Veyra is a Windows GPU pipeline for video, capture and console streaming

One pipeline for local files, capture cards and console streams

Veyra is a real-time GPU video enhancement tool for Windows, described as turning local video, capture cards and game streaming into one real-time GPU enhancement pipeline. The tagline in the README is One engine. Any source. Real-time enhancement. What that means concretely is that the sources are heterogeneous. It handles local video and images, video capture cards, screen capture, PC streaming through Moonlight and Sunshine compatible paths, unofficial experimental Xbox support, and PS5 streaming. The enhancements applied are super resolution, denoising, HDR, frame generation and anti-aliasing, with the implementation offering DLSS, XeSS and FSR options. The primary language is C++, the interface is Qt with a QML layer, and the graphics backend is DirectX 12. It is a desktop application for a single platform rather than a library, so the audience is a person with a gaming PC or a capture setup who wants to watch or record video through hardware upscaling and denoising rather than through the decoder alone.

The node editor is the core of the new interface

The 2.1.0 highlights describe a rebuilt interface with Home, Cinema, Professional List, Nodes, Colour, Export and Settings screens, available in Simplified and Traditional Chinese, English and Japanese. The significant part is what the Nodes screen does. Node editing and layered NR are listed as headline features, with executable connections, independent NR and colour parameters, separate List and Node configurations, and sessions and presets. In other words, the enhancement graph is editable rather than a fixed pipeline of checkboxes. Alongside that, the NR controls gained a lot of detail in 2.0.4. Each NR layer has an Enhancement delta to Total strength control that reaches 5, with a Picture control that defaults off and, when enabled, offers Automatic or Manual modes. Automatic values adapt separately to styles 0, 1 and 2, and nine manual controls cover hue, chroma, neutral tint, source colour retention, lightness retention, shadows, highlights, local compression and temporal stability. The file structure backs this up, with qml/ for the interface, shaders/, services/, src/, include/ and a loop/ directory at the top level.

Strength beyond one is a trade, not a free win

The documentation on the NR strength control is the most honest passage in the README, and it should shape expectations before anyone starts raising sliders. Five extrapolates one inference's residual, and protection reduces the corresponding changes. The text says plainly that it cannot promise full gain five with no artifacts on every video. The Picture control makes the same point in a different form. An amount slider runs from 0, which retains raw strength, to 1, which applies full correction, and switching between modes keeps manual values while turning the control off preserves the original residual algorithm. The publishable claim is therefore about control rather than about quality. You can push the enhancement harder than a default profile would, and the documentation tells you exactly what the push costs. There is also a companion document, docs/NR_STRENGTH_PROTECTION_EXECUTION_2026-10-05.md, that holds the local validation and known limits, and it is the file to read before tuning rather than after.

Cost figures the release notes qualify themselves

The 2.1.0 release announcement includes new-effect cost examples on an RTX5070 with 4K anti-aliasing fixtures. FXAA is given at about 0.15ms and SMAA or MLAA at about 0.30ms. Final single-face mesh beauty is given at 1080p around 0.93ms and 4K around 1.42ms. The next sentence matters as much as the numbers: costs vary with content and hardware, and no local AMD speedup percentage was measured. That last clause is a limit on the AMD path rather than on the NVIDIA one, and it connects to the AMD NR entry in the release notes, which covers RX9000 Small and Original models through a per-layer performance switch that defaults on, and states explicitly that there is no new RX6000 or 7000 support. The AMD feature set is also narrower than the NVIDIA one. The release notes for 2.1.0 live at docs/RELEASE_NOTES_2.1.0.md and carry both the cost examples and the conditions.

