves Collection crashed a few seconds after start on RX 9000 with Neural Rendering on? Fixed in 0.3.5: the game runs its work on small 192 KiB fibers, and HIP's first initialisation (the driver compiles its helper kernels inside the game) overflowed that stack at the first NR frame. The neural bridge's first HIP use now runs on its own large stack ([DlssNr] BigStackCall, auto). If you set [DlssNr] Enabled=false there for 0.3.4.2, turn it back on. Not yet confirmed by a player on RX 9000: send the report zip if it still happens.

A game with dynamic resolution (The Last of Us Part II) flickers with Neural Rendering on? Fixed in 0.3.5: the game changed its render size hundreds of times a minute, and every step rebuilt the network and reset its history. A step that stays inside the allocation is now held (no settle, no warm-up, no history reset, both runtimes); a larger frame or a real drop still reallocates. The report's stats line counts the held steps. Not yet confirmed in that game.

Frame generation: no fps gain, or "restart the game" for ever? In nearly every report, frame generation was simply not switched on - off, not broken. It takes five steps, and since 0.3.5 the Frame Gen tab and the log name the one that is missing:

1. Frame Gen tab: pick both FG Input and FG Output. A DX12 game with its own frame generation: its DLSS FG or FSR 3.1 FG is the input (games with FSR 3.1 FG: "FSR 3.1 FG", not "FSR 3.0 FG"); no FG in the game at all: FG Input = OptiFG (Upscaler), HUD fix on. DX11: OptiFG only. Vulkan: no FSR FG / XeFG output (use the game's own).
2. Save Settings.
3. Close the game completely and start it again - FG cannot switch on mid-session.
4. In the game's own graphics options turn its frame generation on: DLSS Frame Generation (with DLSS as the upscaler) for the DLSSG input, FSR frame generation (with FSR) for the FSR 3.1 FG input.
5. Open the menu again, Frame Gen tab, tick Active under Frame Generation. Nothing is generated until that box is ticked (XeFG may ask for one more restart).

Then turn off in the game: exclusive fullscreen (XeFG needs borderless), V-Sync and frame caps (or cap at twice your base fps), and the game's own XeSS frame generation if it has one (one frame generator per window; a game that loads its own XeSS FG gets a note in the tab). No fps gain = FG is off - the log says ... Enabled is off ...: no frames are generated. Frame generation off, not broken.; half the fps = a cap or V-Sync is holding the generated frames back. If it still fails, press Save report after the failure and post the zip with the game, the pair you picked, which in-game frame-generation option was on, borderless or fullscreen, HDR on or off, and what you saw. The full FAQ is pinned in the Discord support channel.

Other mods (Cyberpunk 2077's RED4ext, Cyber Engine Tweaks) or ReShade beside AMDNR? Since 0.3.5 the report and the log name the other mods' proxy loaders (Mod loaders: in report.txt; RED4ext is winmm.dll, Cyber Engine Tweaks is version.dll through an ASI loader) and warn when AMDNR holds a known loader's name of this game unchained. The AMDNR Launcher never takes or moves a loader's file: it picks another proxy name (Cyberpunk 2077: dxgi.dll), and its Doctor names any loader an earlier INSTALL moved aside (AMDNR_backup). ReShade beside AMDNR is detected by the module's version resource or ReShade's add-on exports, never by a file name, and named in the report and the log ([Game] ReShade detected: <module>); nothing is loaded, hooked or blocked, and making the two share the D3D12 device is designed, not built yet.

No HIP adapter matches D3D12 LUID in the Neural tab or amd_bridge.log? Since 0.3.5 the line names both sides - the game's D3D12 adapter (name, LUID), every HIP device (ordinal, name, gfx, LUID) - and the likely cause with what to do: the game runs on the integrated GPU or another card (Windows Settings > System > Display > Graphics: set the game to the discrete GPU), a HIP runtime without identity (a stray amdhip64_7.dll next to the game: remove it), no HIP device at all (install AMD's own Adrenalin driver), the same card under another identity, or not an AMD adapter. Both runtimes write the same text.

