Trustity VisionX Model V1 / V1.1
First-party on-endpoint VisionX detector; TAO 1.0.36 ships Model V1.1 in dual mode with COCO.
Trustity VisionX Model V1 / V1.1
Starting with TAO 1.0.34, VisionX ships a Trustity-trained detection model that runs entirely on the endpoint - no cloud inference for the detector path.
From TAO 1.0.35 (Stable), VisionX uses dual detection: stock COCO and Trustity both run on each frame. A lock fires when either model confirms a configured class. Endpoint logs and portal alerts record which model(s) scored the hit.
From TAO 1.0.36, the Trustity dual partner is VisionX Model V1.1 - more motion, close-range, and hard-negative training on top of V1.
What ships in the installer
| Artifact | Role |
|---|---|
yolov8n.onnx | Stable baseline (COCO) - strong coverage for phone / handheld imaging at screen |
trustity-visionx-v1.1.onnx | Trustity VisionX Model V1.1 - first-party dual partner |
trustity-visionx-v1.1.labels.json | Class names: phone, camera |
Override for lab/debug only: environment variable TAO_VISIONX_MODEL forces a single ONNX path.
Classes and threat framing
Portal policy emits violation classes phone and camera. Older policies that still say cell phone continue to match phones on the agent (COCO label alias).
Product intent is broader than the label names: VisionX protects against visual-channel exfiltration - phones, dedicated cameras, lenses, and similar imaging devices aimed at a sensitive display. Model labels remain phone / camera today; coverage for additional imaging form-factors improves with later training iterations shipped in TAO builds.
- Phone - covered by both COCO and Trustity V1.1 in dual mode.
- Camera - Trustity class slot for dedicated imaging devices; coverage improves as the training set grows.
Operator notes
- Prefer TAO 1.0.36+ from Deployment for V1.1 + dual detect.
- In agent logs look for
VisionX models loadedwithtrustity-visionx-v1.1. - Lock response in the Cloud Portal (Faster lock / Balanced / More accurate) is the customer-facing control for analysis speed versus confirmation strictness. It maps to performance profile, consecutive frames, and confidence - see VisionX detection tuning.
- Model updates ship with TAO agent builds. They are independent of Cloud Portal version numbers.
Honesty for rollout conversations
V1.1 is a second training pass. Dual mode keeps production detection at least as good as COCO while Trustity weights improve toward (and beyond) that baseline on real Windows desks. Treat dedicated camera / lens coverage as improving over releases, not as finished for every imaging device on day one.