IBB Pose — Run SDPose Estimation

This is the node you actually came for. The two loaders set everything up; IBBPoseProcessor is where the pose estimation happens, and it's the difference between "IBB Pose is a pack I installed" and "IBB Pose is in my ControlNet workflow." Feed it an image or a batch of frames, get back a skeleton overlay and OpenPose-format JSON - the raw material for pose ControlNet, pose-driven video, or OpenPose Editor edits.

The mental model is a three-stage pipeline, and knowing it saves you from blaming the wrong stage when output looks wrong:

1. Detect - find every person in each frame (bounding boxes).
2. Batch infer - crop each person and run SDPose on the crops in true batches (that's what batch_size is for).
3. Reconstruct - map keypoints back to original coordinates, draw the skeleton, build the JSON.

The inputs that matter

Required are few. ibb_pose_model and images are the obvious ones. Then:

• score_threshold (0.1–0.9, default 0.3) - minimum keypoint confidence before a point is drawn. Low it to 0.1–0.2 if you're losing limbs, raise it to clean up jitter.
• overlay_alpha (0–1, default 1) - 1 is pure skeleton on a black canvas (the classic ControlNet conditioning image); 0 would be the untouched photo, which you don't want here.
• batch_size (1–64, default 1) - persons processed per inference batch. For video with many frames, raise it; the tradeoff is VRAM.

The optional slots are where it gets interesting. You can drive detection with yolo_model (the fast default), grounding_dino_model plus a prompt (default "person .") and gd_threshold, or data_from_florence2 JSON from a Florence2 node. Detection priority runs Florence2 → GroundingDINO → YOLO → whole frame. For WholeBody models, keep_face / keep_hands / keep_feet (all default true) let you strip facial/hand/foot keypoints out of the output - handy when hands are noise and you only want the body line.

Outputs

• images (IMAGE) - the skeleton overlay, alpha-blended onto the original at overlay_alpha. Wire this into an OpenPose ControlNet.
• pose_keypoint (POSE_KEYPOINT) - OpenPose-format JSON per frame: keypoints plus confidence scores. Save it to disk for editing by turning on save_for_editor (writes to ComfyUI/output/poses/ with filename_prefix_edit) - the README explicitly points at ComfyUI-OpenPose-Editor-jd for that workflow. enable_confidence_filter governs whether low-confidence keypoints get dropped from the saved JSON.

There's also pose_scale_factor (0.1–10) for line/dot thickness, and scale_for_xinsr for Xinsr ControlNet scaling - niche, leave them alone until you have a reason.

Install and the gotchas

Base install is the pack-wide one: ComfyUI Manager → search "IBB_POSE", or clone https://github.com/IBB666/IBB_POSE into custom_nodes, then pip install -r requirements.txt and restart. Beyond that, which backend you need depends on your model type: pip install ultralytics for Body/OpenPose, onnxruntime (or onnxruntime-gpu) for WholeBody. Models auto-download on first use.

Where people get burned:

• Skeleton drawn over the whole frame as one blob - detection failed and it fell back to "full image as one person." Check your YOLO model and the image. This is the #1 "why is my pose broken" cause.
• Missing limbs / floating dots - raise score_threshold a touch, or check keep_face/keep_hands/keep_feet if you stripped parts deliberately.
• OOM on video - batch_size too high for your VRAM. Dial it back toward 1.
• WholeBody import errors - you skipped onnxruntime. The Body path needs ultralytics; the WholeBody path needs onnxruntime; neither is in requirements.txt.
• First run is slow - models are auto-downloading to ComfyUI/models/IBB_POSE/. Let it finish once.

One honest caveat: SDPose is a diffusion model doing pose estimation, so it's slower than a lightweight DWPose pass - you're trading speed for robustness. For ControlNet work where you're rendering a single conditioning image, that's usually a fine trade. For per-frame video preprocessing, batch wisely.