AMDNR 0.3.5.6

Video : https://youtu.be/v2LynDVtNtI

** First start with Model L: the screen can stay black for up to a minute while d4r prepares its kernels. This happens once; later starts are instant. !!!!!!!!!!**

Neural Rendering for AMD Radeon, built on OptiScaler. Install and update through the AMDNR Launcher 0.3.5.25.

Download: AMDNR-v0.3.5.6.zip - SHA-256 <sha256> ( bytes)

What's new

• Two new upscaler choices (experimental, opt-in): Model E (HelixSR) and Model L (d4r). See the section below before you enable them.
• Full Model Look by default: the default look now runs after upscaling on cards that can afford it. Picking a look in the menu also sets the placement and network size it needs, and the menu tells you if your own setting holds it back.
• Extra detail on faces: always on, on both runtimes.
• Less ghosting with lmxxf: trails behind moving objects are shorter, before and after upscaling, and the flash after a scene cut is about half as strong. Ghosting is reduced, not gone.
• RX 7000 and handhelds: Neural Rendering now runs one network size smaller by default, which costs less time per frame. Set AmdLmxxfFastMode = false in [DlssNr] to turn it off.
• New menu controls: Game exposure (Auto / On / Off) and Anti-halo (formerly Halo protection).
• NR resolution now goes up to 200% on both runtimes.
• The menu only shows what your runtime can use: controls a runtime cannot run are hidden on that runtime.
• The lmxxf version is shown as the runtime's real version.

Fixes

• danielblnc runtime: false "timeouts" are fixed, and the old on-screen text claiming the driver reset the GPU is gone (there was no reset). A wait that gives up no longer turns Neural Rendering off, and the "runtime failed" note is written only after a real failure.
• FSR 3 / FSR 4 should no longer drop back to FSR 2.1.2 when another upscaler choice fails or is not ready. Your own FSR pick is kept.
• Better memory accounting for XeFG frame generation (the 6X setting is now counted correctly).

Model E and Model L (experimental, opt-in)

Both are fetched by the launcher from their authors' own releases when you choose them. AMDNR does not host or ship their files.

• Model E is HelixSR by lonewolf0622 (github.com/lonewolf0622/HelixSR), used under the HelixSR Freeware License. It builds its network on your PC from NVIDIA's DLSS file, which the launcher downloads from NVIDIA's own GitHub (or takes from your game). AMDNR does not host or ship these files. Works with both NR runtimes.
• Model L is d4r by countervolts (github.com/countervolts/d4r). It runs NVIDIA's DLSS runtime, under NVIDIA's EULA, through ZLUDA/HIP. d4r has not been tested by its author on Windows. AMDNR does not host or ship any NVIDIA files.
  • RX 9070 / RX 9070 XT only. Tested on an RX 9070 XT.
  • Model L runs with the lmxxf runtime, or with Neural Rendering off. danielblnc's runtime pairs with Model E.
  • The launcher shows the full disclosure before you enable it.

Known limits

• Ghosting with lmxxf is reduced, not removed.
• The faster default on RX 7000 and handhelds has no game reports yet. Please send yours.
• In some D3D11 and Vulkan games a failed upscaler can still fall back to FSR 2.2.

Credits

• lmxxf by Kien (MIT).
• The Automatic, lmxxf runtime
• DLSS-NR on AMD by Daniel Blanco (danielblnc), used with his permission and installed from his own installer.
• HelixSR by lonewolf0622. d4r by countervolts.
• OptiScaler and its contributors; see Licenses/AMDNR_NOTICE.txt.
• Copyright (c) 2026 3zwr1 (AMDNR). Not affiliated with or endorsed by AMD or NVIDIA.