R13 fixes and verification limits

• The version string is now 0.3.5.8R13.
• danielblnc inline hang fallback (Red Dead Redemption 2 DEVICE_HUNG, Gears of War: E-Day danielblnc reports): when danielblnc-inline-hang.txt exists for the game (in %LOCALAPPDATA%\AMDNR\settings<exe>), AMDNR no longer starts danielblnc in its inline (same-frame wait) mode again; the Neural tab status and amd_presr.log say why ([INLINEHANG-FALLBACK]). The marker only counts when this AMDNR build wrote it, so an update gives one more try. AMDNR never deletes the marker itself; delete the file to try inline again. It is ON by default; [DlssNr] AmdInlineHangFallback=false goes back to advice only, and an explicit Inline=1 or Async=0 in danielblnc's ini is respected. Not yet confirmed in the game.
• Play Anywhere (32-bit games such as Far Cry 3): inside the capture host only, when the NR network needs longer than one host frame and Passes is above 1, the host runs Passes=1 for the rest of the session (runtime only, your ini is not changed) and logs [ANYWHERE-PASSES] once. Each stats window also logs a host [LMX-GHOST] line (repeated/held frames, picture age, history resets). The in-game path and Model B are unchanged. Not yet confirmed in the game.
• Mochi (Phantom Blade Zero): after you change the NR scale, the old Mochi worker could end with a non-zero exit code and the host then kept Mochi off for the rest of the session. Now only a worker that does not end in time counts as a failure; a non-zero exit is logged as [MOCHI-RESIZE] worker ended with exit code ... and a new worker starts. Why that worker exits non-zero is not known. Not yet confirmed in the game.
• Metro (metro.exe): AMDNR no longer raises the D3D11 feature level for metro.exe (SkipD3D11FeatureLevelElevation), and the log shows the first D3D11CreateDevice calls. This is a mitigation attempt for a crash at metro.exe+0x831B95 / +0x831E55: there was no stock-game control run and no crash dump, so it is not proven to help.
• Gears of War: E-Day x3/x4 stall step-down (an experiment, ON by default): in the logs we have, the real Present blocks for about 200 ms on the 3rd present of an x3 batch while Anti-Lag 2 types the presents. The cause is not shown, so this only tries one thing. When 3 such stalls of 120 ms or more happen within 4 s at MFG x3 or above, AMDNR lowers the Anti-Lag 2 frame type tags from full to first-only, and after 3 more stalls to none, and logs [AMDNR-R14-AL2-STEPDOWN]. It starts again at full when the game changes its DLSS frame generation request. x2 and below, and sessions without Anti-Lag 2, are unchanged. [Hotfix] Al2MfgTagAutoStepDown=false keeps the fixed full tags as in R12. [Hotfix] DiagAl2MfgTagMode (1 = first only, 2 = off) pins a level and wins over the step-down. [AMDNR-R13-AL2-TAGS] logs the tags in use. [FIRST-EVAL] begin/end lines are logged around the first FSR dispatch and [FIRST-EVAL-STALL] after 10 s, so a hang can be told from a crash. Whether this helps is not proven, and what dropping the tags does to Anti-Lag 2 is not tested.
• NR not fed (Forza Horizon 6 report): the log now follows [NHUB080-NR-UNFED] with one [NHUB080-NR-UNFED-WHY] line naming the condition that fired. The overlay's NR segment reads "enabled - not fed in this session" and the hotkey toast says "NR is not fed in this session"; the old wording named a cause that was false for 3 of the 5 reasons. Diagnosis and wording only; no fix is claimed.
• Log changes only: the "Frame count jumped too much" warning is logged for the first 8 jumps, then at most once per 5 s with the number skipped. The Wine/Proton text explains a missing amdxcffx64.dll, and the toast for FSR-RR on Super Sampling is clearer. [NR-CARRY] (Model A) also logs passes= and jitterSign=, and the stats line's fill word names the carry that really ran. Behaviour is the same as before.
• Launcher 0.3.5.68 and settings:
  • GPU detection: on a PC with an AMD integrated GPU and an NVIDIA card, the NVIDIA card is now the detected GPU and is shown by name (before, only the integrated GPU showed). Not tested on a real hybrid PC. The wording for non-AMD cards is unchanged: it still says Neural Rendering needs an AMD card, and whether NR runs on NVIDIA without a runtime is not verified.
  • Wine/Proton: the GPU label comes from the PCI id when Wine's DXGI list has no matching adapter (Nobara report).
  • Detection: anti-cheat markers are searched up to 3 parent folders up; a Call of Duty game raises the Ricochet rule, and outdated Call of Duty installs are not offered the update (the Doctor says to uninstall AMDNR from that game). A speed-hack DLL (for example TrSpeedHack) next to the proxy gets a gate note for multiplayer titles and Battlefield V. d3d12.dll beside d3d11.dll with no D3D12 device evidence reads "DirectX 11/12 (mixed imports)". MO2/usvfs and DODI-style repack folders (storage, .v38) get a note, and REFramework is also found in storage\dinput8.dll.
