Skip to main content

Vigil AI Detection

Vigil is O.D.I.N.'s built-in AI print failure detection system. It analyzes camera frames in real time using on-device ONNX inference to detect print issues before they waste filament or damage your printer.

What It Detects​

Detection TypeDescription
SpaghettiExtruded filament not adhering to the model, producing tangled strands
DetachmentPrint separating from the build plate mid-print
First-Layer IssuesPoor first layer adhesion, warping, or inconsistent extrusion
Build Plate EmptyNothing detected on the build plate when a print should be present

How It Works​

Camera → go2rtc → Frame Capture → ONNX Model → Detection → Action
  1. go2rtc streams video from your cameras (RTSP, USB, or Bambu built-in cameras)
  2. Vigil captures frames from the go2rtc stream at regular intervals
  3. Each frame is passed through an ONNX object detection model running locally
  4. If the model detects an issue, a confirmation buffer validates the detection
  5. On confirmed detection, Vigil triggers the configured action (alert, auto-pause, or both)

All inference runs locally on the O.D.I.N. host. No frames or data are sent to external services.

Setup​

Prerequisites​

  • At least one camera configured in go2rtc and streaming
  • The camera must have a clear, stable view of the print bed
  • The printer must be linked to a camera in Settings > Printers > [Printer] > Camera

Enabling Detection​

  1. Navigate to Settings > Vigil AI
  2. Enable detection for each printer that has a camera assigned
  3. Detection starts automatically when the printer begins a print job
info

Vigil runs per-printer in its own thread. Enabling detection on many printers simultaneously increases CPU usage. On low-powered hosts, consider enabling detection only on printers running unattended or high-value prints.

Threshold Tuning​

Each detection type has an independent sensitivity slider (0-100):

  • Higher sensitivity — Catches more issues but may produce more false positives
  • Lower sensitivity — Fewer false alarms but may miss subtle failures

Recommended starting points:

Detection TypeDefaultNotes
Spaghetti70Reliable. Lower if you get false positives from support structures.
Detachment60Can false-trigger on raft edges. Lower if using rafts frequently.
First-Layer50Highly dependent on camera angle and lighting. Tune carefully.
Build Plate Empty80Very reliable. Only lower if transparent filaments cause issues.

Adjust thresholds under Settings > Vigil AI > [Printer] > Thresholds.

Confirmation Buffer​

To avoid false positives, spaghetti, detachment and build-plate-empty alerts need 2 of the last 3 frames above the threshold before Vigil confirms an issue. First-layer alerts fire on a single frame because they are time-critical. Each failure type then has a 60-second alert cooldown.

The confirmation buffer ensures:

  • Transient anomalies (lighting changes, camera vibration) are ignored
  • Only repeated detections trigger spaghetti, detachment and build-plate-empty alerts or pauses
  • You can adjust the buffer length per detection type for more or less conservative behavior

Auto-Pause​

When enabled, Vigil automatically pauses the print on a confirmed detection. This prevents wasted filament and potential printer damage.

Auto-pause behavior:

  • The pause command is sent immediately to the printer via its native protocol
  • An alert notification is dispatched to all configured channels
  • The detection frame is saved for review in the Detections page
  • The print remains paused until a user manually resumes or cancels it
warning

Auto-pause sends a pause command through the printer's protocol (MQTT for Bambu, REST for Moonraker/PrusaLink, SDCP for Elegoo). If the printer's communication is interrupted, the pause may not reach the printer. Auto-pause is not a substitute for physical safety measures like smoke detectors and thermal cutoffs.

tip

Enable auto-pause for unattended prints. For attended prints, you may prefer alert-only mode so you can visually confirm before deciding to pause.

Detection History​

All detections (confirmed and unconfirmed) are logged and viewable at Detections in the sidebar. Each entry includes:

  • Timestamp
  • Printer and detection type
  • Confidence score
  • Captured frame image
  • Whether auto-pause was triggered

Frames are stored at /data/vision_frames/{printer_id}/ inside the container.

Vision Models​

ONNX models are stored at /data/vision_models/ inside the container. Default models are copied from the image on first boot.

To use a custom or updated model:

  1. Place the .onnx file in /data/vision_models/
  2. Register it via Settings > Vigil AI > Models
  3. Assign it to printers as needed
caution

Custom models must follow the same input/output tensor format as the default models. Incompatible models will fail silently or produce inaccurate detections. Test thoroughly before deploying custom models to production printers.

See Also​