AMDNR 0.3.5.2, The Witcher 3, Intel frame generation, driver resets, RE Requiem

Video: Hotfix 0.3.5.2 video

AMDNR 0.3.5.2 is a hotfix: the answers to what you reported on 0.3.5.1, led by two crashes and a black
screen. The network's weights, the danielblnc runtime, the lmxxf kernels and the Anywhere host are unchanged
(LmxxfNrRuntime.pak is byte-identical to 0.3.5.1); OptiScaler.dll, OptiScaler.ini and LmxxfNrRuntime.dll
change. Launcher users: REPAIR / UPDATE, or let it update at start. By hand: drop the zip over 0.3.5.1.
Requires AMD Software: Adrenalin Edition 26.9.1 or newer.

One default changes: Intel's XeSS frame generation now runs exactly as Intel ships it, at 2X, unless you ask
for the multi-frame unlock. If you were running 3X or more, see the first entry under Fixed.

Join the Discord (support, bug reports, test builds): https://discord.gg/AMDNR

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), shipped unmodified with his
permission - https://github.com/danielblnc/DLSS-NR-on-AMD. Copyright (c) 2026 Daniel Blanco, all rights
reserved. It is not AMDNR's work. Please support his project.
The lmxxf runtime is built on lmxxf by Kien (MIT), including the lmxxf 0.37 and 0.39 kernels by Kien (MIT) -
https://github.com/lmxxf/dlss5-on-amd-9070xt-porting.
AMDNR Anywhere runs inside Magpie by Blinue, experimental fork by SAOG0721 (GPL-3.0), fetched from the
author, not redistributed by AMDNR - https://github.com/SAOG0721/Magpie.
Ray Regeneration comes from FSR Ray Regeneration for OptiScaler by Zach Hembree (DarkHelmet) (GPL-3.0),
continued by burak113.

New

• Anywhere: FSR 4 can scale the window, and on V3 it now does so with Neural Rendering on too. The Upscaling row
  has a third entry, V3: V1's single stage with FSR 4 in place of FSR 3, out of the same library AMDNR already
  puts beside the window, so it downloads nothing. Until now, with NR on, AMDNR's own pass scaled the captured
  picture and the FSR stage was skipped - on purpose, because a captured window has no jitter, so an FSR temporal
  pass finds no extra detail in it and only blends the frame before through estimated motion. That reasoning was
  measured on FSR 3, which has no neural network, so on V3 FSR 4 now carries the picture either way, with that
  dispatch's history reset. A new row, Who scales it (Anywhere menu > Upscaling), switches between "FSR 4" and
  "AMDNR's own pass" live. V1 and V2 are untouched, and so is any card or window that cannot take FSR 4 - only a
  window you picked V3 for changes. Two things to know while you compare: it only means anything on a window smaller
  than your screen, and the FSR stage sharpens (RCAS) where AMDNR's pass does not, so part of any difference you see
  is sharpening rather than the upscaler. [DlssNr] AnywhereFsr4Upscale=false is the way back.
• RX 9000: lmxxf snaps the NR size to the nearest network tier by default, as RDNA 3 already did, so 1440p FSR
  Quality (1707x960) runs the 900 tier instead of 1080: Cyberpunk 2077 went from 52.2 to 64.0 fps (+22.7%) on an
  RX 9070 XT with no visible difference in the player's test. [DlssNr] AmdLmxxfTierSnap=false restores the old sizes.
• New, opt-in: [DlssNr] AmdLmxxfFastHistory=true (lmxxf): the network's history pass runs lmxxf's float version
  instead of the exact double-precision one, which costs about 1.8 ms per frame on RX 9000. The picture differs
  slightly, so it is off by default.
• Radeon 760M (8 compute units): lmxxf runs, as an experimental and very slow option. About 75-90 ms per network
  run at the smallest network size (an estimate - no 760M has been measured yet), with the model every 4th frame in a
  game. Inside the Anywhere window it runs on every frame, so expect a low frame rate there. The Z1 and the Radeon
  740M (4 compute units) stay unsupported.

