QproFaceTracking V2.1.1 Hot-Fix

Before you install

A rooted Quest Pro and the latest VRCFaceTracking on Steam are required. Extract the entire QproFaceTracking V2.1.0 Fix Test.zip folder before opening the Hub. GitHub's automatically generated source archives do not contain the ready-to-run app.

Follow the PDF guide included in the ZIP, or use the text setup instructions.

Bug fixes

• AMD ROCm setup: Added checks for the required PyTorch host, device and SDK packages. Missing torchgen or incomplete host-package metadata triggers a targeted repair. Python import failures are now reported separately from GPU detection failures.
• AMD GPU selection: Improved recognition of Radeon driver name variants, including RX6700XT. Qpro checks actual HIP device order, excludes integrated graphics and verifies that the selected card matches the installed GPU target. GPU visibility settings inherited from another application are cleared only within Qpro's ROCm process.
• Cheek calibration: Calibration now reads the original cheek signals and records relaxed, left, right and both-cheek poses. It measures overlap between the sides, supports unequal channel ranges and rejects weak or unstable samples with clearer retry messages.
• Calibrated cheek movement: Improved intermediate puff strengths and uneven two-cheek puffs. Added a short confirmation period when switching sides and reduced movement remaining on the relaxed cheek.
• Hub shutdown: Closing the Hub waits for tracking cleanup and stock eye-model restoration, then stops the configured PC ADB server. Pending status checks cannot restart it. Other Android tools using that shared server will disconnect.
• Gaze preparation diagnostics: Unsupported engines now report the exact firmware build and explain that tracking was left unchanged. A read-only diagnostic mode reports compatibility information without copying headset models. This does not add support for the reported Horizon OS V2.4 engine.
• Setup guide: Updated the PDF and text instructions to explain using one independent gaze method at a time, the temporary headset freeze when applying the Hub's gaze method, and what Stop tracking restores.

Updating and testing

Extract this build into a new folder. Close VRCFaceTracking, install the matching Virtual Desktop or Steam Link module from this Hub, then reopen VRCFaceTracking. The new cheek calibration requires the updated module's raw cheek feed.

Use Calibrate cheek puff on Live tracking and complete all four poses. Each streaming app keeps its own profile. Existing profiles remain compatible, but recalibrating uses the new method.

Working tongue models from V2.0 onward can be exported through Model manager in the old version and imported into this build as a .qptonguemodel file. No tongue retraining is required for these fixes. Preserve your original folder and recordings while testing.

Test status

This is a fix test build. Offline cheek, GPU routing, ROCm package repair, ADB shutdown and gaze preparation checks pass. Live cheek calibration and physical RX 6700 XT testing remain pending. AMD ROCm support remains experimental; a mapped discrete Radeon, a compatible AMD driver, Windows 11 and internet access are needed for a new installation.

QproFaceTracking V2.1.0 — AMD/NVIDIA + Wireless + Steamlink

Before you install

A rooted Quest Pro and the latest VRCFaceTracking on Steam are required. Download the QproFaceTracking V2.1.0.zip release asset and extract the entire folder before opening the Hub. GitHub's automatically generated source archives do not contain the ready-to-run app.

Follow the PDF guide included in the ZIP, or use the text setup instructions.

Close VRCFaceTracking before installing, switching, or uninstalling a Qpro module. Reopen it afterward. Installing the module for your selected streaming app removes the other Qpro source module.

What's in V2.1.0

• Steam Link support: Choose Virtual Desktop or Steam Link in the Hub and install its matching module. Steam Link requires eye and face data sharing with OSC output set to 9015. Its native tongue-out tracking also works when Qpro camera tracking is off.
• Individual cheek tracking: Added separate cheek puff and cheek suck controls. Cheek puff offers Calibrated, 1/0, and Balanced modes, with guided calibration and bundled developer baselines for both streaming apps. Turning either feature off restores the streaming app's original values.
• Eyebrows and smirks: Added an eyebrow adjustment toggle and sensitivity control, plus stronger one-sided smirk response.
• Tongue training: Added Focused diagonals + facial hair capture and expanded Full dataset capture. Quick and Focused refinement extend the tongue model selected in Live tracking.
• Mustachio model: Added an optional, highly experimental model that extends developer v8 with a bearded and moustached wearer's recording. Developer v8 remains the default. Mustachio has limited validation, may perform worse for some people, and cannot recover tongue details fully hidden by hair.

