IAMCCS Cine H3 Fun Control Input · Pose / Depth / Edge

MiniMax H3 can take structural guidance - pose, depth, edges - the way Stable Diffusion takes ControlNet maps. In the IAMCCS shotboard architecture, "H3 Fun Control media" is the name for those frames, and this node is the door they walk through. It has two modes that look identical but are very different jobs:

1. You already preprocessed the frames (pose skeletons, depth maps, canny lines) elsewhere and just want to hand them over.
2. You hand it raw source_video and let it run the preprocessor for you.

Either way, the frames get parked inside the cine_linx bus under iamccs_minimax_h3_control_video so the H3 backend and settings nodes can find them downstream. The author's contract is explicit: this node owns transport and preprocessing; the H3 Settings node owns kind, model, strength and window. Don't go looking for a strength slider here - it isn't this node's problem.

How it works

Feed it cine_linx plus either preprocessed control_video or raw source_video. If you give it raw frames and pick a real preprocessor, it calls into your installed ControlNet Aux implementation (AIO_Preprocessor) at preprocess_resolution. Options: dwpose, depth_anything, canny, hed, mlsd - plus already_preprocessed (default, hand the frames in done) and from_iamccs_settings, which reads which control kind your H3 Settings chose and picks the matching preprocessor automatically.

One nice engineering touch: when you pass in an already-processed control_video, the node keeps the tensor by reference rather than copying it. A long pose batch isn't duplicated just to travel through the bus - on multi-thousand-frame shots that saves real VRAM.

Inputs that matter

• cine_linx - required, the merged bus from CineH3Input / H3 Settings.
• source_fps - 24 default; recorded in the frame metadata so the backend can line control frames up with the timeline.
• preprocessor - already_preprocessed unless you want aux preprocessing.
• preprocess_resolution - 512 default, 64-step increments up to 2048.
• control_video - your finished control IMAGE batch.
• mask - optional inpaint mask if the control path uses one.
• source_video - raw frames to run through the chosen preprocessor.

Outputs

• cine_linx - the bus, with the control media attached. Carry it forward.
• report - a status string.
• control_preview - the exact frames that get sent to H3. This is the output to watch: if the pose looks wrong here, it's wrong downstream, and at least you can see it before burning a render.

Install

Same pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

(ComfyUI Manager: search IAMCCS.) For the live preprocess modes you additionally need comfyui_controlnet_aux installed and enabled - the node will raise a clear error telling you so if it's missing. already_preprocessed needs nothing extra.

Gotchas

• Pick a preprocessor with no comfyui_controlnet_aux installed and you get a hard runtime error, not a silent fallback. By design.
• The deliberate pass-through: leave all media unconnected and the node still returns a bus with a "pass-through" entry in the chain - IAMCCS keeps this insertion point wired even while ControlNet is off. But enabling ControlNet downstream while real frames never got attached hard-fails in the backend. It's a feature (no silent black renders), and it bites exactly once.
• This node is only meaningful inside the IAMCCS H3 shotboard graph. On its own it's a bus with a movie attached.