A Vulkan game (Indiana Jones and the Great Circle) stops at start with "Could not create the Vulkan device (VK_ERROR_EXTENSION_NOT_PRESENT)"? Fixed in 0.3.2: the inherited NVIDIA neural path asked the AMD driver for two NVIDIA-only device extensions. Vulkan titles reach the neural pass through OptiScaler's D3D12 bridge (see Requirements).

lmxxf froze a Vulkan game at the first NR frame? Fixed in 0.3.3; expect one hitch of about 1 s when NR starts. If a Vulkan session ever stops before lmxxf's first answer, the next start runs danielblnc's runtime and the Neural tab says why; press Retry lmxxf there (it removes lmxxf_vk_launch.pending beside OptiScaler.dll) to try lmxxf again.

danielblnc paused for seconds, then stopped NR, on a Vulkan game (Indiana Jones) with 2-3 Neural passes? Fixed in 0.3.3: on Vulkan titles it runs 1 pass, and its 80 ms post-submit wait is gone. The first NR frame of a session still pauses about 5 s; a note under the runtime choice explains its log lines. Testers: [DlssNr] AmdVkLateCopyWait=true (experimental, off by default, not yet tested in a game) is expected to remove that pause; send OptiScaler.log, amd_presr.log and dlssnr_on_amd.log.

lmxxf's RAM use climbed for as long as NR ran? Fixed in 0.3.3 (it was about 45 GB an hour at 60 NR fps). What remains: danielblnc keeps VRAM for each new NR size above about 1 MP (0.3.3.2 rounds its sizes to 64 px away from 100%, so there are only a few); restart the game after many changes with danielblnc. Since 0.3.3.2 lmxxf no longer keeps about 97 MB per NR resolution or DLSS mode change: it makes its network buffers once per network size and reuses them (a small rest of about 10-25 MB of VRAM per change remains).

A Streamline game fails at start with slInit error 0x18 (seen with NBA 2K27 on AMD)? 0.3.3 closes one way OptiScaler's Streamline plugin hooks could cause it, but that is not confirmed as NBA 2K27's cause. OptiScaler.log now records slInit returned ... and [SLINIT] lines: send the log with the report.

You cannot find the Ray Regeneration settings? Since 0.3.5 they are on their own Ray Regeneration tab, right after Upscaling, drawn while Ray Regeneration runs in the title (and kept for the session once it has run); the Neural tab's Ray Regeneration line then offers an Open Ray Regeneration button. While it is not running, the tab is not drawn and that line in the Neural tab says why: the game has not turned Ray Reconstruction on, Ray Regeneration gave this title up and why, or it last ran N seconds ago. On RX 7000 a further dim line adds that Ray Regeneration is not offered there: AMD's denoiser has no provider for RDNA 3, so the game keeps its own denoiser; an experimental opt-in exists in the Upscaling tab (Experimental: Ray Regeneration on this card (restart)), but it is not supported. On RX 6000 and older it says that it is not offered on this GPU, that AMD ships it for RDNA 4, and that [FSR-RR] FfxDenoiserAllowPreRdna4=true offers it anyway. To get it running, in the game's own graphics settings: pick DLSS as the upscaler (not FSR, not XeSS), turn ray tracing or path tracing on, and turn Ray Reconstruction (DLSS-RR) on; the Upscaling tab then reads "FSR Ray Regeneration". Which setting to change for which problem is in the Ray Regeneration settings guide RR-BEST-SETTINGS.md (not in the zip).

