Superside Kling 2.5 Turbo Pro Image-to-Video

This is the Kling 2.1 node's bigger sibling: Kling 2.5 Turbo Pro, which the README describes as the top-tier cinematic model with better motion fluidity and prompt precision than 2.1. Where the 2.1 node makes you pick a quality tier, this one is the top tier - one model, no model_tier dropdown, no decisions to make about quality. If you're after the best motion Kling will give you from a still, this is the node in this pack to reach for.

Same shape as its 2.1 sibling in every other way: closed model, hosted on fal.ai, pay-per-call, image leaves your machine. You feed it a first frame and a motion prompt, and it animates. This is squarely the "the model has no open weights" category of API node - there is no local Kling to install, so a wrapper like this is the way Kling enters a ComfyUI graph at all.

The inputs that matter

• prompt, image, api_key - the essentials. The prompt is a motion and camera brief, not a still-image description; Kling responds to phrases like "camera slowly orbits" and "cloth ripples in the wind."
• duration - 5 or 10 seconds. 10s showcases the motion fluidity but costs more per call.
• negative_prompt (default "blur, distort, and low quality") - leave it, it's a sensible default.
• cfg_scale (default 0.5) - how tightly the model follows the prompt. In Kling's world the useful range sits low, around 0.3–0.7; the default is a good starting point.

The output gotcha (yes, this one too)

Same as the 2.1 node: the output is a STRING - the URL of your finished video on fal's CDN - not a VIDEO tensor. You cannot plug this straight into SaveVideo. Grab the URL and download it, or route it through a URL-fetching node. The pack's own README lists these Kling outputs as "video URL (STRING)" and it's the first thing that trips people up, because it looks like it should behave like a video node and doesn't.

How it works

Standard Superside/fal plumbing: upload the first frame, POST to the Kling 2.5 Turbo Pro image-to-video endpoint with duration, negative prompt, and CFG, then return the result URL. All generation happens on fal's GPUs - nothing runs locally, which is the entire point of a closed-model wrapper. Video calls are the slow ones, so expect the node to sit there for a bit while fal does the heavy lifting.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Superside/comfyui-superside-nodes
pip install -r requirements.txt

Restart ComfyUI and it's under the Superside category. Paste a fal key into api_key - no config file, and a blank key falls back to a FAL_KEY environment variable if you have one.

Which Kling node, then?

If you're iterating on a prompt, the 2.1 node's standard tier is the cheaper sandbox. The moment you have a first frame and a prompt you believe in, this is where you spend the money for the good take. One practical note: because every run is a billable generation, nail the prompt on a 5-second run before committing to a 10-second keeper - the motion quality difference is in the model, not in the duration.