SDPose: Video Multi-Person Detection
Upload a video to detect human poses. Supports detection for both single individuals and multiple people within the same scene
About
This workflow performs multi-person human pose detection on video using the SDPose-OOD whole-body model. It ingests a video with LoadVideo, extracts frame size and FPS via GetVideoComponents, and feeds each frame into an SDPose subgraph (node ID 01b6a731-fb78-4070-9a38-c87146da9604). Inside the subgraph, an object detector proposes person bounding boxes per frame and the SDPose whole-body checkpoint (sdpose_wholebody_fp16.safetensors) estimates keypoints for body, hands, and face. The PrimitiveInt node controls max_detections, letting you cap how many people are processed per frame in crowded scenes.
What you provide
- Reading the workflow
Runs in ComfyUI. This workflow can't run on this page. Open it in Comfy Cloud with everything loaded, or download it for your own ComfyUI.
Try in CloudQuestions & answers
No. It performs per-frame multi-person detection and pose estimation. IDs are not persisted between frames. For tracking, you would add a tracker outside this subgraph.
Increase PrimitiveInt:max_detections, ensure the input resolution is sufficient, and consider unpacking the subgraph to raise detection confidence or adjust NMS. Note that higher resolution and more detections increase compute time.
Yes. Unpack the subgraph or tap the pose map output before ImageBlend and route it to CreateVideo/SaveVideo. You can also save per-frame images by adding a SaveImage node instead of video assembly.
LoadVideo and SaveVideo rely on your ComfyUI/FFmpeg setup. MP4/H.264 is commonly supported. If you encounter format issues, transcode the input to a standard MP4 and keep the default SaveVideo settings.