  • TRY NEXT PROXY: dbghelp.dll is offered last and only with a warning. It stops when the last two crashes share a module and offset across different proxy names, names them and suggests the baseline test. Faults in shared system modules (ntdll, KERNELBASE and similar) no longer count as the same fault.
  • COLLECT LOGS: "AMDNR in the game: loaded" follows the session whose log shows it loaded, and says it is proven only for that earlier session when the last run is another process. The sessions line names the process id per file. The proxy-change history comes from amdnr_crash.log START lines only. The rotated OptiScaler log is filtered before the three newest are kept. The report shows the inline-hang marker and ReShade / BFI files. The runtime inventory ignores linked parent folders of the game folder.
  • Per-title advice: Gears of War: E-Day (Steam) recommends NeuralHub 0.8.0 over 0.6.0 as a recommendation only (x2 advised; x3/x4 stutter under investigation; no default changed). Metro Redux asks for a stock-game test before you report.
  • Repack wording (RE Requiem .v38): the pre-install note says "once AMDNR is installed", and the crash note says AMDNR "may have been loaded"; AMDNR is neither cleared nor blamed.
• Mochi: the Mochi worker in this package and AMDNR-Mochi-0.3.5.8R11-component.zip are unchanged from R11 (the component stays on the Alpha0.3.5.8R11 release). The resize fix above is in the host DLL, not the worker. No NVIDIA DLLs or network weights are shipped.
• Not fixed in R13 (please keep sending COLLECT LOGS):
  • lmxxf Model A ghosting / trails: not fixed and no cause shown. R13 only adds log text for it. If you see trails, send a short clip plus COLLECT LOGS and say which model you use and whether NR is in-game or via Play Anywhere.
  • Gears of War: E-Day DLSS FG x3 stutter / x4 freeze compared with R5: not shown to be fixed. The step-down above is an untested experiment; please send COLLECT LOGS from an x3 or x4 session (look for [AMDNR-R14-AL2-STEPDOWN]). The spoof A/B test is still to be done.
  • Mochi can be slower than the frame time on a scene ("[MOCHI-ASYNC] output pending after 16 ms", then the next frame resets history). That is speed, not a setup error, and R13 does not change it.
  • Cyberpunk 2077 lag on RX 9060 XT: needs a log from a lagging session with ReShade and BFI off.
  • Microsoft Flight Simulator 2024: needs the exact error text or a screenshot.
  • Crimson Desert red tint in dark rooms (NeuralHub 0.8.0): the 0.8.0 host is a separate runtime. Its page has Tone strength, NR style (Natural keeps closer to the game's colour), Tone curve and Black lift. These are workarounds, not a fix. Not reproduced here, and no AMDNR cause was found in the code we checked.
  • Resident Evil 4 Remake: frame generation freezing the game / PC, and NR with frame generation when the game has no DLSS FG: cause not shown; send a report from the same session.
  • "DLSS 0.0.0" in the overlay is cosmetic: there is no nvngx DLL on AMD to read a version from.
  • Still open from R12: Stalker 2 NR noise/grain (turn NR off or lower strength); Battlefield V (send one log with spoofing off and one with it on); Farming Simulator 25 (check whether it fails without AMDNR); Forza (send the log; Store/UWP builds may block injection).
  • Advice only: GTA V Enhanced (use its own FSR/DLSS, no other injectors), The Witcher 3 (verify files, use the game's upscaler setting), Mad Max (no native DLSS/FSR: use only frame-level features, or remove AMDNR), Clair Obscur on Linux (use the game's FSR), ETS2 (AMDNR does not upscale it; the reported crash was in another mod's dxgi.dll).
• Left out of R13 on purpose:
  • A Model A fix: a host-side hold that refuses stale pixels was designed and rejected for now (no evidence it helps, and it would change quality for every Model A player). It stays a candidate for a later release.
  • HelixSR 3.1.0 (Model E from 3.1.0 and the HelixSR runtime): built and unit-tested only. It was never run for real or in a game, the runtime has no launcher UI yet, it is not reviewed, and the licence use is not confirmed. Model E stays on HelixSR 2.0.0.
  • Same-frame NR (testers only, off by default, unchanged since R12) for native DX12 games and Mochi: not built yet.
  • Mochi on RDNA3, RDNA3.5, RDNA2 and handhelds, and a Mochi shader mode: not started.
  • Signed Mochi core and forwarder, and the signed Mochi model extractor: not in R13.
  • The antivirus warning on AMDNR's server address: not fixed in R13.