Ray Regeneration looks grainy or noisy? Judge it with Neural Rendering switched off first (untick Enable Neural Rendering at the top of the Neural tab, press Home while you play, or set [DlssNr] Enabled=false): the neural pass runs after Ray Regeneration, on its output, so a screenshot taken with NR on tells you nothing about the denoiser. Then, per kind of grain - crawling grain in a still scene, bright specks, grain on faces, trails behind moving people - the settings to try are in the same guide, RR-BEST-SETTINGS.md. No denoiser default and no sharpening default changed in 0.3.4.2: the numbers are 0.3.4.1's. What did change is that the sharpening AMDNR adds after Ray Regeneration is its own ini key now, [Sharpness] RrDefaultSharpness (same 0.25 default), so 0.15, 0.10 or 0 is an ini edit instead of a new build. Check your ini before you chase grain with the sharpness slider: a value under [Sharpness] Sharpness does nothing while OverrideSharpness is off, and it applies at once the moment you tick Override in the menu - which is why Image > Sharpness now tags it ("ini Sharpness 1.00 waits for Override"). Two things we will not pretend away: a part of the grain is the game's own ray sampling - AMD's denoiser is not built to repair noise that arrives correlated, and a game that offers DLSS Ray Reconstruction switches its own denoiser off and hands us the raw signal - and the grain that crawls in a still scene has a structural cause on our side that no slider removes completely. That one is a known issue. 0.3.5 puts a number on it: the Ray Regeneration tab's Diagnostics block shows the grain in and out and the still-camera flicker (measured while the tab is open), and the tab's Denoiser backend row can be set to Off, which tells the game Ray Reconstruction is unsupported so it keeps its own denoiser (after a restart). AMDNR's own denoiser is 0.3.6 work.

The game's Ray Reconstruction is on but the Neural tab says "Ray Regeneration is off in this title"? The game does not publish what FSR Ray Regeneration needs: its DLSS plugin passes empty camera matrices (Satisfactory), which NVIDIA's Ray Reconstruction treats as optional and FSR Ray Regeneration needs. FSR upscaling runs in its place and NR takes its normal pre-SR position; the Upscaling tab says the same. Since 0.3.4 it stays off for the whole session in an Unreal title with this signature. Turn Ray Reconstruction off in the game and restore the engine's own denoiser settings.

A Ubisoft Anvil game (AC Black Flag Resynced, Shadows, Mirage) shows "DX12 Error 0x80070057"? Those games carry their own XeSS Frame Generation. Since 0.3.5 the Frame Gen tab shows an advice note there (leave the game's own XeSS FG off, or two generators share one window) and AMDNR's XeFG output still runs; the simplest route is the game's own XeSS FG option with AMDNR's frame generation left off. If it still happens, set [FrameGen] Enabled=false and [fakenvapi] ForceXeLL=false and report with the log.

The Last of Us Part I crashes on boot? That is the game's own Streamline init, a known OptiScaler issue: rename sl.common.dll in the game folder to sl.common.dll.bak and pick FSR 3.1 in the game's settings instead of DLSS.

Full notes for every version: CHANGELOG.md (in the zip and in the repository).

