PCVR streaming for Meta Quest 3, with PyroWave decoding directly on the Adreno GPU.
An experimental ALVR fork for exploring high-quality, low-latency streaming over USB and fast local networks. Vulkan handles wavelet decoding; ALVR supplies SteamVR integration, tracking, controllers and audio.
Download · Get started · Settings · Performance · Troubleshooting · Build / contribute
Research preview. SteamVR video, audio and tracking have worked in tested setups. Sustained native-resolution 120 fresh FPS and lower latency than Virtual Desktop remain unproven. A selected refresh rate or a 120 FPS counter does not establish 120 distinct displayed frames.
| Track | What to expect |
|---|---|
Published preview: alpha.8 / .51 |
Matching APK and Windows server with bounded frame waiting, automatic decode priority, independent resolution controls and safety fixes. Release notes |
Development: .55 |
PR #9 integrates overlay modes, independent game/stream controls, optional foveation profiles and diagnostics. Tested 2b289f8 passed short Quest screens near 119 fresh FPS. Amended 9da8560 passed all five signed CI jobs and artifact review; Bilinear stays the default. Measurements and limits. |
Development: .57 (PR #10) |
SteamVR runs at 240 Hz: the PC switches the Quest 3 panel to its 240 Hz scaled mode over USB when 240 Hz is chosen, and restores it afterwards. Three measured streaming profiles (native 120, 207, 240) and eye-image invalidation by default. Short screens: about 119 fresh FPS at native 120 Hz, 196 at 207 Hz and 223-230 at 240 Hz with 1280x1376-1440x1536 per eye. Results and limits |
| Working experimental target | 2080 × 2208 per eye · 120 Hz · 1000 Mbps · 4:2:0 · no foveated encoding, with Vulkan Compute decoding. Sustained acceptance is still pending. |
| Defaults preserved | Conservative 400 Mbps / 72 Hz candidate; 4:2:0, TCP and Quest 3 Auto → Compute. Development keeps synchronous decoding and the 4 ms selection wait. |
Publication notifications, prerecording and experimental fence handoffs remain off by default. Exact final-color fusion is included as an opt-in experiment; its earlier 120 Hz screen showed no delivery gain. Haar H2 remains excluded. Foveation, thread hints and eye invalidation remain off, and Bilinear remains the default filter. A 15-second no-decode probe reached 206.9 FPS at 207 Hz; this is runtime cadence, not 207 Hz streaming. See the integration review and whole-stack scorecard.
| Capability | What it enables |
|---|---|
| Quest GPU decode | Custom Vulkan wavelet reconstruction on Adreno, with measured decode and completion times. |
| USB and Wi-Fi | ALVR wired TCP forwarding or LAN streaming. UDP is a separate experimental path. |
| Bitrate control | A 5–2000 Mbps slider, high-bitrate presets and latency-driven Auto bitrate. |
| Independent render / encode sizes | Render a larger PC source while keeping the Quest's decoded pixel count fixed. |
| Chroma choices | 4:2:0 by default; optional 4:4:4 for controlled quality comparisons. |
| Local performance overlay | A client-rendered 3D panel: both-thumbstick clicks cycle Compact / Full / Hidden; a hold opens settings in .55. |
| Reproducible research | Pinned upstream sources, matching platform builds, regression checks and published measurements. |
Decoded Vulkan images cross an AHardwareBuffer bridge into GLES/OpenXR eye images for presentation. Pipeline and ownership details.
PyroWave uses a custom wavelet codec and GPU reconstruction. H.264, HEVC and AV1 PCVR paths use conventional video codecs and the headset's hardware video decoder. High bitrate and full chroma are useful research options, but spare bandwidth does not make PyroWave decoding free. This project has not established a matched quality or latency advantage over those codecs or Virtual Desktop.
The latest published preview is v0.1.0-alpha.8.
| Download | Purpose |
|---|---|
| Quest APK | Install on the Quest 3. |
| Windows server ZIP | Dashboard, SteamVR driver and required bundled files. |
| SHA-256 checksums | Verify the downloads. |
Read the alpha.8 release notes for its setup and optional
direct-eye-copy mode. For development builds, choose a successful full run in
GitHub Actions
and download both Quest3-Pyrowave-Android and Quest3-Pyrowave-Windows artifacts
from that same run. Tests-only runs produce no installable pair.
Do not mix an APK and server from different releases or runs. Pull-request APKs use temporary signing keys; main/release signing can differ. See APK signing and updates before switching builds. Keep your previous matching pair for rollback.
- Meta Quest 3, with Developer Mode and USB debugging enabled for sideloading.
