# AMDNR runs DLSS 5 neural rendering on AMD cards through an OptiScaler build

> AMDNR is a patched OptiScaler distribution that brings NVIDIA's DLSS 5 neural rendering path to AMD GPUs, adding model interleave, residual composition, XeSS frame generation and FSR ray regeneration on top. It is aimed at players on RDNA 3 and RDNA 4 hardware, and the release notes show a project iterating weekly while still labelled alpha.

**3zwr1/AMD-NR---OptiScaler** — DLSS 5 Neural Rendering on AMD GPUs, built into OptiScaler: model interleave, residual composition, XeSS frame generation up to 6X, FSR Ray Regeneration. Discord:  https://discord.gg/AMDNR · Ko-fi: https://ko-fi.com/3zinr

- Repository: https://github.com/3zwr1/AMD-NR---OptiScaler
- Stars: 688 · Forks: 16
- Language: C++
- License: GPL-3.0
- Published: 2026-10-21 · Updated: 2026-10-21 · Language: en
- Canonical page: https://hysenlabs.com/projects/3zwr1-amd-nr-optiscaler

## AMDNR works by extending OptiScaler, not by replacing the game

The mechanism starts with a hook that already exists. OptiScaler is a proxy layer that intercepts Direct3D 12 games and substitutes an upscaler, and AMDNR is a build of OptiScaler with the DLSS 5 neural rendering path added, so it works in any Direct3D 12 game OptiScaler already hooks without each title needing native support. The additional capabilities are named in the project description: model interleave for a large frame rate gain, residual composition, XeSS frame generation unlocked up to 6X with an opt-in path to 10X in D3D12 games, and FSR Ray Regeneration for games that use DLSS Ray Reconstruction. The repository layout matches the two halves of the project, with a main directory holding the core build and a Launcher directory holding the desktop front end, plus an Anywhere directory for the games that have no upscaler of their own. The code is C++ and the licence is GPL-3.0, which is a notable choice for a project whose payloads include proprietary runtime files.

## Two runtimes, and the third party work inside the archives

The release notes are explicit about what is AMDNR's work and what is not, which is unusual and worth reading closely. The runtime that does the neural work in the Runtime archives, the files named dlssnr_amd_pass1 through dlssnr_amd_pass3, is Daniel Blanco's work, described as DLSS-NR on AMD, copyright 2026 Daniel Blanco, all rights reserved, and it is shipped unmodified with his permission. AMDNR states plainly that it is not AMDNR's work and asks users to support that project instead. Two other runtime names appear throughout the notes, danielblnc and lmxxf, with lmxxf credited to Kien under the MIT licence. The danielblnc runtime is the recommended one from 0.3.5.1 onwards and also runs on RX 6000 with AMD's HIP SDK installed, while lmxxf 0.41 is the default set on RX 9000. The distinction matters for anyone who wants to know what they are running, and it also explains why the ini keys are prefixed by runtime name, since the two behave differently and the menu hides controls a given runtime cannot use.

## AMDNR Anywhere reaches games with no upscaler of their own

Since 0.3.5, the project has offered a path for titles that ship no DLSS, XeSS or FSR 2 at all. The AMDNR Anywhere preview, as the notes call it, presents a single PLAY ANYWHERE button in the AMDNR Launcher, and the important detail is that nothing is written into the game folder. That is a materially different approach from the standard OptiScaler install, where files are dropped next to the executable, and it changes the risk profile. In the preview release the support is scoped to RX 9000, which the notes identify as RDNA 4, and the Anywhere directory at the top of the repository is where that mechanism lives. The launcher itself was redesigned during the 0.3.5 series, with thirteen languages and an ASK AI feature, and it also handles fetching the optional third party upscaler models from their own release pages, so the launcher is doing network work as well as process management.

## Frame generation at 6X, and why nothing may generate

The frame generation side is where the project is most aggressive. XeSS frame generation is unlocked up to 6X, with an opt-in route to 10X in Direct3D 12 games, and the 0.3.5.6 notes say the memory used by XeFG is now counted correctly, with the 6X setting included in that fix. The same release also changed the fallback behaviour: FSR 3 and FSR 4 should no longer drop back to FSR 2.1.2 when another upscaler choice fails or is not ready, so your own FSR selection is kept. The Frame Gen tab was given a diagnostic job, and the documentation says it names the five steps involved so that when nothing generates, the tab says why. That is a small design decision with a large effect on support, because the most common complaint about frame generation is a black screen with no explanation, and a tool that names the missing step is doing something different from one that silently does nothing.

## Ray regeneration, and the denoiser that comes with it

Ray reconstruction support is a separate tab in the interface, placed right after Upscaling, and its status lines report per card and per API what runs and why not. FSR Ray Regeneration is the mechanism used for games that rely on DLSS Ray Reconstruction. The support has moved through states quickly. On RX 7000 it was offered again as a preview from 0.3.5.1, on by default, with the AMDNR Ray Denoiser, itself a preview, denoising the result from 0.3.5.3, and the Upscaling tab checkbox or the setting to stop it returns the game to its own denoiser. The 0.3.5.6 notes list a menu change that removes guesswork: a control a runtime cannot run is hidden on that runtime, and its ini key keeps working, and the runtime entry names the lmxxf version that actually runs, such as lmxxf 0.41. Whether ray regeneration is worth enabling on a given card is therefore a per-card question the interface now answers instead of leaving you to experiment.

