🎯 Cascade Detector Advanced (Bbox/Segm, Staged)

This is a single node that runs up to three detection passes - each with its own bbox or segm YOLO model - and hands you one combined SEGS output. The author describes it as either a "magnifying glass" or a "big detector," and both halves are right. Chained together, each stage re-detects inside what the previous stage found, which is how you reliably find small stuff that a first pass on the full image would miss. Run the stages in parallel and you've got a single node that merges three different detectors' results with NMS, the way you'd otherwise assemble with a pile of SEGSPreview and SEGSConcat nodes.

The genuinely useful trick is input label filtering. Each stage takes a comma-separated list of labels and only processes segments with those labels from the previous stage - so stage 2 can run eye detection exclusively on the face segments stage 1 produced, instead of pointlessly scanning arms and torsos for eyes. stage2_input_filter_labels = "face" is the whole workflow, right there.

How it works

Each stage has its own detector (bbox or segm), confidence, IoU threshold, dilation, class filter, crop factor, scale mode, and target/max size. In sequential mode, stage 2 detects inside the crops of stage 1; in parallel mode every enabled stage runs independently on the full image and the results are merged with NMS against the top-level iou_threshold; parallel_per_segment runs each detector independently on each segment.

The part worth knowing about is masks. When a downstream stage detects on an upscaled crop ("enlarged fragments"), the node recalculates each mask back into the original image's coordinate space as a numpy array - the author's flagship fix, and the thing that makes it work with non-square images like 1152×1280 and with Impact Pack's SEG objects, which historically hate mismatched mask shapes. There's also simplify_masks (morphological closing with an elliptical kernel) to smooth those recalculated masks out.

The inputs that matter

• mode - sequential vs parallel. Start here; it changes what the stage settings mean.
• stage_N_detector_type plus the matching stage_N_bbox_detector / stage_N_segm_detector - you wire Impact Pack's UltralyticsDetectorProvider into these. That's where the actual YOLO models (like face_yolov8m.pt) come from and download on first use.
• stageX_input_filter_labels - the headline feature. Comma-separated labels; empty means process everything.
• include_masks_in_output - keep this true. The README is blunt: most nodes that consume SEGS error out or misbehave without masks attached.

After the cascade, min_confidence, min_bbox_width, and min_bbox_height prune weak or tiny detections, and drop_size feeds the detectors' own filtering.

Outputs

segs_output_all_stages is the main one - wire it into SEGSPaste or SEGSDetailer from Impact Pack. preview_image draws the boxes (color-coded per stage), cropped_fragments_image is a contact sheet of every crop, masked_fragments_image shows just the masked interiors, and stage1/2/3_segs give you each stage's results separately. image_bypass is a nice fallback: it returns the original image when nothing was found and a black frame otherwise, so a downstream branch can skip gracefully when detection comes up empty.

Installing it

ComfyUI Manager - search "ComfyUI-Impact-BIGdetector-BBOX-SEGS" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/Devourer56/ComfyUI-Impact-BIGdetector-BBOX-SEGS

then restart ComfyUI. The pack ships no requirements.txt: it uses numpy, cv2, torch, and PIL, all already in ComfyUI. What you actually need is Impact Pack installed separately, plus its YOLO models.

Honest caveats

Set expectations before you build a masterpiece on this. The README's foreword admits the node was generated with AI in half a day "for my own needs," the previews are self-described as "so-so" (use Impact Pack's SEGSPreview), and the author's announcement post landed with zero engagement - this is a personal utility that happens to be public, not a maintained ecosystem staple. The stages are hardcoded to three, so don't expect a 5-stage pipeline without editing source. And since it leans on Impact Pack's Ultralytics stack, keep both updated - that dependency chain is exactly how the December 2024 compromised-Ultralytics cryptominer reached ComfyUI installs. All that said, if you've ever stacked three FaceDetailer-style passes manually, this node collapses the whole cascade into one block with per-stage settings you can actually read.