Bug fixes and improvements

• PC runtime setup: Bundled private Python so setup can work with or without an existing Python installation. Added clearer Activity progress and non-interactive PowerShell launches to avoid hidden setup prompts.
• GPU selection: Improved discrete GPU selection and CPU fallback so integrated graphics are not selected for Qpro's GPU training or inference.
• AMD ROCm setup: Added experimental ROCm 10.0 installation for mapped RX 6000, 7000, and 9000 cards. Setup downloads the required packages and marks the environment ready only after GPU training and inference checks pass. Fixed incorrect rejection of valid installations caused by confusing the HIP component version with the ROCm release, and added RX 6700 and RX 6650 XT mappings. Windows 11 and a compatible AMD driver are required; not every card has been live tested.
• Tongue movement: Improved extension and retraction smoothing, with a brief visibility hold to reduce flickering during short detection dropouts.
• Hub: Improved layout and scaling, added dataset deletion, and enabled only the module install button matching the selected streaming app.
• Independent Eye Gaze: Defaults to off, with improved headset model selection and recovery when gaze tracking fails. Firmware compatibility still varies. Applying or restoring gaze briefly restarts the headset's tracking service and can look like a crash.

Tongue model upgrades

If you have a working V2.0, V2.0.1, or V2.0.2 tongue model, export it from Model manager in your current version, then import the .qptonguemodel file through Model manager in V2.1.0. You do not need to retrain but is recommended. Models from versions before V2.0 should be recorded and trained again.

Known limits

Independent gaze and convergence remain experimental. The bundled eye profile passed its gaze checks, Relative pupil dilation is an experimental as well.

For help, share the relevant Activity lines in the community Discord. This server is not owned by Fwooffy.

[Superseded] QproFaceTracking V2.0.2 — Tongue training and ROCm setup hotfix

Before you install

A rooted Quest Pro and the latest VRCFaceTracking on Steam are required. Download the QproFaceTracking.V2.0.2.zip release asset and extract the entire folder before opening the Hub. GitHub's automatically generated source archives do not contain the ready-to-run app.

Follow the PDF guide included in the ZIP, or use the text setup instructions.

What's in V2.0.2

• Tongue training: Restored Python modules missing from the V2.0.1 release ZIP. Their absence caused both Quick refinement and Full dataset training to stop with ModuleNotFoundError: No module named 'tongue_visibility_calibration'.
• AMD ROCm setup: The Hub now reports ROCm as installed only after its GPU training and inference checks pass. The AMD installer also refuses to run on an NVIDIA-only PC.
• Tongue model upgrades: If you have a working V2.0 or V2.0.1 tongue model, export it from Model manager in your current version, then import the .qptonguemodel file through Model manager in V2.0.2. You do not need to retrain. Models from versions before V2.0 should be recorded and trained again.

[Superseded] QproFaceTracking V2.0.1 — AMD/NVIDIA + Wireless

Before you install

A rooted Quest Pro and the latest VRCFaceTracking on Steam are required. Download the QproFaceTracking V2.0.1.zip release asset and extract the entire folder before opening the Hub. GitHub’s automatically generated source archives do not contain the ready-to-run app.

Follow the PDF guide included in the ZIP, or use the text setup instructions.

