Kling V3

Kling V3 is the newer Kling engine, and unlike the plain CivitaiVideoGenKling node - which is hard-wired to text/image generation - this node folds five operations into one: text-to-video, image-to-video, reference-to-video, video-to-video edit, and video-to-video reference. You pick the operation from a dropdown and the node's inputs light up accordingly. It also adds audio controls, which the older Kling node doesn't have.

This is the node to reach for when you want the current Kling generation without juggling a separate node per operation. Same deal as every node in civitai/civitai-comfy-nodes: the job runs on Civitai's cloud, billed in Buzz, no GPU or model download on your side.

How it works

The node is a thin discriminator variant - it submits a videoGen workflow with engine kling-v3 plus the operation you chose to the Civitai Orchestration API, long-polls, and returns the finished MP4 as a native VIDEO. Which inputs actually matter depends on the operation: image-based ops want source_image/images, video ops want video_url, and reference ops want images plus your elements_json or prompts. Because it's one node with a broad input surface, the main skill is not confusing yourself about which inputs the current operation consumes.

Inputs and outputs that matter

• prompt (required) - the clip or the edit instruction.
• operation - text-to-video, image-to-video, reference-to-video, video-to-video-edit, video-to-video-reference. The default is empty, so pick one deliberately.
• mode - standard / professional (cost difference).
• duration - 3–15 seconds, default 5.
• aspect_ratio - 16:9, 9:16, 1:1.
• source_image - start frame for image-to-video (IMAGE socket; accepts URL/DataURL/Base64).
• end_image - end frame for a first-to-last-frame style move.
• video_url - the source clip for video-to-video operations.
• images - reference set for reference-to-video.
• elements_json - structured element/character definitions for the newer Kling editing workflows.
• generate_audio - have Kling V3 synthesize an audio track (default off).
• voice_ids_json - voice selection when audio is generated.
• keep_audio - preserve the source clip's audio in video-to-video (default true).
• multi_prompt_json - multi-shot prompt script.
• api_config - auth override.

Outputs: video, workflow_id, raw_json.

Installing it

Shared pack install: 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

Only requests; no model downloads.

Common issues

• Auth & Buzz. Civitai account with Buzz needed; CIVITAI_API_TOKEN on headless, Auth node or sidebar otherwise.
• Operation/default trap. The operation dropdown defaults empty. Leaving it blank can submit an incomplete job - set it explicitly for the mode you want.
• Wrong input for the operation. Feed video_url to text-to-video and the job may ignore it. Match the inputs to the operation; when in doubt, the KB's rule of thumb applies - text ops want prompt only, image ops want source_image, video ops want video_url/images.
• Audio doubles the job. generate_audio on is a heavier, costlier run. Use keep_audio on edits where the source track is fine.
• Early preview. Behavior can change without notice; pin versions for production.