H3 Crop + Confidence Report (internal)

The most expensive mistake in Mask mode isn't a bad replacement - it's a bad track. If SAM 3 latches onto the wrong thing, or the crop box drifts across the shot, you only find out after the H3 render has already cost you the time and VRAM. H3 Crop + Confidence Report (internal) is the node that lets you check before you commit: it hands the frontend a JSON report of the crop boxes around the tracked subject plus SAM 3's per-object confidence, which the tracking preview draws as an overlay. Green means good; a warning with a concrete fix means don't render yet.

It's deliberately read-only. The docstring says it best: purely additive - it only reads the mask pipeline and writes a UI string, so wiring it into a real generation graph cannot change the output. That's a nice property for a diagnostic node, and it's why the pack can leave it in the graph without affecting results.

Inputs:

• bboxes (BOUNDING_BOX) - the per-frame crop boxes from the MVEx Subject Crop planner, in source pixels.
• track_data (SAM3_TRACK_DATA) - SAM 3's tracking output, including per-object confidence scores.
• masks (MASK, optional) - if connected, enables the subject-clipping and pixel-area readouts. The crop box itself stays meaningful without it.
• confidence_threshold (FLOAT, default 0.4) - the line below which a track is flagged low_confidence.

Output: a single report (STRING) containing the JSON payload, which the frontend consumes to draw the box and paint the verdict in plain words.

The mechanism is simple and grounded: it extracts the first box per frame, collects SAM 3's confidence scores (one value per tracked object, not per frame - 1.0 for painted-mask tracks), computes the minimum, and compares it against your threshold. A score below 0.4 flips the "weak track" flag. If you've got masks connected it also measures how much of the crop the subject actually covers, which is how it catches a crop that cuts the subject off.

This is one of the newest pieces of the pack - the tracking preview it powers landed in the 0.10.0 changelog, with the confidence readout and crop warnings right after. It's internal, so you won't wire it by hand, and it only exists in Mask mode, which needs the MaskVidExperiments pack plus the SAM 3.1 checkpoint (sam3.1_multiplex_fp16.safetensors in checkpoints/) - SAM 3 being the Meta lineage model that made text-and-click-driven concept segmentation standard in ComfyUI. If you want to tune it at all, the one honest knob is confidence_threshold: 0.4 is a reasonable default, but drop it toward 0.25 for low-contrast shots where SAM is honest about being unsure, and the preview stops yelling at you.