Fixed

• Intel's XeSS frame generation runs as Intel shipped it unless you ask for more. CONTROL Resonant crashed with
  it and God of War Ragnarok was unstable. Everything above 2X on our cards exists because AMDNR rewrites five
  places inside Intel's frame-generation library in memory and redirects three of its entries for pacing - a
  vendor's code on a path its maker never tests on a card that is not theirs. That whole path is now off by
  default: [XeFG] UnlockMFG=true in OptiScaler.ini turns it back on, and the ceiling and the pacing are
  unchanged when you do. One line in the log says which way a session ran. In those two titles the pacing stays off
  by itself even with the unlock on; [XeFG] ExtraPacing=true hands it back. A provider build AMDNR has no table
  for is left alone instead of patched blind, and a leftover XeFGUnlock*.asi plugin in the game folder is named in
  the log - with the built-in unlock off, that plugin loads and patches the same library by itself.
• The Witcher 3 no longer crashes to the desktop on the second launch. The game played once, and the next start
  closed about four seconds in until you uninstalled and installed AMDNR again. The game ships three frame
  generations of its own, and AMDNR was holding on to a frame-generation swapchain the game had already given up -
  so when the game asked Windows for a new one for the same window, Windows refused and the game went down. AMDNR
  now lets that one go and takes a fresh one, as it already did for CONTROL Resonant and Black Myth: Wukong.
  Nothing else about the title changed, and the crash was before Neural Rendering ran at all.
• A driver reset no longer turns into a session of them. When the display driver decides the GPU has stopped
  answering while the network is running, it resets the device - a stutter and usually a black flash. AMDNR used to
  try again one second later, every time, for the whole session: one handheld report ended with 172 resets and the
  game closing. Now the wait between tries grows (1, 2, 4, 8, 16, then 30 seconds), and after 8 resets one after
  another - or 32 in a session - Neural Rendering turns itself off for the rest of it and says so in one line. The
  game and its upscaler carry on without it, and starting the game again tries once more. A minute without a reset
  starts the count over, so a game that stumbles twice at a loading screen and then plays normally is back to one
  second. [DlssNr] AmdTimeoutKeepTrying=true never stops - the growing wait still applies, so it costs about two
  resets a minute instead of sixty.
• No more black screen in Resident Evil Requiem (and any game that hands over an exposure value too small for
  the hardware to treat as a number). The value passed the old check for "is this a real number above zero", and
  then dividing by it sent the picture to black. Every place an exposure value enters the upscaler now rejects one
  that small and uses 1.0 instead. Verified by the player who reported it.
• Wuthering Waves no longer shows "Fatal error!" when Neural Rendering starts in the lobby or with NR on at
  launch, with either runtime. 0.3.5.1's note blamed the large-stack helper; the real cause was AMDNR itself: its
  first-frame module check released Windows' own dxgi.dll (the copy AMDNR forwards every DirectX call to) and
  Windows unloaded it. That dll is now pinned for the life of the game, every handout of it is counted, and the
  module check reads version information without loading anything. Other Unreal Engine titles on the same hook path
  are covered too.
• Black Myth: Wukong: no more black screen with Neural Rendering on. The exposure value the game hands over
  turned the network's answer black; AMDNR now ignores it in this game by default - the fix a player found
  ([DlssNr] AmdUseGameExposure=0). A value you set yourself still wins.
• Save Settings works in games installed under C:\Program Files (GTA V Enhanced from the Rockstar Games Launcher
  or the Epic Games Store, among others). The game cannot write there, and the failed save was silently ignored, so
  every change was gone at the next start. Now such a save goes to
  %LOCALAPPDATA%\AMDNR\settings\<game>\OptiScaler.ini and the next start reads it; the log says where it went.
• RX 9060 XT / 9060 / 9050 with the danielblnc runtime: same-frame again by default. 0.3.5.1's asynchronous
  default on these cards caused problems, so it is off. Async=1 in dlssnr_on_amd.ini still asks for asynchronous
  mode (the way out if your whole PC froze with same-frame), and any Async / Inline key you set wins.
• Tone intensity starts at 0 with the danielblnc runtime, in the Anywhere window and in a game. With it at 1 a
  captured picture mostly got darker - "super dark" in Dark Souls Remastered - and could show a glow around
  characters; in a game it is the one knob AMDNR was driving that his own overlay leaves alone, which is why some
  players saw flicker and ghosting "but only on yours". 0 is his own default, so his runtime is now driven the way
  he drives it. The lmxxf runtime is unchanged, and a value you set in the menu or the ini wins everywhere.
• Ray Regeneration: less grain, flicker and colour bleeding. The light that goes around the denoiser (the spatial
  floor, and the raw frame AMDNR blends back for detail) is now held to Ray Regeneration's own picture instead of
  being added on top unfiltered; textured surfaces such as skin, brick and fabric no longer copy the raw frame's grain
  in full; and on strongly coloured surfaces such as green foliage, glow, fog and particles are no longer amplified
  into magenta patches. New defaults: Stability bias 0.75 (was 0.5), Correlation bias 0.5 (was 1.0), handover leash 1.0
  (was 2.0). Ray Regeneration tab > More Ray Regeneration options has Composition repair, Chroma-safe demodulation and
  Correlation bias to compare live; OptiScaler.ini [FSR-RR] lists the settings that give 0.3.5.1's picture back.
• CONTROL Resonant (Ray Regeneration): glass and polished panels are denoised with their surroundings
  (material type 0) and held more tightly to Ray Regeneration's picture (leash 0.75), aimed at the flickering
  rectangles on buildings a player reported.
• Auto-exposure no longer runs on a single GPU thread. lmxxf's auto-exposure, danielblnc's white-point estimate
  and the Anywhere window's dark-fade measure spread their 64x36 sample grid over 256 threads, with the same result:
  about 1 ms of GPU time per frame less at 1440p on an RX 9070 XT (a player's own profile of AMDNR).
• Anywhere on 16:10 and ultrawide screens: the Game window line in the Anywhere menu names sizes in your screen's
  own shape (1440x900 on a 1920x1200 screen) and, when the picked size is as tall as your screen, your screen's own
  size: no upscaling then, and Neural Rendering and frame generation still run.
• A game that cannot start HIP says which of the reasons it was in the menu, instead of pointing at a section of
  the support notes.
• A report from a Skyrim modlist now says ENB is there, and under which file name. ENB and AMDNR both want the
  same file names, and a report that did not mention it cost a round trip on every Lorerim-style setup.