These changes passed the production build with no new compiler warnings (967 exports, the same as R12) and a native startup test of the exact signed DLL (17 cases) with NR, FG and NeuralHub off and same-frame NR not set, so the NR, same-frame, NeuralHub and frame generation log lines were not reached. The Gears of War: E-Day step-down rules passed a unit test that needs no GPU; the code that applies them inside the DLL was compiled but never run. Launcher 0.3.5.68 passed its full test suites (Core 3962, App 1107). The AMD host contract suites were not re-run for R13. None of these changes has been tested in an actual game for this release. Crash, pacing and image-quality fixes need fresh reports with COLLECT LOGS to be confirmed.

Previous release: R12 (0.3.5.8R12, Launcher 0.3.5.67)

• The version string is now 0.3.5.8R12.
• lmxxf ghosting (Model B): Model B no longer carries its previous output into the next frame by default ([DlssNr] AmdLmxxfModelBFreshOnly defaults to true and is not saved). [DlssNr] AmdLmxxfClassicCarry no longer turns the carry back on for Model B, and the menu greys out Temporal stability for Model B. Static scenes may look a little noisier; that is not measured yet. Not yet confirmed in the game.
• lmxxf ghosting (bad motion vectors): while a game's motion vectors are being rejected, NR pauses and its history is reset instead of pasting unwarped old pixels. The status reads "self-heal: NR off while the vectors are rejected" and the log shows [LMX-GHOST]. In a game, NR comes back by itself after 3 stats windows in a row with good vectors. Titles whose vectors are only sometimes bad may see NR switch on and off. Not yet confirmed in the game.
• lmxxf history restart: when a frame is fed two frames after the last one and the frame in between recorded no motion, the network history now restarts instead of chaining stale motion. The log shows [LMX-GHOST] history gap. Late-submit games such as The Last of Us Part II are not affected. Not yet confirmed in the game.
• Still seeing trails? Lower NR strength and send a short clip plus COLLECT LOGS.
• Experimental same-frame NR (testers only, OFF by default):
  • Today NR's answer reaches the screen one frame late and is moved forward with the motion vectors. We think that move is the main source of trails, but not the only one. With [DlssNr] AmdSameFrame=1 (DX11 games), or AmdSameFrame=1 plus AmdSameFrameVk=1 (Vulkan games), the switch is designed so each frame shows either its own NR result or the plain frame, never the previous frame's result. Whether trails are gone on screen is not proven yet.
  • If anything goes wrong, that frame (or the rest of the session, after repeated failures or a stall of 750 ms or more) is shown without NR instead of late NR.
  • While it is on: Image look and CAS sharpening are not applied to those frames; RTGI and debug view 5 are refused; NR's own temporal history is off; danielblnc is never forced into it.
  • Native DX12 games, Mochi, APUs/handhelds and Vulkan games without AmdSameFrameVk=1 have no same-frame route yet and show the plain frame (no NR) while the key is on. Vulkan under Linux/Proton is untested.
  • Logs: [SF-ROUTE], [SF-OK], [SF-RAW], [SF-STALL], [SF-REFUSE].
  • The GPU wait was tested only in a D3D12 test program on an RX 9070 XT, NOT in a game and not on the DX11 or Vulkan paths themselves. The FPS cost is unknown. Testers: please send a short video of a fast camera turn with it on and off, plus your log.
• Starfield (NR streaks / smearing): switching NR on or off (hotkey, Neural tab, NeuralHub menu) now re-creates the upscaler so the provider follows the NR state. Each toggle causes a short hitch while it is re-created. A change seen only in the NeuralHub progress trace waits for two matching reads about 5 s apart; right after game start the old provider can also run about 5 s longer. With NR off, the overlay's FSR label should show the driver's FSR 4. The log shows [NHUB080-NR-FEED-APPLY]. Not yet confirmed in the game.
• Starfield performance: [NHUB080-PROGRESS] now logs the NR GPU time ("pending" until it is measured). Logging only.
• S.T.A.L.K.E.R. 2 with SR/DLSS on NeuralHub 0.8.0 (crash reports): AMDNR tells the host Upscaling.Provider = Passthrough for Stalker 2, so the host stays out of upscaling (partial: it avoids the configuration). Use the game's own FSR, not DLSS. If it still crashes, send OptiScaler.log and the crash file. Not yet confirmed in the game.
