Skip to content

Sitr v0.5.2 — Part thresholds and visual video controls

Choose a tag to compare

@tantour tantour released this 20 Oct 07:45
· 2 commits to main since this release

Choose when an image part should cover the entire part or body, split chest and
belly independently, and control video coverage with the same visual editor as
images. Optional Advanced video mode adds independent body-skin, face and hair
masks while keeping the existing person detector.

New and improved

  • Per-part image thresholds: select an area and percentage, then Hide entire
    part or Hide entire body. The percentage is exposed body skin divided by the
    estimated visible part area, including clothing. A rule triggers strictly
    above its percentage. Set the part to Show for conditional-only coverage;
    ordinary Show/Hide actions still apply.
  • Threshold face action: body rules offer Black (the default) or Show. Show
    reopens the semantic face and uniquely associated face box after expansion,
    overriding that person's ordinary face effect while the rule triggers. Black
    takes precedence if multiple triggered body rules disagree.
  • Chest and belly: separate approximate halves use existing shoulder/hip
    landmarks and dense anatomy to follow body orientation, including sideways,
    lying-down and upside-down cases. Saved torso choices migrate to both areas;
    missing or ambiguous anatomy retains Unresolved torso.
  • Advanced video mode: off by default and explicitly marked heavy on
    performance
    . MediaPipe runs on each analyzed video frame alongside either
    YOLO silhouettes or fast person boxes. Body skin excludes faces; face and hair
    have separate Show/Hide choices. All three masks share one semantic pass.
    Videos do not run BodyPix part segmentation or image thresholds.
  • Advanced video performance: remove the inherited 4–8 Hz cap, analyze
    available frames as quickly as inference permits, reuse reliable person motion for at
    most 250 ms between detector passes, and capture large face inputs only when
    classification is due. Request high-performance MediaPipe graphics and report
    the active semantic GPU adapter. Add a playing-video benchmark.
  • Visual video editor: Female/Male tabs, clickable person diagrams, keyboard
    area selection, action buttons and advanced Show all / Hide all. Separate
    expansion respects shown regions. Standard whole-person effects are retained
    when advanced mode is switched off; image/video policies stay independent.
  • Refreshed README, user guide, English/Arabic screenshots and regression checks.
    Model weights, artifact hashes, required permissions and Chrome minimum are
    unchanged from 0.5.1.

Install or update

Download sitr-0.5.2-chrome.zip and SHA256SUMS.txt, verify its SHA-256,
and extract into a permanent folder. In chrome://extensions, enable Developer
mode and Load unpacked the folder containing manifest.json. Requires
Chrome 148+; models and runtimes are included.

To update, replace the files in the same extension folder, click Reload,
and refresh website tabs. This preserves the unpacked extension identity and
saved settings. Wait for startup preparation to finish. Advanced video mode
remains off until enabled.

Developers: sitr-0.5.2-source.zip contains matching source;
sitr-0.5.2-models.zip contains pinned models, original export inputs and
licenses. The 0.5.1 model files are also compatible because hashes are unchanged;
0.5.0 and older bundles do not meet the restored BodyPix output pins. Publish
all three ZIPs and SHA256SUMS.txt together under tag v0.5.2.

Validation

Initial release preparation passed TypeScript, 189 unit tests, all 18
browser checks
, and refreshed English/Arabic UI screenshots. The performance
follow-up expands unit coverage to 194 tests and reruns affected browser
checks; see the validation record for current evidence. npm audit reported zero vulnerabilities.
Checks include both packaged video detectors with independent semantic masks,
threshold face overrides, cached policy edits, model fallback and local WebRTC
streams. Archives include matching source, pinned models and SHA-256 checksums.

Performance, privacy and limits

Advanced video adds a semantic inference pass and can substantially slow analysis,
especially on CPU or with several videos. Face/hair selections reuse that pass;
they do not start additional models. Standard mode preserves the existing
whole-person video path. This release introduces no new model weights, network
services, analytics or accounts; media processing stays on device.

Skin ratios, anatomy and appearance labels are fallible estimates of visible
content. Uncertain/missing anatomy can prevent a threshold from triggering.
Mask expansion can cover nearby clothing, and motion can outrun sampled video
masks. These controls do not guarantee complete coverage or zero-frame exposure.
Meet validation uses local WebRTC fixtures rather than an actual meeting.

See release validation, area behavior,
user guide, and the
README screenshots.
Original Sitr files use AGPL-3.0-only; third-party assets retain their licenses.

Full Changelog: v0.5.1...v0.5.2