Load YOLO Model (AnotherUtils)

Load YOLO Model (AnotherUtils) is the front door to the pack's detection suite: pick a YOLO model, and it gives you a ready-to-run detector that you then feed to the pack's inference nodes. The part that actually deserves the "seamless" sticker is the download - the README's claim checks out. You select yolov8m.pt, and on first run the pack fetches the weights itself into ComfyUI's models/ultralytics (or models/another_utils) folder. No hunting through HuggingFace, no manual downloads.

The model list is generous: presets cover YOLOv8 and YOLO11 across the n/s/m/l/x size ladder, with detection, -pose, and -seg variants - so object detection, human pose estimation, and instance segmentation are all covered from one dropdown. On top of the presets, it scans your local models/ultralytics and models/another_utils folders and adds anything you drop in yourself, so a custom fine-tuned .pt shows up in the same list. device lets you force cuda, cpu, or let auto decide.

The output type matters more than it looks: it's ANOTHER_MODEL, a pack-internal type that only the pack's own inference nodes accept. So the flow is loader → AnotherYOLOInference, where you set a confidence threshold and get bounding boxes, keypoints, masks, labels and a debug image back. If you try to wire the output into Impact Pack or another detector, it won't connect - this is a walled garden, by design. That's the zero-dependency philosophy in action: the README emphasizes the inference logic runs on embedded internals rather than pulling in the ultralytics pip package, so it won't collide with the packs you already have. Given the December 2024 Ultralytics supply-chain incident that reached ComfyUI users through a poisoned package version, avoiding a second ultralytics install in your environment is a defensible choice, not just paranoia.

The gotchas are the practical ones. First run downloads a model, so you need an internet connection and a little patience - and yolo11x variants are big. Second, the loader is just the loader; it validates nothing about whether inference will succeed until AnotherYOLOInference actually runs. And auto device selection does its best, but on a machine with a dead or busy GPU, forcing cuda or cpu explicitly is the debugging move.

Where it fits in the wider picture: YOLO detection is the backbone of auto-detailing pipelines (Impact Pack's FaceDetailer machinery runs on it), and here it's the front end of a segmentation pipeline - detect with YOLO, then hand results to the pack's SAM2 nodes for precise masks. That chaining is the intended architecture.

Install: ComfyUI Manager → search "AnotherUtils", or cd ComfyUI/custom_nodes && git clone https://github.com/marcoc2/ComfyUI-AnotherUtils, restart. No pip requirements to resolve - that's the whole point of this node.