Installing on Windows 11 with the right GPU package

Installation is a zip rather than a package manager entry. You download either Veyra-2.1.0-NVIDIA-win64-portable.zip or Veyra-2.1.0-AMD-win64-portable.zip for the GPU that is active on your machine, extract it with Windows or 7-Zip into a new writable folder, and run the executable:

veyra_qml_ui.exe

You need one GPU package to run the application, and the source assets are only for rebuilding. The stated platform requirements are Windows 11 x64 and DirectX 12. Qt and the approved runtimes are included in the package, while GPU and capture drivers are installed separately, and the backend hardware requirements vary. There are helper entry points in the repository as well, Veyra.cmd and run-test.cmd, and the build system is CMake with CMakePresets.json and a vcpkg.json for dependency management. The repository also carries a corresponding sources release, v2.1.0-sources, published alongside the binary release, and the README notes that the exact user-tested packages are promoted without rebuilding or repacking, which is why internal 2.1.0-test labels and manifests remain even though the shipped binary reports PE 2.1.0.0.

Capture and export are where the work is

The 2.1.0 changes to capture are extensive and they read like a response to real hardware problems. There is native enumeration with a fallback, driver range, step and custom size checks, a searchable legacy device list, exact rate and pixel format matching, explicit substitution rejection, and capture settings that can be saved and resumed. Network and audio handling got similar attention, with shared buffering and reconnect, a slow-server buffering status, HLS and interlaced damage handling, throttled subtitle scanning, vendor multichannel audio, non-PCM confirmation, complete-payload decoding and bounded recovery. The export workflow is a separate feature worth naming: an editable ordered queue, trimming, MP4 and MKV output, multiple audio and embedded subtitle tracks, cancellation and retry, and a completion sound. For a capture-card owner the export queue is arguably the feature that matters, because it turns a single-shot capture tool into a production pipeline.

What a different tool would have to do instead

The realistic alternative to Veyra is not another open source tool with the same feature list. It is a chain of separate pieces, which is how most people do this today. A capture card feeds a player or recording tool, upscaling runs in whatever the GPU driver or the game provides, and denoising is either on or off. The difference in approach is that Veyra puts the stages in one graph with one set of parameters, so the same profile applies to a local file and to a live console stream, and the export queue carries the settings into the file you keep. That single-graph property is what the node editor and the shared presets are for, and the documentation says presets, List and Node sessions, PNGs and video exports share the settings. The cost of that integration is platform lock-in to Windows and DirectX 12, and the cost of not integrating is that every stage is tuned separately and the settings do not travel with the output.

Licence, releases and the state of the project

The project is released under GPL-3.0, with a LICENSE file at the root and a THIRD_PARTY_NOTICES.md alongside it, which matters for a repository that bundles NVIDIA and AMD runtimes and other third party code. The repository is not archived and the last push was on 2026-10-20. The release pattern shows the project publishing a binary and a corresponding sources tag together, with v2.1.0 and v2.1.0-sources on 2026-10-19 and v2.0.9-sources on 2026-10-17. That pattern tells you something about how this is developed and shipped: the sources are released to match the binary rather than the binary being built from the repository by users, and the README states the shipped packages are promoted without a rebuild. There is also a product specification document, VEYRA_PRODUCT_SPEC_V1.md, and an agent execution playbook at the root, plus a community knowledge base linked from the README that compiles guidance from four group chats on setup, common problems and capture cards, which is a reasonable place to look for the problems the release notes do not cover.

Editorial conclusion

Adopt Veyra when you have a Windows 11 machine with an NVIDIA or AMD card and want a single pipeline for local video, capture cards and console streaming, with an export workflow you can queue and trim. Do not adopt it if you need Linux, if you want frame generation without accepting the added latency and artefacts that the documentation says it can produce, or if you expect the AMD path to match the NVIDIA one, since the release notes state no local AMD speedup percentage was measured. Verify first that you download the GPU package matching your card, that Windows 11 x64 and DirectX 12 are present, that your capture card drivers are installed separately, and that you read the NR strength and protection limits before raising the enhancement delta.

Frequently asked questions

Official sources

1. Issues
2. License: GPL-3.0
3. Likely7/Veyra-NRVideo on GitHub
4. README
5. Releases