run yolo bboxs

This node doesn't run its own detection model - it runs Impact Pack's, but across a whole batch of images at once instead of one at a time. If you already know Impact Pack (ltdrdata's detect-crop-refine node pack, the one behind FaceDetailer), this is the utility node that plugs into that ecosystem to give you raw bounding-box data over a sequence rather than a single frame.

It's from ComfyUI-WJNodes, a personal utility pack - the author, publishing only as 807502278 on GitHub, says outright in the README that this and the sibling YOLO node require Impact Pack installed separately. That's a real dependency, not optional: without Impact Pack in your custom_nodes folder, this node has nothing to plug into.

What it's doing

YOLO (You Only Look Once) is a fast, well-established object detection architecture - Impact Pack uses it as one of its detector backends, loaded through Impact Pack's own BBOX_DETECTOR provider node (commonly UltralyticsDetectorProvider in Impact Pack, or its separately-installed Subpack as of Impact Pack v8.0 and later). This WJNodes node takes that already-loaded detector and an image sequence, runs detection frame by frame, and returns the bounding box results as a single bboxs object rather than making you loop the detector node manually across a batch.

That "batch" part is the actual value-add here. Impact Pack's own detector nodes are built primarily around single-image detection flows; running them across a full video sequence or a large image batch efficiently, and getting the results back in a form other WJNode utilities can consume, is what this wrapper is for.

Inputs and outputs that matter

• image - the batch to run detection over.
• bbox_detector - the loaded BBOX_DETECTOR object from Impact Pack. This node has no detector of its own; it entirely depends on Impact Pack supplying this.
• threshold - confidence cutoff, default 0.5. Lower it to catch more (weaker) detections, raise it to only keep confident ones.
• input_size - resizes images before running the detector, up to 4096px. This trades detection accuracy against speed/memory - a smaller input_size runs faster but can miss small objects.

The single output, bboxs, is a WJNode-specific bounding-box collection type, meant to feed onward into other nodes in this pack that consume that format (crop-by-bounding-box utilities, for instance) rather than Impact Pack's own SEGS type.

Installing it

Two dependencies here, not one. First, WJNodes itself:

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

(or via ComfyUI Manager, searching "ComfyUI-WJNodes"). Second, and required for this specific node to do anything: Impact Pack - install it the same way, through Manager or a manual clone of ltdrdata/ComfyUI-Impact-Pack. Note that since Impact Pack v8.0, the YOLO/Ultralytics detector provider moved out into a separately-installed Impact Subpack - if your bbox_detector loader node is missing, that's the piece to check for, not WJNodes itself.

Where people get stuck

The most likely failure mode isn't this node at all - it's the upstream detector. If Impact Pack (and, depending on version, its Subpack) isn't installed, you won't even be able to build a bbox_detector to feed this node, and the error will point at a missing type or missing node rather than anything obviously about WJNodes.

Beyond that, input_size is worth tuning deliberately rather than leaving at the default if you're detecting small objects in a large frame - downsizing to a small input_size before detection is a common way small subjects get silently missed. And because this pack has no public discussion trail, cross-checking Impact Pack's own documentation and issue tracker for detector-specific quirks (model choice, confidence calibration) is more productive than searching for WJNodes-specific answers that don't exist.