IBB Pose — Load YOLO Model

Most IBB Pose workflows start here, even though this node looks like it's doing nothing. The IBBPoseProcessor doesn't estimate pose on a whole image blindly - it first finds each person, crops them out, and runs the pose model on the crops. That's what lets it handle multi-person frames and video without the skeleton spaghetti you get from one giant pass. The IBBYOLOModelLoader supplies that first step: the detector that answers "where are the people?".

How it works

It's about as thin a loader as this pack gets. The model_name dropdown is populated live from whatever YOLO files you have sitting in ComfyUI/models/yolo/ - there are no hardcoded choices, which is why the node's schema shows an empty list. Pick one, it instantiates the Ultralytics model, and returns a YOLO_MODEL you wire into the processor's yolo_model input.

One thing worth knowing so you don't pick wrong: the processor only uses YOLO for person bounding boxes (class 0, confidence above 0.5). It doesn't use YOLO's own keypoint output at all - the actual skeleton comes from SDPose. So the model you want is one that detects people well. The pack's README suggests the pose variants (yolo11n-pose.pt, yolo11m-pose.pt, yolo11x-pose.pt), and those work fine because a pose model still emits person boxes. For a single video frame or still, the n (nano) is plenty and fast; step up to m/x only if people are consistently missed.

Installing it

Base install is the pack-wide one - ComfyUI Manager, search "IBB_POSE", or clone into custom_nodes and pip install -r requirements.txt. This node additionally needs Ultralytics, which the pack deliberately keeps optional:

pip install ultralytics

Then drop your YOLO weights here:

mkdir -p ComfyUI/models/yolo
# put yolo11n-pose.pt (or similar) in there

The gotchas worth knowing

• License. Ultralytics is AGPL-3.0, which the KB flags as a real consideration if you ship or sell anything built on it. It's the standard choice for ComfyUI detection and nobody bats an eye for personal use, but it's worth knowing it's not MIT.
• Missing detection? The processor silently falls back to GroundingDINO, then Florence2 data, then - worst case - treats the whole frame as one person. If you're getting full-frame crops with sloppy skeletons, your YOLO isn't finding anyone, and that's a detector problem, not a pose problem. Try a bigger model or a better input.
• No ultralytics installed → you'll hit an ImportError the moment the loader runs. That's the missing pip install above, not a pack bug.