ComfyUI Node
Person Selector Multi
Matches multiple reference persons in the current image using InsightFace (ArcFace) embeddings, optionally combined with hair color, head appearance, and outfit color matching. Each reference input accepts a batch of images for one person. All references are optional — with none connected, all faces go to the generic slot. Supports batch input: pass multiple images and get PERSON_DATA for PersonDetailer. Faces are assigned exclusively: each face matches at most one reference. match_weights controls the scoring blend: 'face/hair/head/outfit' (default 50/15/15/20). Set to '100/0/0/0' for pure face matching, or '60/20/20/0' to disable outfit. Outfit matching: connect palette_preview images as a batch to outfit_palettes. Compares palette color distribution with detected clothing colors (BiSeNet). Built-in YOLO detection: select a model from aux_model to detect body parts. Detected parts are assigned to references via body mask overlap and stored as 'aux' masks in PERSON_DATA for PersonDetailer. Drop YOLO .pt models into models/ultralytics/segm/ to add more detectors.
Person Selector Multi
- current_image
- sam_model
- reference_1
- outfit_palettes
- person_mask
- depth_map
- sam3_model
- reference_2
- reference_3
- reference_4
- reference_5
- reference_6
- reference_7
- reference_8
- reference_9
- reference_10
- person_data
- face_masks
- head_masks
- body_masks
- combined_face
- combined_head
- combined_body
- aux_masks
- preview
- similarities
- matches
- matched_count
- face_count
- report
◄auto_thresholdtrue►
◄threshold0.40►
◄guaranteed_refs0►
◄aggregation▾►
◄mask_fill_holestrue►
◄mask_blur0►
◄det_size▾►
◄aux_mask_typenone►
◄aux_model▾►
◄aux_confidence0.35►
◄aux_label►
◄aux_fill_holesfalse►
◄aux_expand_pixels0►
◄aux_blend_pixels0►
◄match_weights50/15/15/20►
◄depth_edge_threshold0.05►
◄depth_carve_strength0.80►
◄depth_grow_pixels30►
◄body_mask_modeauto►
◄depth_sort_orderfront_last►
CategoryFVM Tools/Face
Inputs (36)
| Name | Type | Default | Description |
|---|---|---|---|
| current_image | IMAGE | Image(s) to search for faces. Supports batch input — each image is processed independently. | |
| auto_threshold | BOOLEAN | true | Auto: finds optimal 1:1 face-reference assignment (ignores threshold). Off: uses manual threshold. |
| threshold | FLOAT | 0.400–1 | Minimum cosine similarity to count as match. Ignored when auto_threshold is on. |
| guaranteed_refs | INT | 00–10 | Force-assign the first N references to their best matching face, ignoring threshold. 0 = off (default). 2 = Ref1 and Ref2 always get assigned. Useful when key characters must always be detailed. |
| aggregation | COMBO | How to combine similarity scores across multiple reference images of the same person | |
| mask_fill_holes | BOOLEAN | true | Fill holes inside the mask (closes gaps in segmentation) |
| mask_blur | INT | 00–100 | Gaussian blur radius for mask edges |
| det_size | COMBO | Face detection resolution — higher finds smaller faces but uses more VRAM | |
| aux_mask_type | COMBO | none | Additional mask type for aux_masks output (derived from BiSeNet, no extra cost) |
| aux_model | COMBO | YOLO model for body-part detection. Runs per batch image internally. Select a model from models/ultralytics/ to detect body parts (hands, persons, etc.). Detections are assigned to references via body mask overlap and stored as aux_masks in PERSON_DATA. none = no body-part detection (default) Drop .pt files into models/ultralytics/segm/ to add more models. | |
| aux_confidence | FLOAT | 0.350.05–1 | YOLO detection confidence threshold. Lower = more detections (may include false positives) Higher = fewer, more confident detections 0.25-0.35 recommended for most models. |
| aux_label | STRING | Filter YOLO detections by class label (substring match). Empty = keep all detected classes (default) Comma-separated: 'person' or 'leg,foot' Substring match: 'leg' hits 'Left-leg', 'right_leg', etc. The class list for the selected model is displayed above the Matching section after model selection. | |
| aux_fill_holes | BOOLEAN | false | Fill holes inside YOLO aux masks (closes interior gaps). Off by default — segm models usually produce solid masks. |
| aux_expand_pixels | INT | 00–100 | Dilate YOLO aux masks by N pixels (elliptical kernel). Useful when segm masks hug the silhouette too tightly for inpainting. 0 = no growth. |
