"DLSS-NR on AMD," an unofficial tool that runs NVIDIA's DLSS on AMD Radeon GPUs, has grown into an integrated toolkit that covers Super Resolution and Ray Reconstruction as well. Developer Daniel Blanco released Alpha 0.8.0 on October 18 (Japan time). It folds in the contents of the earlier 0.7.0 release and adds multi-frame generation of up to 4x. The project began with the neural rendering of DLSS 5, which adds lighting and material effects, and it has now reached the stage where users can pick and choose among the major DLSS features. Still, being able to run the models on a Radeon is one thing; being able to feed them the data they need in the games you actually play is another, and each comes with its own conditions.
Four Features, Each With Different Requirements
Alpha 0.8.0 pairs "Supermod," which lets you select each feature from an in-game overlay, with "NeuralHub," which manages installation. Super Resolution and Ray Reconstruction join the existing neural rendering, and frame generation is now selectable too. Based on the author's update notes and the README, the supported scope breaks down as follows.
| Feature | Role in the image | What Alpha 0.8.0 states, and conditions |
|---|---|---|
| DLSS Super Resolution (SR) | Reconstructs a high-resolution image from one rendered at a lower resolution | Supports DLSS 4.5 models M and L. Adds dynamic resolution; model L supports arbitrary scaling factors |
| DLSS Ray Reconstruction (RR) | Reconstructs an image from noisy ray-traced output | Model F has been ported. The game must supply DLSS-RR inputs |
| DLSS Frame Generation (FG) | Generates additional images between rendered frames | Up to 4x MFG. Requires DLSS-FG inputs and a setting that makes the game recognize an NVIDIA GPU |
| DLSS Neural Rendering (NR) | Adds lighting and material effects to a rendered image | Up to 3 passes per frame can be selected, with style and intensity set for each pass |
Even though all four major features are targeted, RR needs DLSS-RR inputs, and FG needs DLSS-FG inputs plus GPU spoofing. The RX 6000 series has only experimental support for NR.
As for GPUs, the author lists the RDNA 4 Radeon RX 9000 series and the RDNA 3 RX 7000 series as targets for all four major features. The RDNA 2 RX 6000 series is also within the project's overall scope, but for now it has experimental support for neural rendering only. Additional features are planned for the next major update.
In other words, just because the RX 6000 series is named doesn't mean it can use the same combination of features. Even for the RX 7000 and RX 9000 series, the supported scope comes from the author's own statements and is not a guarantee of the same image quality or speed across all models and games.
What the DLSS Setting Shows vs. What Actually Runs
OptiScaler acts as an intermediary between the upscaler a game calls and the upscaler that actually processes the image. For example, you can choose DLSS in a game's settings and have its inputs passed to FSR or XeSS. Even if the settings screen says "DLSS," the model running behind it isn't necessarily DLSS.
Blanco explains that for this project he reimplemented DLSS's neural network runtime for AMD from scratch. In addition to HIP kernels that compute on the GPU, he says he rebuilt the memory layout and the connection to the game. That makes the processing the author is aiming for different in kind from approaches that simply route DLSS inputs to another upscaling technology.
That said, this description of the implementation comes from the author. It is not an announcement that a third party has audited the internal code or confirmed output and performance identical to NVIDIA's implementation. Users must also obtain the necessary DLSS DLLs legitimately themselves, and this is not an official feature endorsed or supported by NVIDIA or AMD.
GPU spoofing, meanwhile, plays a role separate from the model computation. Its purpose is to make the game recognize an NVIDIA GPU so that DLSS processing paths and inputs become available. The Alpha 0.7.0 change notes record SR and RR input support using GPU spoofing, and the current setup instructions say FG requires both spoofing and dedicated inputs.
Separating the part that receives information from the game from the part that processes it also makes mixing with FSR and XeSS easier to understand. The author gives an example that combines neural rendering and FSR Super Resolution on top of XeSS input. However, this flexibility does not remove the need for the dedicated inputs that RR and FG require.
Why Ray Reconstruction Needs a Game's Internal Data
NVIDIA's Ray Reconstruction programming guide makes clear that RR is not a filter that simply sharpens the edges of a finished image. Its input requires a noisy ray-traced image. In addition, it takes information such as surface color, reflective properties, the direction a surface faces, and surface roughness.
