SegmDetectorCombined batch

This node is a modified, batch-friendly rebuild of Impact Pack's own SegmDetectorCombined node - the author says so directly in the description. Two changes from the original: it runs across an entire image batch at once instead of one frame at a time, which matters a lot if you're segmenting video frames rather than a single still, and it supports both detector types - bounding-box and full segmentation - simultaneously, rather than making you pick one node for each.

The inputs and outputs that matter

Impact Pack, from ltdrdata, is the pack behind ComfyUI's whole detect-crop-refine pattern - the same family of ideas FaceDetailer uses to find a region, crop it, resample it properly, and paste it back. This node borrows that detection layer specifically: feed it an image (or a batch), a threshold (0.5 default) for how confident a detection needs to be before it counts, and a dilation amount (in pixels, can go negative to shrink) to grow or shrink the detected region before it becomes a mask. Wire in a bbox_detector and/or a segm_detector - both optional here, but you need at least one, and both come from Impact Pack's own loader nodes, not from this pack. Out comes one MASK, batched to match your input.

Installing it

Because the detectors themselves live in Impact Pack, this node is only useful once that's installed too - the README is explicit that you need Impact Pack for anything in this section of the pack. ComfyUI-WJNodes installs the usual way - ComfyUI Manager, search "ComfyUI-WJNodes," or:

cd ComfyUI/custom_nodes
git clone https://github.com/807502278/ComfyUI-WJNodes.git

But this specific node stays dark in your node list until Impact Pack (ltdrdata/ComfyUI-Impact-Pack) is also installed - grab that through Manager or its own repo, and restart once both are in place. You'll find this node under WJNode → Other-plugins → ImpackPack.

Where people get burned

Confirm you've actually got a detector loaded and wired into bbox_detector or segm_detector first - this node does nothing on its own without one. If detections seem too sparse, lower threshold; if the resulting mask hugs the detected object too tightly or too loosely, dilation is your knob, positive to grow the mask outward, negative to shrink it in.

If you're setting up YOLO-based bounding-box detection specifically, know that as of Impact Pack v8.0 that detector provider moved out into a separately-installed Impact Subpack rather than shipping by default - so if bbox_detector has nowhere to plug in, that's the piece to go grab. Worth a quick historical note while you're at it: that detector path depends on the Ultralytics package, which had a genuinely bad moment in December 2024 when a compromised release shipped a cryptominer that reached ComfyUI users through exactly this Impact Pack pathway. It's long since patched and isn't a reason to avoid the node - just a reason to keep Impact Pack and its Subpack updated rather than running a stale pin from that era.