ComfyUI MiniMax H3 Fun ControlNet Union: Pose, Depth, and Canny Control Video

Control MiniMax H3 video motion in ComfyUI with Canny, Depth, HED, MLSD, or Pose control videos, or run mask-based video inpainting with the Fun ControlNet Union patch.

Control the motion of a MiniMax H3 video with a control video instead of text alone. The Fun ControlNet Union patch (alibaba-pai/MiniMax-H3-Fun-Controlnet-Union, packaged in Comfy-Org/MiniMax-H3) is a single checkpoint that conditions H3 on Canny, Depth, HED, MLSD, or Pose control videos, and also runs video inpainting with a mask. It works with both the fl2va and ref2va transformer files.

Run on Comfy Cloud

Open in Comfy Cloud

Download Workflow

Download JSON or search “MiniMax H3 Fun ControlNet Union” in Template Library

Input materials

Upload this file to the matching Load Video node:

dancer_field_pose.mp4

Pose control video used by the example workflow, or use your own.

This template requires ComfyUI 0.35.0 or later. The example uses a pose control video: a built-in SDPose subgraph extracts the pose skeleton from the input video, so the output reproduces the dancer’s motion. You can also connect any preprocessed control video directly to the Apply MiniMax H3 Fun ControlNet node’s control_video input.

Model downloads

Diffusion Model: minimax_h3_ref2va_pruned_int8_convrot

Place in ComfyUI/models/diffusion_models/

ControlNet Patch: minimax_h3_fun_controlnet_union_pruned_int8_convrot

Place in ComfyUI/models/model_patches/

Text Encoder: qwen3vl_32b_minimax_h3_nvfp4_awq

Place in ComfyUI/models/text_encoders/

VAE: minimax_h3_video_vae_fp16

Place in ComfyUI/models/vae/

VAE: minimax_h3_audio_vae_fp32

Place in ComfyUI/models/vae/

LoRA: minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16

Powers the optional Lightning LoRA turbo mode. Place in ComfyUI/models/loras/

Checkpoint: sdpose_wholebody_fp16

Pose extractor used by the built-in SDPose subgraph. Place in ComfyUI/models/checkpoints/

Detector: rt_detr_v4-x-hgnet_fp16

Person detector used by the built-in SDPose subgraph. Place in ComfyUI/models/diffusion_models/

Model storage

ComfyUI/
├── 📂 models/
│   ├── 📂 diffusion_models/
│   │   ├── minimax_h3_ref2va_pruned_int8_convrot.safetensors
│   │   └── rt_detr_v4-x-hgnet_fp16.safetensors
│   ├── 📂 text_encoders/
│   │   └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
│   ├── 📂 vae/
│   │   ├── minimax_h3_video_vae_fp16.safetensors
│   │   └── minimax_h3_audio_vae_fp32.safetensors
│   ├── 📂 loras/
│   │   └── minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors
│   ├── 📂 model_patches/
│   │   └── minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors
│   └── 📂 checkpoints/
│       └── sdpose_wholebody_fp16.safetensors

Prompting tips

1. Video length: The target video runs 5 to 15 seconds on the 17n+5 frame grid at 24fps (124 frames = 5 seconds). A control video longer than the target is trimmed to its first frames; a shorter one holds its last frame. To match the control video’s length exactly, link the example’s computed batch_size output to the length input of the MiniMax H3 Reference to Video node
2. Guidance and strength: Keep guidance_scale at 1.0. Raise the patch strength above 1 only if the output drifts away from the control
3. Video inpainting: Connect a mask to the mask input: regions marked 1 are regenerated under control of the optional source_video input, while the rest of the frame stays fixed
4. Turbo mode (optional): Enable the Lightning LoRA checkbox to use the 4-step turbo LoRA for much faster generation, with slightly lower audio and motion quality
5. Preparing control videos: ComfyUI ships with built-in preprocessing nodes for each control type. See the Detect Edges (Canny) node for edges and the Depth Anything 3 tutorial for depth maps