What's in V2.0.1

• PC runtime setup: Fixed a PowerShell $HOME variable collision that stopped setup while checking an incomplete Python 3.12 installation.
• Independent Eye Gaze: Fixed an undefined received variable that crashed gaze tracking in both windowed and headless modes.
• Tongue model upgrades: If you have a working V2.0 tongue model, export it from Model manager in V2.0, then import the .qptonguemodel file through Model manager in V2.0.1. You do not need to retrain. Models from versions before V2.0 should be recorded and trained again.

[Superseded] QproFaceTracking V2.0 — AMD/NVIDIA + Wireless

QproFaceTracking V2.0 is a Windows release for a rooted Meta Quest Pro, based on Qpro-Enhanced-FT by n0tmast3r. Tongue tracking and relative pupil animation have been live tested on a Quest Pro. AMD ROCm was tested on an RX 7900 XTX, and NVIDIA CUDA hardware has also been live tested and is functional. The Hub supports tracking over USB or wireless ADB.

Before you install

• Required: a rooted Quest Pro and the latest VRCFaceTracking from Steam. Let Steam finish updating VRCFaceTracking before installing Qpro's bridge. Virtual Desktop and SteamVR are also needed for the full VRChat workflow.
• Download QproFaceTracking.V2.0.zip below and extract the whole folder. Open QproFaceTracking.exe from the extracted folder. GitHub's automatic Source code ZIP does not include the runnable program or its packaged assets.
• Follow Quest_Pro_Enhanced_Face_Tracking_Guide.pdf in the ZIP. If you prefer text, use RELEASE_INSTRUCTIONS.md. RELEASE_README.md has technical details.
• Coming from an older Qpro release? Record and train a new tongue model in V2.0. Keep the old model as a backup; importing it does not replace the new capture and training run.

What's in V2.0

• Lots and lots of bug fixes that I can't be bothered naming
• A resizable Hub that opens on First-time setup for a new extracted copy. The PC runtime, wireless ADB, VRCFaceTracking bridge, and optional AMD ROCm setup are handled inside the Hub.
• USB and wireless ADB connection choices, with a success or failure pop-up for wireless connection. Pairing is only needed if the Quest shows a six-digit pairing code.
• Tongue Quick refinement and Full dataset training, model management, and the option to delete old recordings.
• Optional independent eye gaze, convergence, relative pupil dilation, and camera preview controls. Live tracking reports AMD ROCm, NVIDIA CUDA, or CPU when the tongue model loads.
• Start tracking and Stop tracking controls. If the Hub's independent gaze method fails, it disables that option and attempts to restore the stock eye model while other selected tracking can continue.
• A combined Qpro VRCFaceTracking bridge with Install bridge and Uninstall bridge in the Hub. Troubleshooting launchers are in Helpers.

Independent gaze and firmware

Applying the Hub's temporary independent gaze model restarts Meta's headset tracking service. The Quest may appear frozen and briefly lose positional tracking while it restarts. Wait for the Activity log to confirm that tracking has returned. A similar pause can happen when Stop tracking restores the stock model.

Gaze and convergence support can vary by exact Quest firmware build. The tested AMD and wireless Quest Pro setup used build 51503870024400340. If the Hub method fails, the guide explains Singularity's Quest Pro Independent Eye Gaze Magisk module, reported working on Horizon OS v2.7. Use one independent gaze method at a time. Relative pupil values are estimates for avatar animation, not calibrated measurements.

Help and credits

If Windows will not delete an older extracted Qpro folder, close its Hub and camera preview. If the folder is still in use, end a lingering adb.exe under Task Manager > Details, then try again. This temporarily disconnects other tools using that ADB server.

Original project: n0tmast3r. Singularity: Lumince and contributors. The Quest Pro rooting guide was made by glorpette and Fwooffy. This edition and its beginner face tracking guide were prepared by Fwooffy. For help, share relevant Activity lines in the community Discord. The server is not owned by Fwooffy.

