**Video:**→ https://youtu.be/rQ1OyDBej1g
AMDNR 0.3.5.1 is a hotfix: the answers to what you reported in the first day of 0.3.5, and lmxxf 0.39 by Kien
(MIT). The network's weights, the danielblnc runtime and the Anywhere host are unchanged; OptiScaler.dll,
OptiScaler.ini, LmxxfNrRuntime.dll and LmxxfNrRuntime.pak change. Launcher users: REPAIR / UPDATE, or let it
update at start. By hand: drop the zip over 0.3.5.
Requires AMD Software: Adrenalin Edition 26.9.1 or newer.
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.zipfiles
(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.
New
- Faster neural rendering on RX 9000 with lmxxf: lmxxf 0.39 by Kien (MIT). lmxxf's newer kernels load beside
the 0.37 set, each checked by SHA-256, with the same picture as 0.37. Network time per frame on an RX 9070 XT:
11.33 -> 11.03 ms at the 1080 tier, 8.29 -> 8.05 ms at 900, 5.61 -> 5.40 ms at 720, and its new one-wave path
takes about another 0.1 ms at 1080 and 900. The RX 9060 XT runs the same set, not measured there yet. Anything
missing or mismatched falls back to 0.37 by itself;[DlssNr] AmdLmxxfL39=falsekeeps 0.37. RX 7000 and
handhelds are unchanged. - Network style on lmxxf: the Network style row (Neural tab, Model strength) now works with lmxxf too: Style 0
is NVIDIA's default look, Style 1 the one lmxxf has always used (Auto keeps it), Style 2 the third. A change
rebuilds the network, about a second. It is the same[DlssNr] AmdRuntimeStylekey as danielblnc's style, so a
value set for danielblnc now applies on lmxxf too.
Fixed
- Wuthering Waves and other Unreal Engine 4 titles no longer hit a fatal error on the first neural frame: the
large-stack helper now moves the neural pass only when the game evaluates from a job fiber; a plain engine thread runs
it on its own stack as before.[DlssNr] BigStackCalltakesauto(default),true(the 0.3.5 behaviour) orfalse. - The first neural frame no longer stalls for seconds while AMDNR checks which other mods are loaded: the module walk
resolves each folder once and skips Windows' own modules. - The adapter list written to
amd_bridge.logis read once on a thread of its own instead of inside the first neural
frame;[DlssNr] AmdAdapterList=falseskips it entirely. - GTA V Enhanced and every FSR 3.1 title: the game's video-memory readout no longer shows a nonsense figure like
"929756 MB" when AMDNR drives the upscaler; the memory query now gets a real estimate. - "Upscaler failed to run!" no longer appears for a single skipped frame (common with FSR Ray Regeneration at the start
of a session); it shows only when the upscaler has failed for about half a second in a row. - RX 9060 XT / 9060 / 9050 with the danielblnc 0.5.0 runtime: Neural Rendering runs asynchronously by default on these
cards. The same-frame wait that could freeze the whole PC is off;Async=0indlssnr_on_amd.initurns it back on, and
anyAsync/Inlinekey you set still wins on every card. - Games with dynamic resolution: the seamless-resize hold is tighter (only steps within 90 percent of the largest frame
are held, and the log shows what it costs);[DlssNr] AmdHoldResize=falseturns the hold off entirely if your FPS
dropped with it (the 68 to 48 fps reports with frame generation). - Black Myth: Wukong with frame generation: the game's own frame-generation swapchain no longer shares the XeFG
swapchain (the same fix Control Resonant got in 0.3.5); the Frame Generation tab says when a game created a second
swapchain under XeFG. - XeSS multi-frame generation above 2X is labelled experimental (Intel ships 2X only on non-Intel cards; the higher
multipliers run on a patched provider); RX 9000 owners get an advice line: prefer FSR FG (FSR 4 FG) as the output. - "Show Detected UI" (Frame Generation tab) is a debug view that tints the picture pink on purpose: it turns itself
off when the menu closes, is never saved, and says so on screen while it is on. - RX 7000 (RDNA 3): Ray Regeneration is offered again by default, as a preview. The Ray Regeneration tab appears
and the game can turn Ray Reconstruction on; AMD's denoiser has no RDNA 3 provider, so the picture is upscaled
without a denoiser and can look noisier than the game's own. The Upscaling tab's box "Preview: Ray Regeneration on
this card (restart)" is ticked by default; untick it or set[FSR-RR] FfxDenoiserAllowPreRdna4=false. - Handheld preset: the Neural tab's Handheld button applies the tuned handheld settings on both runtimes.
- NR resolution on lmxxf: the slider stops where the network stops growing; a saved value above that is shown as
capped and the cost tag says which percent it prices ("0.64x cost (as at 100%)"). - Anywhere: a steady picture. New FPS limit slider in the Anywhere tab (60 by default; any rate on a FreeSync
display): the Anywhere window holds one rate for the whole session, with or without frame generation, and the launcher
holds the game at the same rate through AMD Software's Radeon Chill while Play Anywhere runs, then puts your own
settings back (Anti-Lag and Boost included). A game running far above the rate no longer judders. - Anywhere tab in plain words: Smoothness / Picture / Status / Advanced, "the Anywhere window" and "AMD Software"
instead of host and driver, and frame generation explained (30 real frames show as 60). - Neural placement (Before / After upscaling) switches at once from the menu; before, it needed a game restart.
- danielblnc 0.5.1 is driven (
v0.5.1-Runtime.zip): picking it no longer falls back to lmxxf; on RX 9060 XT / 9050 it
runs async by default like 0.5.0. - COLLECT LOGS now summarises danielblnc's own log: how many stalls and lost captures it holds and whether it ran
same-frame or async, so a freeze report reads without opening the log by hand.
Known issues
- Ray Regeneration on Proton: unchanged (known since 0.3.4.1).
- Colours against danielblnc's own overlay (Style, tone): being measured; not in this hotfix.
- Black Myth: Wukong with XeSS frame generation: no report with a log yet. Send COLLECT LOGS with
[Log] CrashHandler=true. - lmxxf on RX 9000 in Marvel's Spider-Man Remastered (report): about 20 percent fewer fps than the danielblnc
runtime and a trail behind moving objects. Not a 0.3.5 slowdown: a 1712x960 feed is priced as the 1080 tier (just
above the 900 box), and lmxxf's one-frame-late carried edit trails where his same-frame runtime does not. A/B lines
under[DlssNr], one at a time:AmdLmxxfFastMode=true(fps),AmdLmxxfTierCap=720(fps),AmdTemporalStability=0
andAmdLmxxfHistory=false(trail). The danielblnc runtime is the right pick there today.
Update
- Launcher users: REPAIR / UPDATE, or UPDATE AT START. The launcher itself updates to 0.3.5.2 (INSTALL asks "in the
game or Play Anywhere", the Play Anywhere download shows, RX 9060 XT gets the 720 tier, reports carry the host's crash
files, the 32-bit line is a note and a starter beside the 64-bit game no longer wins). - No launcher yet?
AMDNR-Launcher.exe0.3.5.2 is on this release (signed,72,087,704bytes, SHA256
f2311524ccee478826068594313c76d987a34d41e1c0a5b1d314b031c1632dc3). - By hand:
AMDNR-v0.3.5.1.zip(629,270,812bytes, SHA256
6ede7b984fc02f2285f6355b98d3009391cb082992250292d6e2bbc7d442fc8a),SHA256SUMS.txt,AMDNR-Launcher.exe.
Thanks
To everyone who sent COLLECT LOGS zips on 1 October: every fix above came from one of them. And to Kien for lmxxf
0.39.