Known issues

• Ray Regeneration on Proton: unchanged (known since 0.3.4.1).
• Ray Regeneration on RX 7000 (RDNA 3) runs, but the denoiser quality is the open problem - AMD's denoiser has no
  RDNA 3 provider, so the picture is upscaled without one. Being worked on; not in this hotfix.
• Colours against danielblnc's own overlay (Style, tone): the Tone default above closes one half of it; the rest is
  still being measured.
• FSR 4 against AMDNR's own pass in Anywhere has not been compared yet - by anyone, including us. The new default
  is there because picking V3 is already asking for FSR 4, not because one has been shown to look better. If you
  compare them, the log line names whichever carried the picture; please send it.
• lmxxf on RX 9000 in Marvel's Spider-Man Remastered (report): the tier-pricing half of this is addressed by the
  tier snap above (a 1712x960 feed no longer pays the 1080 tier). The trail behind moving objects is lmxxf's
  one-frame-late carried edit and is unchanged: [DlssNr] AmdTemporalStability=0 and AmdLmxxfHistory=false are the
  A/B lines, one at a time, and the danielblnc runtime is still the right pick there today.
• Black Myth: Wukong with XeSS frame generation: no report with a log yet. Send COLLECT LOGS with
  [Log] CrashHandler=true.

Update

• Launcher users: REPAIR / UPDATE, or UPDATE AT START. The launcher itself updates to 0.3.5.6: it no longer
  starts an installer instead of your game when you added the game's folder by hand and had run one inside it; the
  Doctor spots a standalone copy of the danielblnc runtime installed in the game folder on its own - which loads the
  game a second time - and names the file to delete; a report carries OptiScaler.ini as well, the file that answers
  "it worked once and then it crashed"; the Anywhere window is closed before the launcher writes over its files, and
  a file it still holds is named; and the Anywhere window now closes with a game you closed, instead of waiting - it
  waits the three minutes only when you walked away from a game that is still running.
• No launcher yet? AMDNR-Launcher.exe 0.3.5.6 is on this release (signed, 72,134,400 bytes, SHA256
  2cc7a267b794fa5b3d2e818c2065fa2065ce12978ff854558ba9b828653c5f1a). The certificate is AMDNR's own and is
  self-signed, so Windows SmartScreen still warns the first time; the hash above is there if you would rather check
  than trust it.
• By hand: AMDNR-v0.3.5.2.zip (629,396,152 bytes, SHA256
  e7b147f9c0be5b3c79fd02bbc26e347c5cb145b1ccb4361428e79ff86dd1f15e), SHA256SUMS.txt, AMDNR-Launcher.exe.
  Your OptiScaler.ini is kept as it is - new keys fall back to their defaults until you set them.

Thanks

To the players whose logs named the causes outright: the Witcher 3 report, the handheld session that ended with 172
driver resets, the Resident Evil Requiem black screen (and the A/B that confirmed the fix), and the two Intel
frame-generation reports. And to everyone who sent a COLLECT LOGS zip - every fix above came from one of them.