It also uses data showing depth and object motion. For specular reflections, it needs either the motion of the reflected image or something like the distance to the point the ray hit. The reason is that it reconstructs the image using not only which parts of the screen moved, but also what is shown and what kind of surface the light came from.
This mechanism bears directly on how far the port can be used. The author describes RR as working in "games that provide DLSS-RR ray-tracing inputs" because the information the model is meant to receive has to be assembled first. Support for FSR-oriented RR inputs is listed as a future addition as of the release.
Therefore, even if the tool works in a DirectX 12 game that offers FSR or XeSS, that alone does not tell you whether DLSS Ray Reconstruction will work. The inputs needed to run Super Resolution differ from those needed to reconstruct a ray-traced image. Apart from the hardware limits that porting the model can overcome, the connection to the game determines the practical range.
Up to 4x MFG, and How to Read the 90–100 FPS Figure
NVIDIA's frame generation guide describes 4x MFG as a mechanism that adds three generated images to each frame the game renders. More images are displayed, but that doesn't mean game processing or input responsiveness gets faster by the same factor. The FPS before frame generation and the displayed FPS including generated frames need to be considered separately.
The "up to 3 passes" of neural rendering is a separate matter. According to NVIDIA's explanation of DLSS 5, NR takes one image in and outputs one image, enriching how its lighting and materials look. The newly added multiple passes likewise apply different styles and intensities to the same frame and do not represent an MFG multiplier.
In the current README, Blanco cites 90–100 FPS as performance on an RX 9070 XT at 1080p using the full HT model. However, the statement gives no game name or test scene, and it doesn't indicate the internal resolution or whether frame generation was on. It can't be read as a result with all four major features running together, nor as a measurement of the load from RR alone. Performance on par with the RTX 5070 Ti remains a goal the author has set.
The improvement figures in the update notes also have specific targets. Changes included in 0.7.0 are said to improve NR by 6–16% in DirectX 12 environments under heavy gameplay load, and 0.8.0 lists a 25% improvement for NR on RDNA 3. The comparison builds and measurement conditions are not detailed, so the two figures cannot be added together into a single improvement rate. Native DirectX 12 operation has also been introduced, reportedly reducing stalls caused by coordination between APIs.
With Super Resolution too, a model running and a model running efficiently on a given GPU are separate things. According to NVIDIA's official explanation, DLSS 4.5 uses a second-generation Transformer, with model M optimized for Performance mode and model L for 4K Ultra Performance mode. Both models can be used in other quality modes, but NVIDIA itself warns that the load is heavier on older RTX 20 and 30 series cards. This can't be applied directly as a performance forecast for Radeon, but the reason is the same: whether a feature is supported shouldn't be the only factor in your choice.
Easier Installation With NeuralHub, but Requirements Remain
NeuralHub manages in one place the roughly 150MB of weight files and DLLs that previously had to be copied into each game. The flow is to import the DLLs, detect games, and place a loader in each one, and Game Pass games are now supported as well. You need to prepare "nvngx_dlss.dll" for Super Resolution, "nvngx_dlssd.dll" for RR, and "nvngx_dlssg.dll" for FG. For NR, "nvngx_dlssnr.dll" build 310.8.0.0 or later is recommended.
FSR 4.1.1 Super Resolution is bundled with the tool, so it is no longer necessary to obtain an FSR implementation from the driver. That does not make the Radeon driver itself unnecessary, however. The README still lists Adrenalin 26.1.1 or later as a requirement. The supported environment is also Windows 11 and DirectX 12 games that support DLSS, FSR, or XeSS; Vulkan and Linux support for this project is planned for the future.
As an alpha release, it also has limitations. The new Epic Games launcher is not supported, and games sometimes can't be added manually. The author also warns that anti-cheat software may block the DLLs. Even though installation has become simpler, this has not turned into a tool that guarantees support for every game.
For Radeon users, the value of this update is that there are now more options for choosing Super Resolution and Ray Reconstruction feature by feature and comparing them in the same game. To turn that choice into practical use, it's worth checking not only fine detail in still images but also ghosting and flicker when the camera moves. If you can compare under matched conditions, including the speed before frame generation and the stability of frame pacing, it becomes easier to pick a combination that balances image quality and responsiveness on the Radeon you already own.