kling

Kling has no open weights. If you want it, you call it - and this node is how you call it from inside ComfyUI, through Civitai's official civitai/civitai-comfy-nodes pack. It's the textbook "model isn't on your machine" case from the KB's external-API-nodes doc: the whole job runs on Civitai's fleet, billed in Buzz, and the clip lands back in your graph as a native VIDEO.

Kling is Kuaishou's video model, and the reason people pay for it is physicality - it's one of the strongest commercial generators for motion realism, and it has been a benchmark for everything from camera control to video-to-video reference work for a while now. Because it's closed, you can't inspect the weights or fine-tune it; you get the hosted engine and its settings. This node covers text-to-video and image-to-video (via the source_image input), with a camera_control_json field for scripted camera moves.

How it works

Same orchestration plumbing as every node here: assemble a videoGen workflow with the Kling engine, submit to the Civitai Orchestration API, long-poll until succeeded, download the MP4, return video + workflow_id + raw_json. Your prompt (and source image, if wired) goes to Civitai's servers - that's the price of using a model you can never run locally. The api_config input is just an auth override; everything else is engine settings.

Inputs and outputs that matter

• prompt (required) - the clip.
• cfg_scale (required) - default 0.5 (range 0–1). Yes, 0.5 - Kling is not an SD-class CFG scale. This is a normalized guidance strength; don't crank it to 4 or 7 out of habit or you'll fight the model.
• model - engine version: v1, v1.5, v1.6, v2, or v2.5-turbo. Newer usually means better; turbo trades quality for speed/cost.
• negative_prompt - what to avoid.
• mode - standard or professional. Professional mode costs more; for most experiments standard is fine.
• aspect_ratio - 16:9, 9:16, 1:1.
• duration - 5 or 10 seconds.
• camera_control_json - JSON for scripted camera motion, the Kling signature feature. Check Civitai's docs for the accepted schema; malformed JSON just gets ignored or errors server-side.
• source_image_url / source_image - for image-to-video. source_image is the IMAGE socket (accepts a URL, DataURL, or Base64 string per the tooltip); source_image_url lets you pass a plain URL string.
• api_config - optional auth override.

Outputs: video, workflow_id, raw_json.

Installing it

ComfyUI Manager → Custom Nodes Manager → "Civitai Comfy Nodes" (publisher civitai) → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Just requests; nothing downloads.

Common issues

• Auth & Buzz. Civitai account with Buzz required. CIVITAI_API_TOKEN for headless; Civitai Auth node or sidebar connect panel otherwise.
• CFG confusion. 0.5 is the default and a normalized scale - treat it as guidance, not CFG. People paste SD habits in here and wonder why Kling ignores them.
• Professional mode bills more. mode=professional and duration=10 are the pricey combos. The node's status line shows per-run cost.
• Camera JSON is fiddly. Get the schema right before relying on it; the error shows up as a server rejection, not a local one.
• Early preview. The pack is under active development; APIs and behavior may change without notice.