• Cyberpunk 2077 + REDmod with Real DLSS FG (crash at start, partial workaround, root cause not found): when a start with DLSS-FG on NeuralHub 0.8.0 dies before the game is running, AMDNR leaves a per-game marker and the next start skips the host. The marker is cleared once the game shows its first frames (or runs 60 s), on a normal quit, or when 0.8.0 is picked again (in-game menu or launcher); turning Real DLSS FG off also lets the next start use the host. Known gap: some crashes (a code-0 exit after a non-fatal error) still clear it. If it crashes again, try once with REDmod off and send both logs. The log shows [NHUB080-FG-START]. Not yet confirmed in the game.
• NBA 2K27 and other games that end themselves with TerminateProcess: when AMDNR's crash note is on (COLLECT LOGS turns it on) and AMDNR NR is running, the crash log now says "END clean exit [CRASH-SELF-TERMINATE]" unless a fatal fault happened in the 30 s before. On RX 6000 cards: install the HIP SDK 7.2 runtime for NR, or turn NR off. Not yet confirmed in the game.
• Shadow of the Tomb Raider on the Radeon 780M (Phoenix) ("FSR 4.1.2 not available") and handhelds: the FSR 4 opt-in box is also shown when the automatic INT8 Lite pick has no FSR 4 provider, a notice explains the fallback, and the log shows [FSR4-PIN-HINT]. The Fast FSR 4 row names the provider that really runs under NR. Workaround (DX12): set Fast FSR 4 shaders = Exact (4.1.1), or tick FSR 4 (INT8) Experimental, then restart the game. Not yet confirmed in the game.
• Launcher 0.3.5.67 and settings:
  • FSR 4 (INT8) box: it now reads and writes the key the installed AMDNR reads. For R12 that is [FSR] Fsr4GpuSupport, and the old Fsr4ForceModel line is removed. A game kept on an older build that only knows Fsr4ForceModel keeps that key. In-game texts now name Fsr4GpuSupport=2.
  • Mochi settings: saving now works when your Windows user folder path is very long (launcher and in-game), and the in-game menu also works in game folders with long paths. Not yet confirmed in the game.
  • Manual game rows: a crash in "+ Add a game" while reading the game's name is fixed, and the folder search below the picked exe runs only for hand-added rows.
  • Cyberpunk 2077 GOG: it can be added by hand (pick the folder that holds bin\x64\Cyberpunk2077.exe).
  • Crash reports (Euro Truck Simulator 2 report): COLLECT LOGS no longer shows an older run's crash (from before the current install, or a reused process id) as the last crash. It names the newest crash since the install and says whether AMDNR was in that run. ETS2 also gets a note: in that report the crash was inside another mod's dxgi.dll; remove that mod or use TRY NEXT PROXY.
  • Clearer texts: startup crash, Wine/Proton, The Last of Us Streamline backup plugins, the Doctor's host frame generation note, the frame generation display in game settings, and the crash recovery report draft.
• Mochi: the Mochi worker in this package and AMDNR-Mochi-0.3.5.8R11-component.zip are unchanged from R11 (the component stays on the Alpha0.3.5.8R11 release). The worker was not rebuilt, so a worker-side long-path change is not in R12; only the launcher and in-game parts above are. No NVIDIA DLLs or network weights are shipped.
• Not fixed in R12 (please keep sending COLLECT LOGS):
  • Stalker 2 NR noise/grain: turn NR off or lower strength.
  • Battlefield V: send one log with spoofing off and one with it on.
  • Farming Simulator 25: check whether it fails without AMDNR and send a failing log.
  • Forza: send the log; Store/UWP builds may block injection.
  • Advice only: GTA V Enhanced (use its own FSR/DLSS, no other injectors), The Witcher 3 (verify files, use the game's upscaler setting), Mad Max (no native DLSS/FSR: use only frame-level features, or remove AMDNR), Clair Obscur on Linux (use the game's FSR), ETS2 (AMDNR does not upscale it, so you can remove AMDNR there; the reported crash was in another mod's dxgi.dll).
• Left out of R12 on purpose:
  • HelixSR 3.1.0 (Model E from 3.1.0 and the HelixSR runtime): built and unit-tested only. It was never run for real or in a game, the runtime has no launcher UI yet, it is not reviewed, and the licence use is not confirmed. Model E stays on HelixSR 2.0.0.
  • Same-frame NR for native DX12 games and Mochi: not built yet.
  • Mochi on RDNA3, RDNA3.5, RDNA2 and handhelds, and a Mochi shader mode: not started.
  • Gears of War: E-Day DLSS FG x3 stutter / x4 freeze (Anti-Lag 2): still diagnostics only; the change waits for an A/B test.
  • The Last of Us: renaming backup-named Streamline plugins automatically. R12 gives advice text instead.
  • Signed Mochi core and forwarder: not in R12. Signed Mochi model extractor: needs a launcher change first.
  • The antivirus warning on AMDNR's server address: not fixed in R12. Possible fixes (a new server address, false-positive reports to antivirus vendors, a clearer launcher message) are not done yet.