• 0.3.5 (this build) — AMDNR Anywhere (preview, RX 9000, through the launcher); the Ray Regeneration tab with status lines per card and per API, the Denoiser backend Off row and the noise number; the neural pass after upscaling (AmdPlacement); the Frame Gen tab and the log say why nothing generates; lmxxf 0.37 by Kien (MIT) on RX 9000; the 0.3.5 module set (about 10 percent faster on RX 7000 and Z1 Extreme-class handhelds, same picture); dynamic-resolution steps held without a network rebuild; the handheld carry fix; the shared runtime folder (AmdRuntimePath); the HIP-adapter refusal that names both sides; other mods' loaders and ReShade named in the report; anti-cheat and crash-reporter processes left alone; fixes for Uncharted (RX 9000), Assetto Corsa, F1 25 and Kingdom Come: Deliverance II (Game Pass), Control Resonant with frame generation, Tainted Grail: The Fall of Avalon and Dead Space, GTA V Enhanced, Half-Life 2 RTX and other Vulkan titles, and the Linux / Proton texts; AMDNR Launcher 0.3.5.1.
• 0.3.4.2 — hotfix: the Assetto Corsa menu (the menu key toggles once per press, clicks shorter than a frame are replayed, the runtime chooser answers to keys and closing the menu counts as Decide later), the runtime chooser no longer opens the menu by itself, the Ray Regeneration section is always in the Neural tab and says why it is not running, one more accepted danielblnc runtime layout, Wine / Proton text corrections, README additions (proxy names, Uncharted, hybrid PCs, AmdLmxxfTierSnap on RX 9000, the two Ray Regeneration FAQ entries); NR byte-identical to 0.3.4.1 except that one accepted-layout row.
• 0.3.4.1 — hotfix: Ray Regeneration sharpening when the game sends none (Windows; none by default on Linux / Proton), Control Resonant's false "Upscaler failed to run!" popup, the Linux / Proton menu attaches to the game window, Save report names vkd3d-proton / DXVK; AMDNR Launcher 0.3.4.1 (nine languages, search, favourites, hide, rename, CHOOSE GAME .EXE, PLAY, a full UNINSTALL); NR unchanged.
• 0.3.4 — the new menu (the Neural tab rebuilt, the same look on every tab, Save report); lmxxf faster on RX 7000 (the network size tier by default) and on RX 9070 / 9070 XT (lmxxf 0.31 kernels); lmxxf on handheld APUs (experimental; new 360p and 576p network sizes); lmxxf gets Network output, Encoding, Residual edge fade, the native character mask and an opt-in Fast mode; AMDNR Screen GI (preview); danielblnc runtime settings (Network style, Tone curve, Black lift, Game exposure) and a highlight colour guard; Ray Regeneration tuning and diagnostics; fixes for Assetto Corsa's menu input, Shadow of the Tomb Raider, Marvel's Midnight Suns and The Last of Us Part II, clean exit and logs.
• 0.3.3.x — lmxxf on RDNA 3 (RX 7000; AMDNR's own backend); RenoDX colour composition (experimental, opt-in) on both runtimes; lmxxf: Full network option, the RAM leak fixed, Vulkan titles fixed (lazy weight upload inside the Vulkan bridge), 0.29 kernels (bit-exact, faster); danielblnc on Vulkan titles: 1 Neural pass, clearer messages, an opt-in late copy wait; XeFG up to 10X (opt-in, D3D12); Streamline start-up hardening and diagnostics; FSR Ray Regeneration path-traced profile and skin smoothing; UE5 robustness.
• 0.3.2 — the 0.3.1 reports: Vulkan titles start and run with lmxxf, lmxxf colours matched to danielblnc's (auto-exposure), the runtime combo, Ray Reconstruction status and tuning; Nukem9's dlssg-to-fsr3 in the zip for frame generation on Vulkan.
• 0.3.1 — fixes from the first 0.3.0 reports (lmxxf alone never ran, Where Winds Meet's silent NR, the crash on a DLSS-quality change, GTA V Legacy) and NR style presets with three custom slots.
• 0.3.0 — the lmxxf HIP neural runtime (RDNA 4) as a selectable runtime beside danielblnc's, shipped as LmxxfNrRuntime.dll + LmxxfNrRuntime.pak: network history, real Neural passes, the edit shaper, the after-Ray-Regeneration placement, per-title diagnostics and self-healing. Many thanks to TheAutomatic, whose DLSS 5 AMD project work this integration builds on.
• 0.3.6 — RX 6000 (RDNA 2): AMDNR's own kernels behind a hardware gate, with the 360 tier and Model interleave as the preview on Navi 21; the AMDNR denoiser (ARD) preview, a denoiser of AMDNR's own behind the Ray Reconstruction call for the cards AMD's denoiser refuses; AMDNR Anywhere on RX 7000 once tested there.
• 0.4.0 — the 1440p network tier; Ray Regeneration on Vulkan titles through the bridge; cross-adapter Neural Rendering (the game on one card, the network on the AMD card); Anywhere beyond the preview (titles without an upscaler of their own, where OptiScaler supplies the upscaler and the neural pass together).
• Later — AMDNR on any window (the desktop).

This build is a wiring job over other people's work. If you find it useful, the thanks belong upstream.

