H3 Fun ControlNet — pose transfer workflow

Files behind Pose transfer works: the woman in the dress follows the briefcase poses.

Read this before you load them

These are API-format graphs, not UI workflows. ComfyUI's drag-and-drop expects the UI format; load these with Workflow → Open (which accepts API JSON) or POST them to /prompt directly. They are what our pipeline submits, so they are exactly what produced the clips in the write-up — but they will not render as a laid-out node graph.

They will not run unmodified on your machine. They name our specific checkpoints, several of which are quantized builds rather than stock releases:

• minimax_h3_ref2va_pruned_int8_convrot.safetensors (UNET)
• qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIP)
• minimax_h3_video_vae_fp16 / minimax_h3_audio_vae_fp32 (VAEs)
• minimax_h3_turbo_v4_step600_ema.safetensors (Turbo LoRA)
• minimax_h3_fun_controlnet_union_pruned_bf16.safetensors (ControlNet)

Swap those filenames for whatever your install has. The ControlNet is the one that is not interchangeable: the loader takes only Kijai's curve-form pruned variant. Check the safetensors header — adaln_proj.linear of [96768, 8] with adaln_basis metadata is the right one; [96768, 2688] is Alibaba's full-width release and gets refused.

Requires wyzborrero/ComfyUI-H3-FunControl for the H3FunControlLoader and H3FunControlApply nodes.

The control video

H3FunControlApply takes an IMAGE batch. LoadVideo alone returns VIDEO, so the graph goes LoadVideo → GetVideoComponents → control_video.

The control must match the generation's width, height and frame count exactly — the node raises when the control token count diverges from the video segment, and reports both numbers.

If you already run DWPose or another preprocessor, use it. make_pose_control.py exists because our ComfyUI had no pose preprocessor installed; it draws an OpenPose-style coloured skeleton with MediaPipe, offline:

pip install mediapipe opencv-python
curl -sLO https://storage.googleapis.com/mediapipe-models/pose_landmarker/pose_landmarker_heavy/float16/1/pose_landmarker_heavy.task
mkdir -p models && mv pose_landmarker_heavy.task models/

python make_pose_control.py source.mp4 pose_control.mp4 \
  --width 1312 --height 736 --frames 294 --thickness 8 --autocrop

Then drop pose_control.mp4 in ComfyUI/input/.

--autocrop crops the source to the figure before extracting. That matters more than it sounds: raising resolution alone does not raise how much of the frame the subject occupies, and subject size — not strength — is what governs how much authority the control has. At 0.4 MP and 0.98 MP our skeleton occupied the same ~7% of frame and the prompt won outright; cropping took it to 10.6% and the pose took over.

Settings that produced the published clips

1312×736, 294 frames (12.25 s at 24 fps), 8 steps, simple scheduler, strength 0.8, control window 0.0–0.6. About 18 minutes per clip and 20 GiB of VRAM on a 3090.

The prompt deliberately describes only who and where — it never mentions kneeling, a briefcase or a tripod. That is the point of the test: any action in the output has to come from the control. If you write the action into the prompt as well, you can no longer tell which signal produced it.