- Windows PC, SteamVR and a Vulkan GPU/driver supporting the required Vulkan–D3D11 interop. SteamVR and the encoder must use the same GPU.
- USB data cable and preferably a USB 3 port, or a fast LAN with the PC connected to the router by Ethernet.
- Stop SteamVR before changing server installations. Extract the Windows ZIP into a new folder so the previous installation stays available.
- Run
ALVR Dashboard.exefrom the extracted folder. Register this server using its driver controls and enable the ALVR SteamVR add-on. - Install
Quest3-Pyrowave-dev.apkwith SideQuest or ADB. The app has its own package,io.github.jms1717.quest3pyrowave. - On Quest, open Quest3 PyroWave from Unknown Sources. On PC, explicitly trust the discovered headset in the dashboard, then start SteamVR.
- In Settings → Presets, start with Quest 3 PyroWave 400 Mbps / 72 Hz candidate. Once that works, try the 600 Mbps / 90 Hz candidate before the higher-rate experiments.
These are starting candidates, not guaranteed performance levels. The full installation and rollback guide covers prerequisites, controller emulation and restoring another driver.
| Connection | Setup | How to confirm it |
|---|---|---|
| USB / TCP | Plug in a data cable; enable Devices → Wired Connection. Use PyroWave TCP and the custom client package above. | The wired client is Streaming, the server peer is 127.0.0.1, and ADB forwards include ports 9943/9944. USB guide |
| Wi-Fi / LAN | PC and headset on the same LAN; prefer wired PC Ethernet and a nearby access point. | Confirm the active peer and measure actual video payload. Bitrate and network budgets |
A connected cable alone does not prove the stream uses USB. USB link speed and Wi-Fi PHY speed are also different from usable payload throughput. USB avoids the radio hop, but still includes ADB forwarding, GPU decode and compositor work.
| Setting | Where / recommended starting choice |
|---|---|
| Bitrate | Settings → Presets. The slider sets a fixed payload cap; with Auto it sets a maximum. |
| Auto bitrate | Uses network/encoder latency feedback to lower the requested rate when needed. It cannot guarantee FPS or remove a GPU bottleneck. Details |
| Profiles | Settings → Presets → Streaming profile: the three "(measured)" profiles set refresh, stream size, wavelet, chroma and bitrate together; the game render resolution is kept. Measurements |
| Refresh | Select a runtime-confirmed mode. Restart SteamVR after changing refresh. 240 Hz needs the Quest 3 scaled panel mode (1552x1664 per eye); choosing 240 Hz turns on Quest 3: switch to 240 Hz over USB. Capability detection |
| Decoder / wavelet | Quest 3 Auto chooses Compute. Haar/Compute is the measured development recipe; follow the matching build's instructions. Decoder findings |
| Chroma | Keep 4:2:0 for the baseline. 4:4:4 increased decode cost in the recorded comparison. Chroma comparison |
| Foveation | Off by default. Light peripheral encoding is optional development work, with sustained and in-headset acceptance pending. |
| Overlay | On by default. Click both thumbsticks together, release both, then click again to toggle. Metrics and overrides |
| Controllers | Settings → Headset → Controllers → Emulation mode → Quest 3 Touch Plus; restart SteamVR. |
The 600 / 800 / 1000 / 1500 / 2000 Mbps high-refresh profiles are experiments. At 120 Hz, 1000 Mbps permits approximately 1.04 MB per encoded stereo frame; 2000 Mbps permits approximately 2.08 MB. These are payload ceilings, not guaranteed utilization or quality. More bitrate can increase decode and network cost. Profile math and measured bitrate comparisons.
90/120 Hz and extended 144/207 Hz requests are runtime-gated. An accepted refresh
does not imply an equal fresh-frame rate. 240 Hz exists only in the developer
display-scaling mode, which the PC applies over USB (or tools/quest3/refresh_scaling.py
by hand); do not assume every OS/runtime exposes the same modes.
The two resolutions are independent:
| Per-eye size | Role |
|---|---|
| SteamVR render recommendation | Game source quality, subject to SteamVR percentages and game settings. |
| PyroWave encoded resolution | Stream dimensions and Quest decode workload. |
From the source checkout, with the matching server running:
python -m tools.quest3.control resolution --profile supersampled3072This requests 3072 × 3216 for the PC source while keeping encode/decode at 2080 × 2208 per eye. ALVR aligns the source recommendation to 3072 × 3232. Restart SteamVR afterward; games may need restarting too. Record the game's actual render size rather than assuming it equals the recommendation.
