ComfyUI Node
H3MD Motion Control (pose / depth)
H3MD Loop motion control: pose skeleton (SDPose / DWPose) or depth (Depth Anything V2) + MiniMax H3 Fun ControlNet on the low-res stage. Connect to the Loop's depth_control input. Tick 'depth' on a video in the Director editor to use it as the motion source.
H3MD Motion Control (pose / depth)
- vae
- motion_control
◄control_modelminimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors►
◄strength0.70►
◄start_percent0.000►
◄end_percent0.600►
◄control_kindpose►
◄pose_enginesdpose►
◄pose_handstrue►
◄pose_facefalse►
◄pose_feetfalse►
◄depth_modeldepth_anything_v2_vitl.pth►
◄depth_resolution512►
◄sam3_enabledfalse►
◄sam3_threshold0.30►
◄sam3_fill20►
◄sam3_grow2►
◄sam3_blur2►
◄sam3_outputsilhouette►
CategoryMiniMax H3/H3MD
Inputs (18)
| Name | Type | Default | Description |
|---|---|---|---|
| control_model | COMBO | minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors | MiniMax H3 Fun ControlNet file in models/model_patches (int8). Loaded by the Loop only for segments with a depth video, released after their low-res stage. |
| vae | VAE | Video VAE used by core's Fun ControlNet to encode the control video (passed through as a reference, nothing is loaded here). | |
| strength | FLOAT | 0.700–10 | — |
| start_percent | FLOAT | 0.0000–1 | — |
| end_percent | FLOAT | 0.6000–1 | — |
| control_kind | COMBO | pose | pose = skeleton (body motion only, any character / outfit; Fun ControlNet 2.0 recommended). depth = Depth Anything V2, same as H3MD Depth Control. |
| pose_engine | COMBO | sdpose | pose only. sdpose = ComfyUI core SDPose wholebody (models/checkpoints/sdpose_wholebody_fp16.safetensors). dwpose = comfyui_controlnet_aux DWPose (yolox_l.onnx + dw-ll_ucoco_384_bs5.torchscript.pt; uses depth_resolution). |
| pose_hands | BOOLEAN | true | pose only: draw hands / fingers. |
| pose_face | BOOLEAN | false | pose only: draw face points (off: the prompt / reference decides the face). |
| pose_feet | BOOLEAN | false | pose only (sdpose): draw feet points. |
| depth_model | COMBO | depth_anything_v2_vitl.pth | Depth Anything V2 checkpoint (comfyui_controlnet_aux). Loaded per depth segment, unloaded right after. |
| depth_resolution | INT | 51264–2048 | Depth inference short side (aux keeps the aspect, no crop). |
| sam3_enabled | BOOLEAN | false | SAM3 mask on depth videos that have a SAM3 prompt (Director editor): only the prompted objects keep their depth, the rest is black. Empty prompt = full depth. The SAM3 prompt never enters the H3 prompt. |
| sam3_threshold | FLOAT | 0.300–1 | SAM3 detection score threshold (core SAM3_VideoTrack detection_threshold). |
| sam3_fillopt | INT | 200–256 | SAM3 mask: fill enclosed holes up to this size (px of the source video). Open gaps between legs / arm and body are never filled. |
| sam3_growopt | INT | 20–256 | SAM3 mask: grow outward (px of the source video). Keep small: a silhouette wider than the body leaves a translucent ghost (e.g. on the floor). Default 2. |
| sam3_bluropt | INT | 20–256 | SAM3 mask: gaussian feather radius (px of the source video). Default 2. |
| sam3_outputopt | COMBO | silhouette | With SAM3: silhouette = the cleaned mask as a white figure on black (author's workflow) feeds the ControlNet AND <Video N> - only shape and motion, no depth shading, so the reference video carries no picture at all. masked_depth = the depth inside the mask (grey figure). |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| motion_control | H3MD_DEPTH | — |