Person Selector Multi

This is the node the whole pack orbits. Person Selector Multi detects every face in your image, matches each one against up to ten reference slots with exclusive 1:1 assignment - a face can match at most one reference, so no duplicate assignments - and generates the full mask set for every matched person: face, head, body, plus optional hair, skin, eyes, mouth, neck, and accessories, all from a single BiSeNet run. The output is PERSON_DATA, which is exactly the structure Person Detailer eats.

Why exclusive assignment matters: in a scene with three people who resemble each other, naive max-similarity matching happily assigns the same face to two references. This does a proper 1:1 assignment, so each character gets their own mask and their own detailing slot. If you've ever watched a multi-person detail pass "fix" the same face twice, you know why this is the feature.

How it works

The differentiator is appearance-enhanced matching. Pure ArcFace struggles when faces look similar, so the node blends four signals - face identity, hair color (BiSeNet HSV), head appearance (histogram), and outfit color - weighted by match_weights, default 50/15/15/20. Faces too similar? Crank the outfit weight. Need pure face matching? Set 100/0/0/0. That tuning string is the single most important thing to learn on this node.

The quality-of-life wiring is deep. auto_threshold (default on) finds the optimal assignment and ignores the manual threshold; guaranteed_refs force-assigns the first N references so key characters always get detailed. A depth map (from Depth Anything V2 or similar) enables edge-carved masks, cross-reference deconfliction, and back-to-front render ordering - the preview even draws an isometric depth-layer diagram. A person_mask (BiRefNet, RMBG-2.0) hard-clips every mask to the foreground silhouette. YOLO models dropped into models/ultralytics/segm/ add body-part detection, stored as aux masks.

The inputs that matter

• current_image - supports batch input, each image processed independently.
• reference_1 through reference_10 - all optional; with none connected, every face goes to the generic slot.
• match_weights - the four-signal blend.
• guaranteed_refs - force-assign your key characters.

Outputs

person_data (→ Person Detailer), per-reference face/head/body masks, combined masks, aux_masks, an annotated preview, plus similarities, matches, matched_count, face_count, and a report string.

Install

Face deps as usual:

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

buffalo_l auto-downloads; parsing_bisenet.pth is manual; SAM/SAM3 and YOLO are optional.

Common issues

The wrong person gets the wrong slot → tune match_weights, or connect outfit_palettes from the color generator so the outfit signal has something to compare against. Everyone landing in the generic slot → no references connected, which is by design. Masks bleeding into the background → connect a person_mask or a segmentation detector. And if auto_threshold gives surprising results with strongly similar faces, the manual threshold path with guaranteed_refs gives you back control.