ZIP SHA-256: 0d046f8b4c0e95a91dcc68629d358018e23ca468cf9a520c31b71cb38d9bbd79

[Superseded] QproFaceTracking 0.1.10 AMD/NVIDIA Wireless Edition - ROCm installer fix

ROCm installer fix

Fresh AMD setup now continues when its new Python environment does not yet have OpenCV, NumPy, or PyTorch. Windows PowerShell 5.1 no longer treats that expected probe failure as a fatal installer error. Follow the setup guide.

Verified on an AMD Radeon RX 7900 XTX with ROCm 7.2.1 and PyTorch 2.9.1: both GPU training and GPU inference smoke tests passed. The release ZIP contains the same 74 clean files as the previous AMD/NVIDIA Wireless Edition package; only Install-QproRocm.ps1 and SHA256SUMS.txt changed. It includes no installed Python environment, captures, or training data.

When installed under a path containing spaces, AMD's offload-arch.exe helper may still print an android: Unknown command line argument warning. The training and inference tests passed despite that warning.

ZIP SHA-256: 9b7a9fccaf2f28d2deded89efb178fe6d36cf7722e1747fa9810225448090c5a

[Superseded] QproFaceTracking 0.1.10 AMD/NVIDIA Wireless Edition - wireless Hub fix

QproFaceTracking 0.1.10 AMD/NVIDIA Wireless Edition — wireless Hub fix

Download the ZIP asset below and extract the whole folder. GitHub's generated source archives do not include the runnable Hub. Follow the setup guide.

This patch fixes the Hub's “an authorized Quest connected by USB” prerequisite. The rebuilt Hub now checks the selected wireless ADB headset when showing connection status, preparing gaze, recording, and applying tracking. It also passes the saved wireless target to child scripts. The wireless connection and launch helpers recover a stale offline ADB transport.

On a Quest Pro with USB unplugged, the updated Hub showed Quest ADB: Connected, and Prepare gaze completed over Wi-Fi. The clean ZIP passed its self-test and contains no personal captures, saved headset address, or generated eye patch. A complete live tracking session was not rerun for this patch.

AMD ROCm support was previously tested on Radeon RX 7900 XTX. The NVIDIA CUDA and CPU paths remain available, but NVIDIA hardware was not live tested for this edition. Root and Magisk Shell access are required; use wireless ADB on a trusted private network.

ZIP SHA-256: 363f729cb1657e9140968ce388482d57a97e4f0237c5732fc9bbf04e79db859d

[Superseded] QproFaceTracking 0.1.10 AMD Wireless Edition

Superseded: Download the wireless Hub fix. This earlier ZIP can still require an authorized headset over USB when applying tracking.

QproFaceTracking 0.1.10 AMD Wireless Edition

An experimental fork of n0tmast3r/Qpro-Enhanced-FT for a rooted Quest Pro. Download the ZIP asset below, extract the full folder, and follow the included README.md or the online setup guide. GitHub's generated source archives do not include the runnable hub and models.

• Optional AMD ROCm runtime for tongue-model inference and training. Tested on Radeon RX 7900 XTX with PyTorch 2.9.1 and ROCm 7.2.1.
• Retains the original NVIDIA CUDA and CPU paths. NVIDIA hardware was not available for a live test of this edition.
• Adds wireless ADB setup, pairing, connection, and launch helpers. Wireless ADB/root, camera relay, and live AMD inference were tested with a Quest Pro on build 51503870024400340, with USB unplugged after setup.
• The ZIP excludes personal captures, image arrays, training caches, personalized model files, saved headset IP addresses, and generated headset eye patches. It includes the upstream developer's v8 demo tongue model.
• The three interface sounds were replaced with generated tones; see generate_sfx.py and THIRD_PARTY_NOTICES.md.

Root and Magisk Shell access remain required. Use wireless ADB only on a trusted private network. Eye convergence behavior can vary by firmware and headset; the newer firmware was probed, but a complete convergence session was not established as a general compatibility guarantee.

ZIP SHA-256: c10accfc3383f93b40787243fdb99766f5b6635657e84759436c2160ec4dc3e0