Kling 3.0 Turbo Text to Video

Text to video is the most impressive demo and the most expensive way to work, and both of those things are true here. But it earns its place for one job: figuring out what you even want. You've got a shot in your head and no image of it yet - this node is the fastest route from a sentence to "oh, that's the angle I meant."

It's the sibling of the pack's Kling 3.0 Turbo Image to Video node, same unified /tasks API, same 3–15 second range, different starting point.

Inputs

auth from the Kling AI Authentication node, then prompt - the whole creative brief in one field.

resolution is 720p or 1080p. aspect_ratio gives you 16:9, 9:16 or 1:1; the vertical option is the one people forget exists and then shoot themselves in the foot with a 16:9 clip they wanted for a phone screen. duration is an integer, 3 to 15 seconds.

That's the full set. There's no negative prompt, no seed, no motion strength dial. What you get is a prompt and four constraints.

Outputs

video is the decoded IMAGE batch and video_file is the filename saved into ComfyUI's output directory - a name, not a path to hand to a loader node. audio is an AUDIO stream extracted from the clip with torchaudio, best-effort by design: no audio track, or no torchaudio in your environment, and you get a silent placeholder with a warning in the console instead of an exception. url is Kling's temporary hosted link, and task_id is the job id.

How it works

The node builds a request for Kling's unified /tasks API with your prompt and a settings object carrying resolution, aspect ratio and duration. It submits, then polls /tasks with adaptive intervals - frequent while the job is young, backing off to eight seconds - printing status lines to the console as it goes, up to a 1200-second ceiling. On success it scrapes the video URL out of the task result, downloads it into the output folder, and decodes it into frames plus audio. A failed task throws with Kling's own failure message attached, which is where you want to look first.

Installing

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

Just requests and PyJWT for the Kling service, no model downloads. Manager users: search "API Toolkit". The moment you queue this, remember that ComfyUI custom nodes are unsandboxed Python that reaches the network holding your key - that's true of every API pack, and it's the reason to stick with widely-installed ones and read what you're installing.

Getting value out of it

Set up authentication first. Both Turbo nodes require the newer single API key on the Kling AI Authentication node (KLING_API_KEY works as an environment fallback). Access key plus secret key is not enough here, and if the key is missing the node raises before making a call.

Then stop treating it as your main generator. The pattern that actually saves money: use T2V once or twice to find a composition you like, then take the still you'd have generated locally - your own checkpoint, your own LoRA, free, unlimited retries - and animate that with the Image to Video node. You get first-frame control and you stop paying to re-roll aesthetics.

Why not just do video locally? Because you can't, for this model. Kling's weights aren't distributed - closed video services are the one category where there is genuinely no local door, which is exactly why the community tolerates API nodes here and resents them everywhere else. Open video exists and it's real (Wan, Hunyuan, LTX and the recent open releases have closed a lot of ground), and if you're rendering a hundred clips a week the local route amortises. But nobody local ships Kling.

Long generations look like hangs. They aren't - watch the console output, not the canvas, and let the poll finish. And keep tests at 3 seconds and 720p until the prompt is right.