SAMDetector (combined)

SAMDetector (combined) takes the rough regions a bbox detector found and uses Segment Anything to turn each one into a mask that follows the actual outline of the thing. Boxes in, one clean silhouette-shaped mask out. It's the node that fixes the biggest weakness of box-based detection: a rectangle around a face also grabs a lot of not-face, and when you detail or composite through that rectangle, the edges show.

The canonical use is the combination from the Impact Pack's own examples: a BBOX detector finds the face and hands you SEGS, then this node runs SAM at each detected location to extract the precise segment - the face-shaped region, not the box. Intersect the two (a bbox mask AND a SAM silhouette) and you get a mask that's both correctly placed and correctly shaped. Feed that into a detailer and the fix aligns to the face's contour without bleeding into the background.

How it works

Segment Anything (SAM) is a promptable segmentation model: give it a point or a box on an image and it returns the object at that spot. This node walks the incoming SEGS, uses each detection as a prompt for SAM, and merges all the resulting segments into one unified MASK. "Combined" means exactly that - every detected region collapses into a single mask (the Segmented sibling keeps them separate). The detection_hint controls how the box is translated into SAM prompts - center point, a few points, the rectangle, and so on - which nudges SAM toward the right object when a plain center point would be ambiguous.

The inputs and outputs that matter

• sam_model (required) - a loaded SAM model, from SAMLoader (Impact). This is the model that does the segmenting.
• segs (required) - the detected regions to segment, from a BBOX/SEGM detector.
• image (required) - the picture SAM looks at.
• detection_hint (default center-1) - how each detection becomes a SAM prompt. center-1 (one center point) is fine for compact subjects like faces; the multi-point and rect options help when the center point lands on the wrong object.
• dilation (default 0) - grows or shrinks the final mask. Negatives erode.
• threshold (default 0.93) - SAM's confidence gate for including a segment. High by design; lower it if SAM is being too conservative and clipping the subject.
• bbox_expansion (default 0) - pads the box SAM is prompted with, useful when the detector boxed the subject too tight.
• mask_hint_threshold (default 0.7) and mask_hint_use_negative - advanced tuning for using the existing mask as a positive/negative hint. Defaults are fine to start.

Single output: MASK - one unified silhouette mask, ready for an inpaint node, a mask operation (like intersecting with the bbox mask), or a detailer's mask input.

How to install it

Install the pack via ComfyUI Manager (search ComfyUI Impact Pack, install, restart), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack comfyui-impact-pack
cd comfyui-impact-pack
python -m pip install -r requirements.txt   # ComfyUI's python; python_embeded on portable

then restart. The SAM weights (sam_vit_b_01ec64.pth) auto-download to ComfyUI/models/sams on first install; load them with SAMLoader (Impact). The SEGS feeding this node still needs a detector, which means the Impact Subpack for UltralyticsDetectorProvider. Linux: libgl1-mesa-glx and libglib2.0-0 for OpenCV.

Common issues & troubleshooting

The mask grabbed the wrong thing. A single center point can land SAM on a background object or the wrong part of the subject. Switch detection_hint to a multi-point option (rect-4, diamond-4) so SAM has more to go on, or tighten the upstream detector so the boxes are cleaner.

The silhouette is clipping the subject's edges. Lower threshold a little, or add a few pixels of dilation / bbox_expansion. SAM at 0.93 is deliberately strict.

It's out-of-memory or slow. SAM is a real model on top of your generation stack. If VRAM is tight, use the smaller ViT-B weights (the default) rather than ViT-H, and remember you can run the SAM editor on CPU via impact-pack.ini if you only need it occasionally.

Do I even need SAM here? Only when mask shape matters. For a face that's going to be feathered heavily anyway, the box mask from the detector alone is usually enough - reach for SAM when seams or background bleed are the actual problem.