KIE Kling 2.6 Motion-Control (I2V)

Ordinary image-to-video asks the model to invent the motion, and the model gets to make stuff up. Kling 2.6 Motion-Control flips that: you hand it a motion-driver video and a still image, and it transfers the video's movement onto your image's subject. Identity comes from your photo; choreography comes from the clip. It's the closest thing in this pack to a directed camera-movement tool, and it's the move if your local I2V keeps ignoring how you wanted the camera to behave.

This is the API-backed cousin of the workflow idea where you feed a reference video as a motion template - except there's no ControlNet, no keyframes, no local motion model to tune. Kling does the transfer on KIE's servers and bills you in credits.

How it works

Standard KIE job lifecycle with a video-shaped twist. The node uploads your source image and the driver video, then submits a createTask with both, plus the Kling 2.6 motion-control model id and your chosen settings. It polls recordInfo every 10 seconds; video jobs get the long 2000-second timeout because these runs take minutes, not seconds. The finished clip comes back as a ComfyUI VIDEO.

The inputs that matter

Only a handful:

• images - your reference still. First image in the batch is used. This is where the subject comes from, so make it a clean, centered shot.
• video - the motion driver. A single clip; this is the movement you're borrowing. It's a required input - no video, no node.
• prompt - optional-but-advised text describing what should happen ("the subject turns to camera and waves").
• character_orientation - video (default) or image. Default keeps the character's orientation matched to the driver video; switch to image if the subject's facing in the still should win.
• mode - 720p (default) or 1080p. Higher res costs more credits; iterate at 720p.

Output: video, ready for preview or a saver.

Installing

Part of the ComfyUI-Kie-API pack, installed once. ComfyUI Manager: search "ComfyUI Kie API". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/gateway/ComfyUI-Kie-API
# restart ComfyUI

No models to download. API key lives in config/kie_key.txt inside the pack folder (copy from kie_key.example.txt).

Where people get burned

The motion transfer only works if Kling can actually see motion in the driver clip - a static video as your driver just produces a confused still, and you've paid for it. Pick drivers with clear, unambiguous movement. Also: the pack accepts a VIDEO input, and where that video comes from matters - a clip saved by another ComfyUI video node works, but a raw path you typed into a string field won't wire in. Give the run the full timeout window; motion-control jobs are among the slowest in the pack, and killing them at two minutes is a guaranteed waste of the credits you already spent. Failures show up state-by-state in the console and in full at kie.ai/logs.