• TheAutomatic — DLSS 5 AMD project — https://github.com/TheAutomatic/dlss-5-amd-project
• danielblnc — DLSS-NR on AMD by Daniel Blanco — https://github.com/danielblnc/DLSS-NR-on-AMD (the *Runtime.zip files, unmodified)
• lmxxf (Kien) — https://github.com/lmxxf/dlss5-on-amd-9070xt-porting (the network port, kernels and HIP runtime, MIT)
• TheAutomatic — LmxxfNrRuntime.cpp, LmxxfNrApi.h, LmxxfProductionOptions.h: portions contributed to lmxxf by TheAutomatic (MIT)
• lmxxf 0.31 kernels in LmxxfNrRuntime.pak (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)) — lmxxf's (Kien, MIT), built by AMDNR from lmxxf's sources and build recipe; AMDNR's part is the loading, SHA-256 pins, per-GPU gating and fallbacks
• lmxxf 0.37 kernels in LmxxfNrRuntime.pak (the lmxxf037-* modules for RX 9000) and their launch code in LmxxfNrRuntime.dll — lmxxf 0.37 by Kien (MIT), shipped as lmxxf built them; AMDNR's part is the loading as one pinned group, SHA-256 pins, per-GPU defaults, off switches and fallbacks
• c32w kernels (0.3.3.2) — AMDNR's own one-wave RDNA 4 kernels for lmxxf's network, Copyright (c) 2026 3zwr1 (AMDNR); ideas from AMD's public RDNA 4 WMMA docs (GPUOpen, ROCm matrix instruction calculator)
• AMDNR's RDNA 3 backend (0.3.3; the handheld builds gfx1103 / gfx1150 in 0.3.4), the network size tier policy and the small network sizes (0.3.4) — Copyright (c) 2026 3zwr1 (AMDNR)
• Matheus / dlss-5-amd — https://github.com/MatheusGViana/dlss-5-amd-project
• Dagherbou / OptiScaler_DLSSNR — https://github.com/Dagherbou/OptiScaler_DLSSNR
• wilsjo2 / OptiScaler-DLSSNR-PreSR-Multipass — https://github.com/wilsjo2/OptiScaler-DLSSNR-PreSR-Multipass
• Nukem9 — dlssg-to-fsr3 — https://github.com/Nukem9/dlssg-to-fsr3 (GPLv3, unmodified)
• Magpie by Blinue, experimental fork by SAOG0721 (GPL-3.0), fetched from the author, not redistributed by AMDNR — the window-capture host AMDNR Anywhere runs inside — https://github.com/Blinue/Magpie (the fork: https://github.com/SAOG0721/Magpie)
• 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
• Zach Hembree (DarkHelmet) — FSR Ray Regeneration for OptiScaler, the origin of AMDNR's Ray Regeneration path, continued by burak113, whose branch AMDNR ported from (OptiScaler branch ffx-denoise-experimental, GPL-3.0)
• Screen-space GI (the inherited effect; retired from the menu in 0.3.4, [AmdRtgi] Enabled in the ini) — an effect AMDNR inherited from the OptiScaler-AMD-PreSR lineage; the credit belongs to its original authors. It needs the experimental_lighting folder of the danielblnc package, which AMDNR does not ship.
• 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; and the others listed in CHANGELOG.md and Licenses/AMDNR_NOTICE.txt)
• OptiScaler — Overclockers — https://github.com/Overclockers/OptiScaler-Releases

AMDNR Launcher (new in 0.3.4) is a Windows 10 / 11 program that can also run under Proton on Linux (experimental, not tested by us yet: see "Linux / Proton"). It installs and updates AMDNR per game: it finds your games (Steam, Epic, the Xbox app, Ubisoft Connect, the EA app, GOG, Rockstar, Battle.net and Amazon Games), picks the DLL name, downloads the build and the danielblnc runtime you choose, checks every install with its Doctor, runs AMDNR Anywhere for games with no upscaler of their own (PLAY ANYWHERE), and updates itself. Download AMDNR-Launcher.exe, the AMDNR Launcher from the latest release, from the releases page: https://github.com/3zwr1/AMD-NR---OptiScaler/releases

New in Launcher 0.3.5.1 (on the Alpha0.3.5 release; it updates itself first, then your games):

