Detect Regions (YOLO)

This is the "find things" node of the LF region pipeline. LF_DetectRegions runs a YOLO object detector over your image, finds the objects, and hands you back structured metadata - bounding boxes, labels, confidence - ready to be turned into masks by LF_RegionMask or used to drive targeted editing. It's the automation side of ComfyUI: instead of painting a mask over a face, you let the model find the face, then decide what to do with it.

The workflow shape is detection → mask → edit. Detect the regions with this node, carve the mask with LF_RegionMask, then feed that mask into an inpaint or a regional-prompt setup. It's the same "targeted masking" idea as GroundingDINO+SAM, done with a single lightweight ONNX model.

How it works

Two required inputs: image and model_path - the latter expects an ONNX_YOLO file, ideally picked via the pack's LF_ONNXSelector so you get a dropdown instead of typing paths. Files live under ComfyUI/models/onnx. The node runs the detector through onnxruntime, applies non-maximum suppression, and decides which region to treat as "the selected one."

The knobs that matter:

• confidence_threshold (default 0.25) - detections below this get discarded. Raise it if you're catching junk; lower it if things you want keep disappearing.
• iou_threshold (default 0.45) - overlap cutoff for NMS, i.e. how aggressively duplicate boxes get merged.
• max_regions (default 5) - cap on regions kept per image.
• class_filter / class_labels - filter to specific classes by name or ID, and optionally override the label names (useful if your model exports generic class0 labels).
• preferred_label + selection_strategy - tell the node which region to auto-select. If a box's label matches preferred_label, it wins; otherwise it falls back to confidence or area selection.
• select_index - the escape hatch: -1 auto-selects, or set a concrete index to pick the Nth region manually.

Outputs

region_meta (the selected region's metadata) and region_meta_list (all detections as a list). Wire these into LF_RegionMask, which turns one of them into an actual MASK. The compare widget gives you a visual overlay of what got detected.

Installing it

lf-nodes installs as a pack:

cd ComfyUI/custom_nodes
git clone https://github.com/lucafoscili/lf-nodes

restart ComfyUI, or ComfyUI Manager → "LF Nodes". The dependency that matters is onnxruntime (in the pack's requirements). The YOLO model itself is not included - export one or download an ONNX detector and drop it in ComfyUI/models/onnx.

Common issues

• Model not found errors - the ONNX file isn't where model_path points. Use the ONNX selector to guarantee a valid path.
• Too many or too few boxes - that's confidence_threshold doing its thing. Tune it per image content; 0.25 is a floor, not a law.
• Wrong region auto-selected - set preferred_label to your target class, or bypass auto-selection entirely with select_index.