GD BBox Detect

GroundingDINO is genuinely good at one thing: "find the thing I just described in words" - and for pose work, that's "find the person." GD BBox Detect wraps that into a single node that detects, filters, and hands you clean bounding boxes plus a visual preview. It's the middle link of the pack's pose chain: GD_ModelLoader → this → the official SDPose processor.

The node exists because raw detector output is messy. A video frame often has three "person" boxes, and the official SDPoseOODProcessor wants the person. So detection is only half the story - the rest is the mode/strategy filtering that decides which box survives:

• pass_all - keep everything the detector found. For a single-subject clip this is often fine.
• by_area - keep the largest or smallest box per frame (strategy: largest/smallest/highest_score). Right for "the main character is always the biggest person on screen."
• by_position - keep only boxes in a region (left_half, right_half, center_third, top_half, bottom_half). Good when your subject is predictably on one side.
• track - pick the biggest person on the first frame, then follow that same subject across frames. The one to use when your main character isn't always the biggest.
• by_index - pick the Nth detected box. The author warns it's unstable across frames, so this is really for single-frame debugging.

The inputs that matter

• gd_model - from GD_ModelLoader. Required.
• image - your frame batch.
• prompt - default "person ." (note the trailing period; GroundingDINO wants full stop-delimited phrases). You can get fancy: "man in red shirt", "person wearing glasses".
• threshold - confidence cutoff, default 0.3. Lower catches more with more false positives; raise it if you're tracking a stray "person" that isn't one.
• batch_size - frames per inference batch. The tooltip has real VRAM guidance: 4GB→1, 6GB→2, 8GB→4 (default), 12GB+→8-16. This is a GPU memory dial first, a speed dial second.
• preview_count - how many frames get previewed. Doesn't affect filtering.

Outputs: bboxes (a BOUNDING_BOX list for SDPoseOODProcessor), preview_images (frames with boxes drawn, so you can eyeball whether detection worked), and preview_info (a string summary). The preview outputs are the sanity check that saves you from piping garbage into a 20-minute pose pass.

Install

Part of user2318/ComfyUI-CustomNodeKit. ComfyUI Manager: search "CustomNodeKit". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/user2318/ComfyUI-CustomNodeKit.git
cd ComfyUI-CustomNodeKit
pip install -r requirements.txt

Make sure groundingdino-py and transformers are installed (the pack's install.py attempts it automatically). The GroundingDINO weights auto-download on first use via GD_ModelLoader - about 700MB for SwinT.

The gotcha

Watch the preview before you commit. The three silent failure modes are all visible in preview_images: a threshold too high gives you no boxes at all (output degrades to empty), a prompt that's too clever returns boxes for the wrong thing, and a multi-person clip with by_area + largest will happily track the wrong person when the main character is small in frame. When a pose animation looks like it's following the background dancer, this node - not SDPose - is where the mistake lives. Check the preview once, and track mode will save you the repeat.