Py-feat Live — Changelog¶

Py-feat Live v0.8.15¶

Released 2026-06-18 · v0.8.15

What's new¶

Facial-region overlays redesigned (Detectorv2 / mesh detectors). The AU and ARKit-blendshape mesh overlays now use py-feat 2.0.2's non-overlapping, per-triangle region maps: whole, left/right-symmetric region silhouettes drawn on the MediaPipe-478 face mesh, with the eyes and mouth left open. The live overlay fills each region's triangles directly at its detected intensity (monochrome heatmap), replacing the previous subdivision-smoothed rendering — sharper region boundaries and lower per-frame cost.

AU overlay: 19 AUs (left+right merged).
Blendshape overlay: 33 skin-deforming ARKit blendshapes (left/right independent).

Under the hood. Pins py-feat==2.0.2; the bundled macOS runtime lock is recompiled for Apple Silicon (no CUDA wheels), with the platform requirement documented to prevent the 0.8.13-class first-run install regression.

Py-feat Live v0.8.14¶

Released 2026-06-17 · v0.8.14

Fixes¶

macOS first-run failure (regression in 0.8.13) — 0.8.13's bundled runtime lock was compiled on a Linux/CUDA host, so it pinned CUDA-only torch deps (cuda-bindings/nvidia-*) that have no macOS wheels; the app couldn't install its Python runtime on first launch. The lock is now resolved for macOS (CPU/MPS torch). A CI gate dry-run-installs the lock on macOS to prevent recurrence. 0.8.13 is superseded — use 0.8.14.

Facial-region overlays (Detectorv2)¶

New ARKit-blendshape mesh overlay — toggle Blendshapes in Live and the Viewer (monochrome heatmap, Left/Right split at the facial midline).
AU mesh overlay now uses py-feat's non-overlapping AU region map.
Pinned py-feat 2.0.1.

Py-feat Live v0.8.12¶

Released 2026-06-16 · v0.8.12

Live and Viewer fixes, plus a Detectorv2 head-pose yaw correction.

Live¶

Overlay alignment — the mesh, landmarks, box, and gaze now line up with the video. Fixed a vertical stretch that pushed the overlay up off the face ("eyes too high") on both detectors.
Facebox — no longer drawn too long around Detectorv2 faces.
Detector switch — switching Detectorv1 ↔ Detectorv2 keeps the right landmark style (mesh for v2, contour lines for v1) instead of getting stuck on lines.
Gaze direction — corrected for both detectors.
Head pose (Yaw) — Detectorv2 yaw is now derived from the stable 478-point mesh; it no longer flickers or reads the same sign when turning left vs right. Pitch and roll are unchanged.
Stop — stopping the camera now reliably clears the overlay (no frozen facebox/mesh).
Recording frame rate — recording no longer drops fps as severely; the recorder's frame conversion was moved off the detection path.

Viewer¶

Overlay/video sync — overlays now track the video for the full clip on variable-rate live recordings (mapped by the video's real per-frame timestamps instead of a fixed fps). Fixes overlays freezing partway through while the video kept playing.
First frame — fixed a mangled first-frame overlay and wrong identity thumbnail on live recordings.
Valence/Arousal toggle added to the overlay toolbar.
Session details — click the info icon on a session for metadata and a "Reveal in Finder" action.

App¶

Logs button now works on every tab (Live, Extract, Viewer), not just Live.
Dragging a video that misses the Extract dropzone no longer replaces the app with a bare video player.
Logo and title moved into the header so they stay visible on every tab.

Py-feat Live v0.8.11¶

Released 2026-06-15 · v0.8.11

Py-feat Live v0.8.10¶

Released 2026-06-14 · v0.8.10

Installation¶

Requires macOS on Apple Silicon (M-series).

macOS Apple Silicon: Py-feat.Live_*_aarch64.dmg

First launch downloads ~1.5GB of Python dependencies and ~1GB of model weights into your user data directory. Plan for ~10 minutes on a typical broadband connection. Subsequent launches are instant.

Py-feat Live v0.8.9¶

Released 2026-06-14 · v0.8.9

Installation¶

Requires macOS on Apple Silicon (M-series).

macOS Apple Silicon: Py-feat.Live_*_aarch64.dmg

First launch downloads ~1.5GB of Python dependencies and ~1GB of model weights into your user data directory. Plan for ~10 minutes on a typical broadband connection. Subsequent launches are instant.

v0.8.8¶

Released 2026-06-14 · Release notes

Stability + robustness release ahead of going public.