| aux_blend_pixels | INT | 00–100 | Gaussian blur radius for YOLO aux mask edges (symmetric). Blurs ~N pixels inward AND outward from the current edge. If you want the grown shape to stay fully opaque, set aux_expand_pixels at least as large as aux_blend_pixels. |
| match_weights | STRING | 50/15/15/20 | Matching weight blend: face/hair/head/outfit. Controls how much each signal contributes to the final similarity score. 50/15/15/20 — balanced with outfit (default) 60/20/20/0 — no outfit matching 100/0/0/0 — pure face matching 40/15/15/30 — heavy outfit weight 3 values also work (outfit=0): 60/20/20 equals 60/20/20/0. Hair = BiSeNet hair color (HSV). Head = head crop histogram. Outfit = clothing region vs. palette color distribution. Values are auto-normalized, so 3/1/1/1 equals 50/17/17/17. |
| sam_modelopt | SAM_MODEL | SAM2 model from Impact Pack SAMLoader for body masks. Optional when sam3_model is connected. | |
| reference_1opt | IMAGE | Reference image(s) for person 1. Pass a batch of images of the same person for better matching accuracy. Optional — if no references connected, all detected faces go to the generic slot in PersonDetailer. | |
| outfit_palettesopt | IMAGE | Palette preview image batch for outfit color matching. Image[0] = palette for reference 1, image[1] = for reference 2, etc. If batch is smaller than number of references, remaining refs skip outfit matching. Connect palette_preview outputs from Color Palette Generator. Use match_weights with 4 values to control outfit weight (e.g. 50/15/15/20). | |
| person_maskopt | MASK | Foreground person mask (BiRefNet, RMBG-2.0, or similar). When connected, acts as a hard clip for ALL mask types — face, head, body, aux, and BiSeNet label seeds are zeroed outside this silhouette. Body masks are replaced entirely by the BiRefNet envelope (SAM/seed_grow skipped). Single mask [1,H,W] or per-image batch [B,H,W]. If batch size is 1 but current_image has a larger batch, the mask is broadcast to all images. Multi-person split: the foreground is divided per reference using depth (closest depth wins) when depth_map is also connected, or by face-center distance otherwise. When no references are connected, the split falls back to detected face count — each detected face becomes a pseudo-reference. | |
| depth_mapopt | IMAGE | Depth map batch from Depth Anything V2 or similar. Improves masks via edge carving and cross-reference deconfliction. | |
| depth_edge_thresholdopt | FLOAT | 0.050.01–0.3 | Depth gradient threshold for edge detection. Lower = more edges detected. |
| depth_carve_strengthopt | FLOAT | 0.800–1 | How strongly depth edges cut masks. 0=off, 1=full cut. |
| depth_grow_pixelsopt | INT | 300–200 | Gap filling between depth edges. 0 = no growing. |
| body_mask_modeopt | COMBO | auto | Body mask strategy: - auto: seed_grow (recommended) - seed_grow: BiSeNet + SAM seed, carved by image/depth edges - sam: legacy SAM-only body segmentation |
| depth_sort_orderopt | COMBO | front_last | Rendering order for PersonDetailer: - front_last: closest person rendered last (correct for depth_map with bright=near) - front_first: closest person rendered first (for inverted depth maps) - off: no sorting, uses slot order |
| sam3_modelopt | SAM3_MODEL_CONFIG | SAM3 model config (alternative to SAM2). Connect LoadSAM3Model output here to use SAM3 for body masks. When connected, SAM3 is used instead of SAM2 (sam_model). SAM3 supports text+point prompts and may produce tighter masks. | |
| reference_2opt | IMAGE | — | |
| reference_3opt | IMAGE | — | |
| reference_4opt | IMAGE | — | |
| reference_5opt | IMAGE | — | |
| reference_6opt | IMAGE | — | |
| reference_7opt | IMAGE | — | |
| reference_8opt | IMAGE | — | |
| reference_9opt | IMAGE | — | |
| reference_10opt | IMAGE | — |
Outputs (14)
| Name | Type | Description |
|---|---|---|
| person_data | PERSON_DATA | — |
| face_masks | MASK | — |
| head_masks | MASK | — |
| body_masks | MASK | — |
| combined_face | MASK | — |
| combined_head | MASK | — |
| combined_body | MASK | — |
| aux_masks | MASK | — |
| preview | IMAGE | — |
| similarities | STRING | — |
| matches | STRING | — |
| matched_count | INT | — |
| face_count | INT | — |
| report | STRING | — |