Skip to content

Releases

DLSSNR-RDNA3 0.0.9

Choose a tag to compare

@github-actions github-actions released this 13 Nov 15:22

Linux/Proton release of the DLSSNR-AMD 0.0.9 port for AMD RDNA3, from the main rdna3 branch.

Changes

  • Ported upstream 0.0.9 Linux/Proton processing and runtime changes to RDNA3.
  • Added Interleaved fast passes with two presets for RenoDX and wilsjo2 processing.
  • Updated HUD FIX, HUD capture, frame history handling, and diagnostics.
  • Integrated RenoDX, Preprocess, and reduced-resolution processing optimizations.
  • Adapted Ultra INT4: QKV quantization is restricted to calibrated encoder layers; decoder QKV remains FP16.
  • Preserved Ditto SR, temporal reuse, NR Control cursor fixes, and XeSS MFG telemetry.
  • Added APU test support, RDNA3/RDNA4 CI builds, and a fix for stale Vulkan Loader caches after repository renaming.

Downloads and installation

  • RDNA3, 64-bit games: choose the archive ending in x86_64-rdna3.tar.gz (includes the OptiScaler route).
  • RDNA3, 32-bit games: choose i686-rdna3.tar.gz (ReShade routes).
  • Archives without the -rdna3 suffix are for RDNA4.
  • SHA-256 checksums are attached as matching .sha256 files and SHA256SUMS.txt.
  • Extract the package and run bash install.sh "/path/to/game" --dll "/path/to/your/nvngx_dlssnr.dll".
  • NVIDIA models are not included. Use your own 310.9.16 DLL for both base models and all derived variants.
  • Review existing NRFastMethod / FastMethod settings: 0 Interleaved, 1 Extrapolation, 2 Feedback, 3 Hybrid, 4 Edit strength.

Validation

On Z1 Extreme, Silver, Small, and legacy INT4 frames remain bit-identical to the previous fork at 1080p, 1440p, and 4K. All seven network choices ran deterministically at 1080p. Historical strict frame-golden differences and independent numerical references for fitted/optimized modes remain documented in the validation report.

Source commit: c518f6508dac6ffacda96331ec6df56e6ef7a281.

DLSSNR-AMD-Vulkan-Linux-0.0.8.3-v.3-x86_64-rdna3

Choose a tag to compare

@saaaaaaaaaaaaaaaaaas saaaaaaaaaaaaaaaaaas released this 10 Nov 17:55

What's New

New Neural Upsampler: Ditto SR (OptiScaler route)

NR Control → NR settings → Resolution → Upsampling → Ditto SR

Added Ditto SR as a new neural upsampling option, based on Vancete/DittoSR.
It's a lightweight upscaler that has very impressive results.

The integration uses the following pipeline:

neural → Ditto SR → up_neural → standard resolve → up_proxy

Ditto SR receives the game's actual motion vectors, depth information, and per-frame jitter, allowing it to make use of temporal data rather than relying on image-only reconstruction.

Each pass maintains its own history buffer, keeping its temporal data independent and avoiding interference between processing stages.

Image quality and stability

Based on my testing, Ditto SR demonstrates significantly better color consistency and temporal image stability compared to the other three upsamplers currently available in this project.

Its most impressive advantage becomes apparent at very low render resolutions. Even when the neural input resolution is heavily reduced, Ditto SR delivers remarkably stable image reconstruction and preserves visual consistency. This makes it particularly interesting for aggressive upscaling scenarios where conventional upsamplers tend to struggle.

For handheld devices and hardware with limited GPU resources, this could be a particularly useful option.

Tested on Rog Ally Cachy OS

Credits and licensing

Ditto SR is developed by Vancete and distributed under the MIT License.

This integration also retains the required third-party notices for the underlying Qualcomm Snapdragon Game Super Resolution 2 components.

Please visit the original project for more information and development details.

About

Upstream from https://gitlab.com/mochizuki0323/DLSSNR-AMD using GPT 6.1 Sol based on mauri870: https://github.com/mauri870/DLSSNR-RDNA3.

2 new models work including CC_SMALL_FP8_EXPERIMENTAL and CC_SILVER_AARDWOLD_OPTIMIZED.

16-bit history works witch halves VRAM usage.

Usage

End key opens NR Control

Install

bash install.sh "<folder with the game's exe>" --dll <nvngx_dlssnr.dll or zip>

DLSSNR-AMD-Vulkan-Linux-0.0.8.3-cursor-fix-x86_64-rdna3

Choose a tag to compare

Bugfix

Cursor now shows in NR Control. Tested on Rog Ally gamescope session.

About

Upstream from https://gitlab.com/mochizuki0323/DLSSNR-AMD using GPT 6.1 Sol based on mauri870: https://github.com/mauri870/DLSSNR-RDNA3.

2 new models work including CC_SMALL_FP8_EXPERIMENTAL and CC_SILVER_AARDWOLD_OPTIMIZED.
16-bit history works witch halves VRAM usage.

Usage

End key opens NR Control

Install

bash install.sh "<folder with the game's exe>" --dll <nvngx_dlssnr.dll or zip>

DLSSNR-AMD-Vulkan-Linux-0.0.8.3-x86_64-rdna3

Choose a tag to compare

@saaaaaaaaaaaaaaaaaas saaaaaaaaaaaaaaaaaas released this 09 Nov 08:13

About

Upstream from https://gitlab.com/mochizuki0323/DLSSNR-AMD using GPT 6.1 Sol based on mauri870: https://github.com/mauri870/DLSSNR-RDNA3.

2 new models work including CC_SMALL_FP8_EXPERIMENTAL and CC_SILVER_AARDWOLD_OPTIMIZED.
16-bit history works witch halves VRAM usage.

Usage

End key opens NR Control

Install

bash install.sh "<folder with the game's exe>" --dll <nvngx_dlssnr.dll or zip>