Detectors¶

Tracks py-feat v0.7-dev (3855e07): the modular detector is now Detectorv1 (renamed from Detector); old sessions/presets that recorded "Detector" still work via a compatibility alias.

Fixes (from the pre-public review)¶

Analyze: a detection/decode error no longer leaks the recorder (writer thread + CSV handle) or orphans a session dir; a corrupt video.mp4 is no longer left behind; one bad frame no longer fails a whole job.
Analyze WebSocket: guarded JSON parsing + auto-reconnect (the queue no longer silently freezes on a dropped connection).
Serving: session video/CSV stream instead of buffering the whole file into memory; malformed Analyze input returns 422 (not 500); uploaded files are cleaned up on "clear done".
Viewer / Live: resilient session load + update-check (no unhandled rejections); overlay guards; the timeseries last-frame off-by-one is fixed.

Cleanup¶

Removed dead Streamlit-era files/config/env and stale references.

This is also the first release to verify the corrected Info.plist (no longer requests local-network access).

v0.8.7¶

Released 2026-06-14 · Release notes

Fixes a launch crash in v0.8.5 / v0.8.6. Those packaged builds aborted on startup (a webview with label 'main' already exists) — a duplicate window definition. This release restores normal launch.

Same app features as v0.8.5/v0.8.6 (Detectorv2 v2.5, adaptive Live overlay, Viewer panels, grouped timeseries, Detectorv2-default extraction).

Fixed¶

A label-less window in tauri.conf.json defaulted to the main label and collided with the window the app builds in code, crashing launch. Removed the duplicate.

v0.8.6¶

Released 2026-06-14 · Release notes

Hotfix release. Same app features as v0.8.5 (Detectorv2 v2.5, adaptive Live overlay, Viewer panels, grouped timeseries, Detectorv2-default extraction) — this corrects the release pipeline so the auto-updater manifest (latest.json) publishes.

Fixed¶

The macOS updater-signature step searched a tripled bundle path (target/<triple>/release/bundle/macos) while the native build outputs to target/release/bundle/macos, which failed the deploy and skipped the latest.json publish in v0.8.5. Corrected.

v0.8.5¶

Released 2026-06-14 · Release notes

Upgrades to Detectorv2 v2.5 and brings the Viewer in line with Live.

Highlights¶

Detectorv2 v2.5 — 52 MediaPipe/ARKit blendshape coefficients, safetensors weights, corrected 478-mesh decode, and head-pose pitch now reads +up (matching the classic Detector / Detectorv1).
Live — the overlay pipeline is resolution/aspect-adaptive; it tracks correctly on any camera (no more mesh/overlay distortion on non-16:9 webcams).
Viewer — emotion, valence/arousal, and pose panels now match Live and sit beside each face; the timeseries variable picker is grouped into collapsible sets (Emotions, V/A, Pose, Gaze, AUs, Blendshapes).
Analyze — extraction defaults to Detectorv2 · standard; the classic detector is relabeled Detectorv1.

Fixes¶

Mesh overlay no longer collapses/distorts under the v2.5 model.
Head-pose pitch sign corrected to the canonical +up convention.
index.html served with no-store to prevent stale app bundles.

Py-feat Live v0.8.4¶

Released 2026-06-04 · v0.8.4

Installation¶

Requires macOS on Apple Silicon (M-series).

macOS Apple Silicon: Py-feat.Live_*_aarch64.dmg

First launch downloads ~1.5GB of Python dependencies and ~1GB of model weights into your user data directory. Plan for ~10 minutes on a typical broadband connection. Subsequent launches are instant.

Py-feat Live v0.8.3¶

Released 2026-06-03 · v0.8.3

Installation¶

Requires macOS on Apple Silicon (M-series).

macOS Apple Silicon: Py-feat.Live_*_aarch64.dmg

First launch downloads ~1.5GB of Python dependencies and ~1GB of model weights into your user data directory. Plan for ~10 minutes on a typical broadband connection. Subsequent launches are instant.

Py-feat Live v0.8.2¶

Released 2026-06-03 · v0.8.2

Installation¶

Requires macOS on Apple Silicon (M-series).

macOS Apple Silicon: Py-feat.Live_*_aarch64.dmg

First launch downloads ~1.5GB of Python dependencies and ~1GB of model weights into your user data directory. Plan for ~10 minutes on a typical broadband connection. Subsequent launches are instant.