AI Models (BYOM — Bring Your Own Model) is where you register a model, point it at a live camera or a recording, and start inference. You’ll find it under AI & Detections → AI Models (BYOM) in the sidebar.

[!NOTE] This tab requires the byom.manage permission. Creating, editing, deleting, and running models is superuser-only — non-admins can view the list but the form and action buttons are disabled.

[!IMPORTANT] The Cloud AI dialog on this page (Hugging Face inference for individual models) is not the same as the top-level Cloud tab, which handles RTMP restreaming and S3 upload of recordings. For that, see Cloud.

What you can do here

• Add a model that runs one of the local adapters (YOLOv8, YOLOv11, BLIP, InsightFace) or a configured cloud model.
• Assign a task (person detection, counting, scene description, face detection/recognition/verification, watchlist check).
• Point a model at a live camera feed or a recorded video.
• Start / Stop live inference, or Analyze a recording as a one-time job.
• Enable/disable, edit, or delete configured models.
• Manage Cloud AI credentials and cloud model definitions (Hugging Face).

Detections produced by a running model appear on the Detection Results page.

Add or edit a model

1. In the Add New Model form, enter a Model Name (placeholder e.g., Person Detector 1).
2. Pick an AI Model from the dropdown — see the model list below. Enabled cloud models appear under a separate ☁️ Cloud Models group.
3. Pick a Task (see the task list below).
4. Leave Enabled ticked.
5. Choose a Source Type: Live Camera Feed or Recorded Video.
6. Fill in the source-specific fields (see below).
7. Optionally add Additional Config (JSON) — placeholder {"confidence": 0.5}.
8. Click Add Model (or Update Model when editing; Cancel exits edit mode).

To edit an existing model, click Edit in its table row — the form scrolls into view pre-filled. To remove one, click Delete (asks for confirmation).

AI Model options

Value	Label (shown in dropdown)
default	YOLOv8
yolov11	YOLOv11
blip	BLIP (Image Captioning)
insightface	InsightFace (Face Recognition)

[!NOTE] The YOLOv8 option submits the value default — that is the KAI-C adapter slot the YOLOv8 adapter occupies, not a placeholder. Governed inference routes on the slot name, so the label and stored value intentionally differ.

Task options

person_detection (Person Detection), person_counting (Person Counting), scene_description (Scene Description), face_detection (Face Detection), face_recognition (Face Recognition), face_verify (Face Verification), watchlist_check (Watchlist Check).

Fields

Field	Required	Example	Notes
Model Name	✓	Person Detector 1	Free-text label
AI Model	✓	YOLOv8	Local adapter or an enabled cloud model
Task	✓	Person Detection	One of the tasks above
Enabled	—	(checked)	Model is active
Source Type	✓	Live Camera Feed	Live feed or recorded video
Additional Config (JSON)	—	{"confidence": 0.5}	Optional JSON of extra params; validated on save

Run against a live camera

When Source Type is Live Camera Feed:

1. Pick a camera in Assigned Camera (cameras with no RTSP URL are flagged (No RTSP URL)).
2. Set Inference Interval (seconds) — how often to run inference, 1–60 (default 2).
3. Save the model.
4. In the models table, use the ▶ Start / ⏹ Stop button in the Inference column to toggle live inference. A running model shows ● Running.

Fields

Field	Required	Example	Notes
Assigned Camera	✓		Select a camera to enable automatic inference
Inference Interval (seconds)	—	2	1–60 seconds

Run against a recording

When Source Type is Recorded Video:

1. Click Browse next to Recording File to open the recording browser and select a session/segment. The path fills in and shows ✓ Recording selected: ….
2. Save the model.
3. In the table, click 🎬 Analyze to start a one-time analysis job (shows ● Processing while it runs, and ⏹ Stop to cancel).

Recording analysis runs in the background; when it finishes you’ll see “Recording analysis completed! Check AI Detection Results page.”

[!WARNING] You must select a recording via Browse — an empty or invalid recording path is rejected on save.

Cloud AI (Hugging Face)

Click the Cloud AI button (top-right of the tab) to open the Cloud AI Providers dialog. It has two tabs: Credentials and Models. Cloud models you configure here become selectable in the model form’s ☁️ Cloud Models group.

Add a credential

1. On the Credentials tab, click Add Credential.
2. Choose a Provider — Hugging Face.
3. Paste your API Token (placeholder hf_xxxxxxxxxxxxxxxxxxxx). Tokens are encrypted at rest.
4. Optionally add Account Info (JSON, optional) — placeholder {"email": "user@example.com"}.
5. Click Add.

Deleting a credential also deletes any models linked to it (you’re warned first).

Field	Required	Example	Notes
Provider	✓	Hugging Face	Only Hugging Face is offered today
API Token	✓	hf_...	Encrypted at rest
Account Info (JSON, optional)	—	{"email": "user@example.com"}	Free-form metadata

Configure a cloud model

1. On the Models tab, click Add Model (disabled until at least one credential exists).
2. Select a Credential.
3. Enter a Model Name (placeholder My BLIP Model).
4. Enter a Model ID (placeholder Salesforce/blip-image-captioning-base).
5. Pick a Task: Image to Text, Image Classification, Object Detection, or Text Generation.
6. Optionally add Config (JSON) — placeholder {"num_beams": 5}.
7. Leave Enabled ticked, then click Add.

Field	Required	Example	Notes
Credential	✓		Links the model to a saved token
Model Name	✓	My BLIP Model	Free-text label
Model ID	✓	Salesforce/blip-image-captioning-base	Hugging Face model repo id
Task	✓	Image to Text	One of the four cloud tasks
Config (JSON)	—	{"num_beams": 5}	Optional inference params
Enabled	—	(checked)	Only enabled cloud models appear in the model form

📸 Screenshots to capture

The two existing screenshots (ai byom.jpeg and cloud ai byom.jpeg) are reused above. Optionally add:

• byom-model-form-live.png — the Add New Model form with Live Camera Feed selected, a camera assigned, and interval 2.
• byom-models-table.png — the Configured Models table with a running model showing ● Running.