Restore native geometry with --profile native2080. These controls preserve
bitrate, refresh, chroma and decoder settings. Apply overall dashboard presets
first, then independent geometry, since those presets can set both sizes together.
Resolution semantics and A/B procedure.
Recent stationary native-resolution USB screens used 120 Hz, 1000 Mbps, 4:2:0, no foveation, Haar/Compute, one synchronous worker and a 4 ms selection wait.
| Observation | Evidence / interpretation |
|---|---|
| About 117 distinct targets/s | Recent .48/.50 screens, despite an ALVR client counter around 120. Sustained fresh 120 remains unmet. Whole-stack scorecard |
| GPU decode p50 5.90 ms; conversion p50 0.77 ms | .50 classification window. Native decode-to-fence p50 was 7.98 ms; these measure different intervals. Stage findings |
| PC source about 120.7 frames/s | Two .48 control windows, with encode median about 2.1 ms and network estimate 4.3–4.5 ms. These are not additive optical latency measurements. Stage breakdown |
| 4:4:4 was substantially slower | Earlier matched comparison: about 100 fresh FPS at 1000 Mbps / 4:2:0 versus 66 at 2000 Mbps / 4:4:4. Keep 4:2:0 as default. Full comparison |
| Larger PC source, unchanged Quest decode | Separation worked in short native/larger/native screens; no sustained FPS or repeatable quality/latency win established. Resolution experiment |
These are short screening measurements, not sustained gameplay, optical FPS or motion-to-photon tests. Earlier pacing screens still had p1 near 60. ALVR's estimated pipeline latency is not a measured motion-to-photon result.
In these captures, frames were still decoding when selection expired, while the PC supplied enough frames. Longer waits, publication notifications and prerecording did not establish a delivery gain; defaults stay unchanged. New shader candidates need correctness evidence before promotion. Read the whole-stack scorecard, producer findings and reviewed results.
| Symptom | First action |
|---|---|
| Dashboard will not open | Launch it from the extracted server folder with its bundled files. Avoid mixing installations. |
| Audio works, but video is black or corrupt | Check APK/server provenance, selected codec and foveation settings. Preserve logs and report exact settings. |
| USB is plugged in, but video uses Wi-Fi | Check the wired peer and ADB forwards using the USB verification steps. |
| Game never appears in the headset | Check SteamVR's active headset and the game's OpenXR runtime. OpenXR setup and VD rollback |
| FPS is below selected refresh | Compare fresh-frame delivery, GPU decode and completion timing. Increasing bitrate alone may make it worse. |
| Overlay will not hide | A benchmark's forced-visible override can take precedence over the controller chord. Clear the override. |
| Wrong controller model | Select Quest 3 Touch Plus. Older saved sessions or game-specific meshes can differ. |
| APK update fails | Compare signing certificates; PR builds can use different keys. Follow the signing guide. |
Returning to Virtual Desktop: stop SteamVR, disable ALVR in SteamVR's Manage Add-ons, and preserve Virtual Desktop's installation and registration. If you changed OpenXR selection, follow the runtime rollback guide. Keep the previous matching APK/server pair when changing project installations.
| I want to… | Start here |
|---|---|
| Build an APK and Windows server | Pinned build recipe, or run Actions → Quest3-Pyrowave → Run workflow in your fork. Full runs generate matching artifacts, hashes and license notices. |
| Benchmark a change | Benchmarking guide: short controlled screens first, then sustained and image acceptance for promising candidates. |
| Understand the pipeline | Architecture, decode pipeline, presentation research. |
| Continue development | Engineering handoff, whole-stack scorecard and agent restrictions. |
| Report a bug | Open an issue with build/run identity, GPU/driver, connection, per-eye sizes, refresh, bitrate and visible symptom. Remove secrets and private identifiers from logs. |
Pinned inputs live in sources.lock.json, modifications in
patches/, native bridge code in tools/pyroclient/, Quest controls and analyzers
in tools/quest3/, and regression checks in tests/. Upstream trees are
reconstructed from these pinned inputs during builds.
Documentation-only updates do not create new APK/server artifacts. Build success, on-device correctness, sustained performance and in-headset quality are separate gates. Contributions should make clear which were actually checked.
Maintained by JMS1717. Built on Terminal-ennui's Galaxy XR ALVR/PyroWave research, ALVR, and PyroWave / Granite by Hans-Kristian Arntzen (Themaister). Upstream authors retain credit for their work.
This project's changes are MIT-licensed; dependencies retain their own licenses. See LICENSE, NOTICE and the preserved upstream README. Independent project; not affiliated with Meta, Valve, Qualcomm or ALVR.
If this research is useful to you, help fund development and testing:
