Kling v2.1 Pro Image-to-Video (fal)

Kling is Kuaishou's closed video model and one of the genuinely good hosted image-to-video options - smooth, physically believable motion is its whole reputation. The Pro tier is the sensible middle: better than the cheap endpoints, not as pricey as Master, and the one most people should actually be reaching for day to day. This node runs Kling v2.1 Pro from ComfyUI - hand it a still and a prompt, fal animates it on its servers, you pay per clip.

There's no weight to download here; Kling is API-only. So a bridge node is the only way to get it into a graph, and this is it.

What it actually does

Your image and prompt go to fal.ai, fal runs Kling v2.1 Pro, and you get a clip back - no VRAM, no downloads, billed per generation. The nice thing about the v2.1 Pro node over the bare-bones tiers is the extra control it exposes: a real cfg_scale for prompt adherence, and a tail_image so you can pin the last frame.

The inputs that matter

• image (required): your first frame - the clip animates outward from it.
• prompt (required): describe the motion and camera in plain language.
• duration: 5 or 10 seconds. Ten roughly doubles the cost.
• cfg_scale (0–1, default 0.5): how hard Kling sticks to your prompt versus letting the image lead. Nudge up if it's ignoring your direction, down if the motion's getting stiff or artifacty.
• tail_image (optional): a target last frame. Supply it to steer the clip toward a specific ending; leave it out for free animation.

There's a negative_prompt (pre-filled with "blur, distort, and low quality"). The output is a single STRING - the video URL, not a playable clip in ComfyUI. Route it into the pack's Fal Save Media from URL node to download the file into output/, or into Load Video from URL to pull the frames into your graph.

Installing it

ComfyUI Manager → search ComfyUI-fal-API → install → restart. Manual:

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

Restart ComfyUI. No model in models/; you need fal-client and a fal API key - config.ini under [API], or the FAL_KEY env var. No key, no video.

Where people get burned

The cfg_scale range trips people who expect the 1–15 CFG they know from image sampling - here it's 0–1, and 0.5 is a good neutral. Crank it near 1 only when the model's genuinely ignoring your prompt, because high values can make motion look forced. The tail_image works best when it's tonally close to the start frame; a wildly different end forces an ugly transition.

Otherwise the usual API notes: no FAL_KEY means every run fails, and a content-policy refusal (common with real people and likenesses) comes through as readable error text since the pack passes fal's real message along. Iterate cheap - single 5-second runs while you dial the prompt - and lean on the cache, which serves identical calls for free.