Shibiko AI - Cascade Detection

Need to find faces, eyes, or anime faces in an image and turn them into masks or boxes - without pulling in a YOLO model? That's this node. It wraps OpenCV's built-in Haar cascades, the old-school machine-learning detectors that have shipped with OpenCV for over a decade, and hands you an image with boxes drawn, a mask, and the bounding boxes themselves.

The licensing angle in the README is worth taking seriously: YOLO weights are AGPL, which is a non-starter for commercial work unless you buy a license. Haar cascades are bundled with OpenCV (BSD-licensed) and there's no model download at all. The author's pitch - "you can use this as an alternate way for inpainting or in combination with other masking techniques" - is accurate. It pairs naturally with the pack's BboxInsertImage, or with mask-based inpainting from the KB's playbook, and it runs on CPU in milliseconds.

How it works

At startup the node scans two places for .xml cascade files: whatever you've dropped into ComfyUI/models/cascades, and OpenCV's own embedded cv2.data.haarcascades folder - that's where the 17 options like frontalface_default, eye, frontalcatface, fullbody, and license_plate_rus_16stages come from. Detection is classic detectMultiScale (scaleFactor 1.1, minNeighbors 5, min size 32px). For each hit it draws a rectangle on a copy of the image, and builds a mask by filling each box, then applying your chosen blur (gaussian, median, or box) and dilation to feather and grow the region.

The inputs that matter

• cascade - the dropdown. frontalface_default is the safe starting point; eye is handy for close crops.
• blur (default 32) and blur_type (box/gaussian/median) - how much the mask edges are softened. This is your feather control.
• dilation (default 4) - grows the mask beyond the detected box, so an inpainting pass covers the whole face.
• padding (default 4) - how far the drawn rectangle sits outside the box.

Three outputs: image (boxes drawn), mask (feathered, dilated regions), and bbox (the coordinates, ready for BboxSplit or BboxInsertImage).

Install and the fresh-install gotcha

cd ComfyUI/custom_nodes
git clone https://github.com/Shibiko-AI/ShibikoAI-ComfyUI-Tools

Or ComfyUI Manager → "Shibiko". Here's the real-world trap: the node's startup code tries to copy built-in cascades from an assets/cascades folder into ComfyUI/models/cascades - but that folder isn't shipped in the repo, so on a truly fresh install the whole pack can fail to load with a FileNotFoundError if models/cascades doesn't exist yet. If that happens, create the folder yourself before launching:

mkdir -p ComfyUI/models/cascades

One more README-vs-reality note: the README says Nagadomi's lbpcascade_animeface is bundled, but as of the current tree it isn't in the repo. It's a free download from Nagadomi's GitHub - grab the XML and drop it in models/cascades, and the anime face option appears in the dropdown. Also set expectations: Haar cascades only handle frontal, roughly upright faces. Profiles, heavy angles, or heavy occlusion and it'll miss them - that's the trade for being free and CPU-fast.