KlingAI 2.5 Turbo Standard

Kling is Kuaishou's paid video model, the one people run on web dashboards and API platforms rather than their own GPU - and this node is how you get it on a ComfyUI canvas without moving the model to your machine. It's a cloud inference node: you type a prompt (or drop in a reference image), Runware's servers do the denoising, and the finished clip lands back as a native VIDEO on your canvas. No checkpoint download, no VRAM calculation, just a billed generation.

Kling 2.5 Turbo is the fast tier, and the community has strong opinions about it: in the Kling-vs-Wan 2.2 comparison threads, the recurring line is that 2.5 Turbo is "light years better than Kling 2.1" while still being a closed, paid model. If you're coming from the open Wan stack and want a quick take on what the paid tier gets you, this is a cheap experiment.

What you actually set

positivePrompt is the only required input - keep it concrete, same habits as local video gen. The settings that matter:

• frameImages - wire in an image (or images) for image-to-video. It's an IMAGE socket, so any ComfyUI image node feeds it directly.
• duration - 5 or 10 seconds, chosen by the model's frame rate. Longer isn't automatically better; 10s clips are where motion starts to drift on most models.
• CFGScale - defaults to 0.5, which is unusual. Kling Turbo is guidance-distilled, so don't crank this like an SD checkpoint; the tooltip warns that a negativePrompt is ignored whenever CFGScale is at or below 1.
• numberResults - how many clips to generate, each with a different seed. This is also how you multiply your bill, so start at 1.

The rest are output plumbing (outputFormat MP4/WEBM/MOV, outputQuality), safety checking (safety.checkContent with fast/full modes), and the ttl toggle that controls how long a URL-hosted result stays alive.

The output is a single video socket. In current ComfyUI it carries the native VIDEO type, so wire it into a video Preview or Save node. Older setups may fall back to a file path.

Installing

This node ships inside the Runware/ComfyUI-Runware pack, which exposes the whole Runware cloud catalog as nodes. Install it once and Kling (plus every other model node) shows up together.

ComfyUI Manager: open the Custom Nodes Manager, search Runware, install, restart.

Manual:

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

There are no model files - the Python deps are just runware-sdk, pillow, and soundfile. You do need a Runware API key from runware.ai/api-keys, pasted into ComfyUI Settings → Runware API key or set as RUNWARE_API_KEY.

Where people get burned

It's a paid API, so the obvious trap is billing: each numberResults is a separate generation, and the node's title bar shows your per-run cost after the fact - worth actually reading before you queue a 10-clip batch. The subtler trap is treating it like local tooling: there's no seed control exposed here and no local cache, so results aren't reproducible across runs. For reproducibility you'd want a seed parameter, and this node deliberately doesn't expose one. Use it for what it's good at - sampling the closed frontier model without renting an 80GB GPU.