KIE Kling 3.0 Motion-Control (I2V)

This is the motion-transfer node of the pack, and it's a genuinely different flavor of image-to-video. Instead of "here's an image, make it move somehow," you hand it two things: a reference image (the subject) and a motion-driver video (the movement). Kling 3.0 then produces a new clip where your image's subject performs the driver video's motion. Identity comes from the still, choreography comes from the video.

It's the kind of thing you reach for when you have a specific camera move or a specific action you want, not just "a cinematic shot." Want a portrait that turns its head exactly like a clip you shot earlier? Want a product shot that follows the pan from a reference video? That's this node's lane. If you just need a video from an image, the plain Kling 2.6 I2V node is simpler and cheaper.

How it works

Same transport as the rest of the pack - upload both media files to the Kie API, submit the task, poll, download the result mp4 - but the payload has an extra control you won't find on the standard I2V nodes: character_orientation. It tells Kling which source defines the character's facing direction:

• video (default) - orientation follows the motion-driver video. KIE documents this as supporting up to 30s of source video.
• image - orientation matches the reference image. KIE documents this as supporting up to 10s of source video.

If you've ever watched an I2V result where the subject's facing flips arbitrarily, you'll understand why this toggle exists. It's a real quality lever, not decoration.

Inputs

• images (required) - your source image batch; only the first image is used. It must be larger than 300px in both dimensions and its aspect ratio must sit between 2:5 and 5:2.
• video (required) - the motion-driver clip, and it must be under 100MB. Nothing exotic about that; it's an upload limit.
• prompt - optional, up to 2500 characters. Use it to steer the motion ("turns head slowly, looks at camera") rather than describe the scene.
• character_orientation - image or video, as above.
• mode - output resolution: 720p or 1080p.

Output: a single video (VIDEO type) feeding straight into SaveVideo.

Install and first run

Standard pack install - ComfyUI Manager (search "ComfyUI-Kie-API") or:

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

Restart, then put your Kie.ai API key in config/kie_key.txt (copy it from config/kie_key.example.txt). No model files, no GPU requirements - it's all Kie's servers doing the work.

Where it bites

• Validation failures are quiet and early. The 300px minimum, the 2:5–5:2 aspect ratio, and the 100MB video cap are checked before submission. If your run dies instantly with a validation error, measure the image and shrink the driver video - that's the usual culprit.
• It's a Kling 3.0 credit burn. Motion control isn't the cheap lane. The pack's own docs recommend checking remaining credits before long sessions.
• .jpg in the docs. The KIE callback example for this endpoint shows a .jpg result URL, which the node author flags as a docs inconsistency - the node treats the endpoint as video and downloads the first result as a VIDEO. If you ever see something odd in the output, that's a known quirk, not your wiring.
• Long jobs. Same 2000s async timeout as the rest of the pack. If something looks stuck, https://kie.ai/logs shows you the actual job state before you decide it failed.

The motion-control node is the most "directed" video generation this pack offers - when it works, it feels less like gambling with a video model and more like giving directions.