## The 0.3.5.6 pass changes defaults for RX 7000 and handhelds

The newest release is a good illustration of how this project is developed, because almost every item is a behavioural change with an ini key to turn it back. Full Model Look becomes the default neural rendering style, a fuller look on both runtimes, run after upscaling on cards that can afford it, with the cost gate and the TDR guard moving it back before upscaling where they must and saying so. Extra detail on faces is always on on both runtimes, disabled with AmdFaceDetail=false. Ghosting with lmxxf is reduced rather than gone, with shorter trails behind moving objects and the flash after a scene cut about half as strong. On RX 7000 and handhelds, neural rendering now runs one network size smaller by default in lmxxf Fast mode, which costs less time per frame, and AmdLmxxfFastMode=false switches it off, while RX 9000 keeps that mode off. There are also new controls for game exposure, an anti-halo setting that was formerly halo protection, refine passes on danielblnc, and a neural resolution that now goes up to 200 percent on both runtimes.

## Alpha tags, matched file pairs and a known-limits list

The distribution method carries its own constraints. Releases are tagged as alpha versions, with Alpha0.3.5.6 dated 2026-10-17 as the newest in the list, alongside Alpha0.3.5.5R4 from 2026-10-14 and Alpha0.3.5.5R3 from 2026-10-12, which is a release cadence of days rather than months. The install instruction in the notes is exact: replace every file of the zip together, with OptiScaler.dll and LmxxfNrRuntime.dll coming from the same release, and launcher users get that handled for them. Mixing files from two builds is the kind of mistake that produces crashes that look like driver problems. The notes also keep a Known limits section per version, referenced as still open for 0.3.5.6, alongside a CHANGELOG.md in the repository and release notes on GitHub. A REPORT-README.md at the top level sits next to the main documentation, and the project asks for bug reports with logs through its Discord, which is how it says the next build gets better.

## Where AMDNR is the wrong choice

The clearest case against is a Radeon user whose games already handle upscaling well and who wants a configuration their vendor supports. This is a community project publishing alpha builds, redistributing runtime files it does not own, and changing behaviour weekly, which is a fair trade for a player who wants a feature nobody else offers and a poor trade for a studio that has to certify what ships. The second case is a Direct3D 11 or Vulkan title, since the hooks described are for Direct3D 12 games and the Anywhere feature is scoped to that path. The third is hardware outside the tested range: the notes distinguish RX 9000, RX 7000, RX 6000 and RDNA 3 handhelds in different ways, and the defaults differ per card, so an unsupported card is likely to land on the slower path. The fourth is the licensing question itself, since a GPL-3.0 repository shipping proprietary derived binaries creates a distribution arrangement a company lawyer will want to look at, and the project is candid that the neural runtime is someone else's work.

## Conclusion

Owners of RX 7000 or RX 9000 cards who want neural upscaling and frame generation in a Direct3D 12 game and are willing to swap DLLs into a game folder are the people AMDNR serves, and the presence of an AMDNR Anywhere path means even a title without any upscaler can be reached through the launcher. It is the wrong tool for a Radeon user whose games already run FSR 4 well and who wants a supported configuration, and for anyone who cannot accept that the project ships NVIDIA derived runtime files alongside the third party work, since the danielblnc pass libraries inside the Runtime archives are Daniel Blanco's DLSS-NR on AMD build redistributed unmodified. Check three things before you install: that you replace OptiScaler.dll and LmxxfNrRuntime.dll together from the same release, that your card model is covered by the default network size the runtime will pick, since RX 7000 and handhelds run one size smaller by default while RX 9000 does not, and that the release you read about is the one you unpack, since the project publishes under alpha tags such as Alpha0.3.5.6 dated 2026-10-17 and the last push was on 2026-10-18.

## FAQ

### What does AMDNR do that plain OptiScaler does not?

It is an OptiScaler build that adds DLSS 5 neural rendering for AMD GPUs, plus model interleave, residual composition, XeSS frame generation unlocked up to 6X, and FSR Ray Regeneration. It works in any Direct3D 12 game OptiScaler already hooks.

### Can AMDNR work in a game with no upscaler at all?

Yes, through the AMDNR Anywhere preview introduced in 0.3.5, which uses a single PLAY ANYWHERE button in the AMDNR Launcher and writes nothing into the game folder. That preview release covers RX 9000, or RDNA 4.

### Which files must be replaced together when updating AMDNR?

Every file of the zip, with OptiScaler.dll and LmxxfNrRuntime.dll coming from the same release. Launcher users get the update applied for them.

### Who wrote the neural runtime that AMDNR ships?

The runtime in the Runtime archives, with the files dlssnr_amd_pass1 through dlssnr_amd_pass3, is Daniel Blanco's work, described as DLSS-NR on AMD and copyright 2026 Daniel Blanco. AMDNR states it ships it unmodified with his permission and asks users to support that project.

### What are Model E and Model L in AMDNR 0.3.5.6?

They are two experimental, opt-in upscaler choices made by other authors, Model E from HelixSR and Model L from d4r. The AMDNR Launcher fetches them from their own releases, and Model L is limited to the RX 9070 and RX 9070 XT.

## Sources

- [3zwr1/AMD-NR---OptiScaler on GitHub](https://github.com/3zwr1/AMD-NR---OptiScaler)
- [Issues](https://github.com/3zwr1/AMD-NR---OptiScaler/issues)
- [License: GPL-3.0](https://github.com/3zwr1/AMD-NR---OptiScaler/blob/main/LICENSE)
- [README](https://github.com/3zwr1/AMD-NR---OptiScaler/blob/main/README.md)
- [Releases](https://github.com/3zwr1/AMD-NR---OptiScaler/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/3zwr1-amd-nr-optiscaler