Person Selector SAM3

PersonSelectorMulti can build body masks a few ways - SAM2, seed-growing, YOLO-driven carving. This node throws out most of that machinery and does the whole mask job with SAM3's text grounding. You supply the concept - "person", "face", "head", "upper body", even "backpack" - and SAM3 segments it. Non-overlapping masks by design, because text grounding produces them that way. No SAM2, no BiRefNet, no depth-overlap resolution needed.

It slots into the same position as PersonSelectorMulti: takes reference images, does the ArcFace matching with the same appearance-blend (match_weights, default 50/15/15/20), and emits PERSON_DATA for Person Detailer. The difference is entirely in how masks get made. Instead of a segmentation model plus carving tricks, SAM3 grounds each mask type with its own text prompt and confidence threshold - body at 0.15, face at 0.40, head at 0.30, hair at 0.40. BiSeNet still runs per-face for the fine facial subtypes (skin, eyes, mouth, neck, accessories), because those aren't SAM3-groundable.

The aux system is the fun part

aux_preset gives you upper_body / lower_body, clothing, hands / feet / arms / legs, headless_body, or custom - and custom takes any noun phrase. Want a "shoes" mask? Type "shoes". Regional detailing with zero extra model downloads beyond SAM3 itself. There's also an optional YOLO path (aux_yolo_model) that overrides the text-prompt aux pathway, with aux_yolo_sam_refine turning bbox-only detections into real pixel masks via SAM3 with SAM2 as fallback.

The inputs that matter

• sam3_model - from the LoadSAM3Model node. Note the pack doesn't ship the loader; you need whichever SAM3 pack provides it.
• current_image - supports batch input.
• auto_threshold, threshold, guaranteed_refs, match_weights - the same matching controls as Multi.
• aux_preset + aux_custom_prompt - the text-driven aux masks.
• aux_threshold - confidence cutoff for aux detection.

Outputs

person_data, face_masks, head_masks, body_masks, aux_masks, an annotated preview, plus similarities, matches, matched_count, face_count, and report.

Install

Same face deps as the rest of the pack, plus the SAM3 model and its loader:

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools
pip install insightface>=0.7.3 onnxruntime-gpu>=1.17.0 opencv-python>=4.8.0 numpy>=1.24.0

Common issues

Empty or mushy masks usually mean the threshold for that mask type is off - the per-type config is fixed internally, so for aux masks adjust aux_threshold and lean on a well-lit reference for the rest. If you connect both aux_yolo_model and a preset, YOLO wins - it overrides the text pathway entirely. And if the selector errors at load, the SAM3 model config isn't connected or the loader pack isn't installed; the pack happily tells you sam3_model is required.