These changes passed the production build with no new compiler warnings (967 exports, the same as R11) and a native startup test of the exact signed DLL (17 cases) with NR, FG and NeuralHub off and same-frame NR not set, so the NR, same-frame, NeuralHub and frame generation log lines were not reached. Launcher 0.3.5.67 passed its full test suites (Core 3882, App 1107). The AMD host contract suites were not re-run for R12. None of these changes has been tested in an actual game for this release. Crash, pacing and image-quality fixes need fresh reports with COLLECT LOGS to be confirmed.

AMDNR 0.3.5.8R13 Public View

Neural Rendering for AMD Radeon, built on OptiScaler. Paired launcher: 0.3.5.68.

Video

https://youtu.be/PGMeJsf4utY — existing 0.3.5.8 overview, not an R13 validation video.

Retained R5 changes

• Fixed startup crashes caused by applying menu colours before the ImGui context exists.
• Corrected RR discovery and late-load GPU capability queries that could re-enter the loader or wait on each other.
• Reject NGX providers that forward their exports back into AMDNR, while allowing recovery to a valid provider.
• Resolve existing module handles and loader caller/diagnostic policy before waiting for GPU discovery.
• Added custom DLSSG provider support: custom/<name>/dlssg/<dll>, with menu selection, saved settings and the active provider in the overlay.
• Extended file discovery for Phantom Blade Zero's steam directory layout.
• Launcher: validate lmxxf packages before replacing a working installation; preserve NR on/off, quality and independent SR/FG choices during profile restoration and runtime recovery.
• Mochi settings now use per-installation storage shared consistently by the launcher, menu and authored worker. Updates preserve other keys and reject unreadable settings instead of silently restoring stale choices.
• The authored worker uses a writable working directory and defers cleanup until its process exits and the job is idle. The runtime author's core files remain unchanged.
• Clear AMDNR data now removes the AMDNR-Uninstall helper workspace after successful cleanup. Game files outside the fixed LocalAppData product folders remain untouched.

