Kling VIDEO 2.6 Pro

This is the Kling node with no required inputs, and that's the point. Kling VIDEO 2.6 Pro takes every kind of reference the model supports - referenceImages, frameImages, even referenceVideos - plus a prompt, plus native audio generation, and decides what to do with what you give it. Text-to-video works with just a prompt; image-to-video works the moment you feed a frame; character work happens when you add a reference clip. It's the "one node does all of it" slot in the modern Kling lineup.

Why you'd reach for it

Because the new Kling isn't one workflow, it's several, and you shouldn't have to swap nodes for each. 2.6 Pro is the current-generation Pro tier - this is where the quality and the features live. The provider settings are what make it interesting: you can generate native sound in the clip, keep the original audio from a reference video, and tell the model whether its character-orientation reference should come from your image or your video. That's real capability, not widget dressing.

How it works

Everything optional rides to the Runware REST API via the runware SDK: referenceImages/frameImages tensors are base64'd to PNG data URIs, while referenceVideos is a typed input (URL/path/UUID references). The rendered clip returns as a native ComfyUI VIDEO object with cost and NSFW flag on the title bar.

Inputs that matter

• positivePrompt - technically optional, practically your starting point. Describe the scene; 2.6 will fill in what you leave out.
• referenceImages (IMAGE) / frameImages (IMAGE) - first-frame and style-reference sources for image-driven work.
• referenceVideos - point at a source clip to steer character and motion.
• providerSettings.klingai.characterOrientation - image or video: which reference defines the character's look and facing.
• providerSettings.klingai.sound - enable native audio generation in the output.
• providerSettings.klingai.keepOriginalSound - keep the source clip's audio when using a reference video.
• duration - 5 or 10 seconds. width/height default 1024; CFGScale 0.5 default.
• numberResults (1–4), outputFormat (MP4/WEBM/MOV), outputQuality (95), plus the safety group.

Output: one video socket (VIDEO type).

Install and the API key

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

Restart ComfyUI (or Runware via ComfyUI Manager). Key from the Runware dashboard → Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

• No required inputs is a trap disguised as freedom: an empty node will happily generate from just a prompt, which may not be what you meant. Feed it a reference or write a real prompt.
• The media sockets aren't all the same kind - images are tensors, videos are URL/path/UUID references. Wrong kind = confusing error.
• Native sound costs more and isn't always right; try sound on for the shots that need it and off for the ones that don't.
• Closed Kuaishou model, per-run pricing, cost on the node.

If you only add one Kling video node to your graph, 2.6 Pro is a strong candidate - current-gen quality plus every input shape the modern model offers, all in one socket-happy box.