MiniMax H3 Attach Motion to Shot (Plan v2)

Motion transfer is the feature that made people describe H3 as "like Kling" - you feed it a clip of someone doing a dance or a fight, and it makes your character perform the same movement. The trap in hand-written H3 prompts is that a video reference tends to leak everything: the person's face, the location, the lighting, the composition, all of it. Attach Motion to Shot (Plan v2) exists to extract only the motion.

It does this with a clever trick that's documented right in the node's description. The clip keeps its native Video route (so H3 actually gets the pixels), but it's compiled as a reusable action Subject, not a plain video reference. The guide requires that distinction for motion transfer, and the pack's compiler handles the split automatically: the motion becomes a Subject that an upstream character performs, while identity, setting, and composition are explicitly not copied from the clip.

Inputs that matter

• shot_handle - from the exact Shot receiving the motion. This is a per-shot attachment, like Attach Keyframe.
• video_frames - a 2–15 second IMAGE frame batch (decoded video). It supplies motion/action only and gets prepared on H3's native frame grid.
• target_subject - the already-defined Subject that performs the referenced motion inside this Shot. If this field is empty or points at a non-existent Subject, you get a validation error - the target has to exist upstream.
• reference_name - human-readable name (default "shot motion").
• description - describe the motion/action evidence to transfer in this Shot. Be specific: "the leap and shoulder roll," not "movement."
• source_fps - the real FPS of your frame batch. Get this right or the motion timing drifts.

Outputs

• h3_plan - continue to another attachment or the next Shot.
• shot_handle - forwarded so you can attach a keyframe and motion to the same Shot.
• h3_video - the prepared native-grid motion-reference frames.
• attachment_preview - resolved Shot, reusable action Subject, target Subject, source Video, duration, and route. The preview is your friend here: it shows exactly what the compiler decided to do with your clip.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

Restart, browse MiniMax H3 → Plan v2/Shot Composition. No extra dependencies beyond the usual native-H3 ComfyUI build.

Gotchas

• Motion references are per-shot. If you want the same movement in two shots, attach it twice (or via the appropriate reference binding). A single attachment binds to exactly one shot.
• Source timing is capped. Reference videos must be 2–15 seconds, and cumulative video references cap at 15 seconds per project. A motion clip beyond that gets rejected, not truncated silently.
• Don't expect the clip's style to come along. That's the feature, not a bug. If you want the lighting or location from the source clip, declare those separately as Image/Video references. The whole point of this node is that a motion clip supplies motion only.

The pack's "Animate Keyframe with Motion Reference" example workflow is the canonical setup and it's the one marked Ready in the repo - load it, swap in your keyframe and motion clip, and you're running the intended pattern without assembling it from scratch.