Retained features

AMDNR Neural controls, menu languages, SR/RR/FG separation, ReShade controls, FSR 4.1.2 and the separate Exact/Lossy experimental choices remain. Includes the earlier R4 swapchain/resize, FG completion and shader-preparation corrections, plus the October 29 OptiScaler FFX SDK 2.4/custom-provider, XeSS mask, BeamNG and XeFG detection updates. They are retained changes, not new performance claims.

Install / update

Close the game, update AMDNR Launcher to 0.3.5.68, then update AMDNR for the game to 0.3.5.8R13. The paired authored worker and launcher fixes are required; replacing only OptiScaler.dll does not apply all changes. Keep your selected runtime installed separately. Save settings and restart where requested.

For custom FG providers, place a compatible replacement DLL under custom/<name>/dlssg beside OptiScaler, select it in FG Nvngx / FG Nvngx Replacement, save and restart. If multiple DLLs are present, nvngx_dlssg.dll takes precedence. Provider availability does not guarantee compatibility with every game or GPU.

Use COLLECT LOGS after reproducing a problem. Include game, GPU, runtime/model, upscaler, FG and NR placement/interleave settings. For visual artifacts include a short moving-scene video. Send dumps privately.

Files

• AMDNR-v0.3.5.8R13.zip: mod, paired authored worker, dependencies, licence notices and documentation.
• AMDNR-Launcher.exe: Launcher 0.3.5.68.
• Mochi (mochizuki 0.0.8.1 core): unchanged from R11 and not attached to this release. AMDNR-Mochi-0.3.5.8R11-component.zip stays on the Alpha0.3.5.8R11 release and the launcher installs it from there as before (AMDNR's forwarder nvngx.dll_dlssnr.dll and the matching worker; no NVIDIA DLLs or weights).
• SHA256SUMS.txt: hashes of the published files.
  Runtime and Anywhere packages remain separate. No NVIDIA DLLs or NR weights are included in this main mod package.

Known limits

These changes target reported failures; several are precautions or diagnostics only, none is confirmed in a game yet, and they do not certify every game or close every report.

• Phantom Blade Zero DLSS MFG pacing/timeouts remain under investigation. Neither the R5 directory-discovery fix nor the R9 DLSS-G MFG timing changes establish a pacing fix.
• S.T.A.L.K.E.R. 2 artifacts/ghosting require a controlled scene comparison and remain unresolved.
• F1 25's recorded R4 menu-theme crash is covered by the code correction; in-game acceptance and separate spoofing/driver crashes remain unconfirmed.
• Cyberpunk native FSR FG/AL2 and Phantom Blade Zero FSR FG need fresh R13 player confirmation. Flight Simulator and other game-specific failures remain under investigation.
• Long-session Darktide XeFG failures, foreground/background pacing, high-interleave latency, remaining ghosting and other game-specific reports are not claimed fixed.
• Launcher 0.3.5.63 and later stop the Anywhere exclusive-full-screen minimise loop; Metro Anywhere still needs player confirmation. The separate capture-host crash is outside this update. Existing Runtime/Anywhere packages stay separate.
• Loader paths involving unknown NVAPI callers and some FSR4/custom RDNA 2 capability paths, and worker deferred/concurrent/resize scenarios, are not fully qualified.
• No general FPS improvement, image-quality guarantee or new universal GPU compatibility claim is made. Stock FSR, Exact/Lossy and experimental options remain distinct.

Credits

DLSS-NR on AMD by Daniel Blanco (danielblnc) · lmxxf by Kien (MIT) · mochizuki by mochizuki0323 (MIT) · built on OptiScaler (GPL-3.0). HelixSR by lonewolf0622; TheAutomatic integration lineage; AMD FidelityFX, Intel XeSS and other dependency authors. Experimental shader overrides by zgauthier2000 (cyboman); see included licences. Copyright (c) 2026 3zwr1 (AMDNR). Not affiliated with AMD or NVIDIA.

Support

https://ko-fi.com/3zinr