• PLAY ANYWHERE on a game with no upscaler of its own (see "AMDNR Anywhere"): the launcher fetches the capture host from its author's release, starts the game and runs Neural Rendering on its window, with a read-only summary of the host settings beside the button (the settings themselves are in the menu's Anywhere tab). RX 9000 in this release. Games the mod cannot load into (32-bit, DirectX 9 / OpenGL without an upscaler) are refused at INSTALL with a clear line and offered Anywhere.
• UPDATE ALL and update-at-start; package mirrors; RETRY / OPEN DOWNLOAD / IMPORT PACKAGE when a download fails.
• COLLECT LOGS turns the crash handler on for one run and carries amdnr_crash.log, the newest dump and danielblnc's dlssnr_on_amd.ini.
• PLAY starts Xbox / Microsoft Store games by their app id.
• One PLAY with a route chooser (the game's own upscaler or AMDNR Anywhere, remembered per game) and the graphics API (DX9 / 10 / 11 / 12 / Vulkan / OpenGL) on every game page.
• Twelve languages: Turkish, Korean and Hungarian join the nine below.
• Rockstar games start through their store (Steam, Epic or the Rockstar Games Launcher), never by their exe.
• AMDNR Anywhere: the game runs below normal GPU priority so the host and Neural Rendering go first; the capture host's tray icon is hidden.
• Resident Evil 2 / 3 / 4 (2023) / Village: a setup notice (they need PureDark's upscaler plugin with REFramework). More games with no DLSS, XeSS or FSR 2+ are marked unsupported, with the reason, before any download.
• FSR 4 (INT8) experimental opt-in on the game page of RDNA 3 handhelds.
• Other mods' loaders are never taken or moved (RED4ext, Cyber Engine Tweaks and the like): the launcher picks another proxy name and its Doctor names any loader an earlier INSTALL moved aside.
• GTA V Enhanced: the game page carries the route line (FSR 3.1 picked in the game is the input; Neural Rendering runs before it) and a note about settings.xml.

New in Launcher 0.3.4.1: you asked, we built it (from the first Discord feedback):

• Nine languages (twelve since 0.3.5.1): English, Arabic, Chinese (Simplified), French, Spanish, Portuguese, Italian, Russian and Polish, plus Turkish, Korean and Hungarian. The launcher follows your Windows language (else English); pick another under LANGUAGE or in SETTINGS, where it is also offered on first start. Doctor findings, install messages and the COLLECT LOGS report stay in English so support can read them.
• Search the library; favourites (a star, and starred games first); hide games (HIDDEN shows them again); rename a game.
• CHOOSE GAME .EXE: pick the game's exe yourself when the launcher picked the wrong one or none. Cyberpunk 2077 and The Witcher 3 (REDengine) are now found in the right folder without it.
• PLAY and OPEN FOLDER on every game's page; STORES switches whole stores on and off; folders you add by hand stay listed, also while their drive is unplugged, until you remove them.
• UNINSTALL asks first and removes everything the mod placed, including what it wrote while the game ran (logs, caches, crash dumps, unfinished reports); it keeps finished Save report zips and a DLL under the proxy name that is no longer an OptiScaler (the game's own).
• COLLECT LOGS on any game, installed or not, with a scan report.
• Handhelds and APUs (ROG Ally Z1 Extreme and other Ryzen APUs) are recognised, with a note that AMDNR on them is still in testing.
• Linux (experimental, not tested by us yet): the same Windows exe can run under Proton, shows a notice with what to do there and also looks for games in your Linux Steam library; steps in "Linux / Proton". Tell us on Discord if it works.

Its source is in Launcher/OpenSource/ of this project's GitHub repository, with its own licence, Launcher/OpenSource/LICENSE.txt. It is not covered by this repository's GPL-3.0 LICENSE: it is source-available, all rights reserved, Copyright (c) 2026 3zwr1 (AMDNR). The launcher's manifest is Launcher/manifest.json. See also section 7 of Licenses/AMDNR_NOTICE.txt.

AMDNR is Copyright (c) 2026 3zwr1 (AMDNR). It is a fork of OptiScaler, distributed under the GPL-3.0 licence in LICENSE.

AMDNR's own work carries an additional term under GPL-3.0 section 7(b) (see Licenses/AMDNR_NOTICE.txt): any copy, fork or derivative that uses it must keep its notices and credit AMDNR by 3zwr1 (https://github.com/3zwr1/AMD-NR---OptiScaler).

AMDNR menu copyright. 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.

The upstream work credited above stays with its authors, under their own licences; AMDNR claims no copyright over it.

The source code will be published with AMDNR 0.5.0.

This build is distributed under the GPL-3.0 licence in LICENSE; third-party library licences are in Licenses\. The AMD neural runtime and its weights are redistributed under their original authorship as credited above, for convenience only, with no ownership claimed and no warranty offered.

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