Kling Omni Video-to-Video Reference (fal)

This is the most powerful - and most fiddly - of the Kling nodes in this pack. Instead of generating from a still, it takes an existing video plus a prompt and reference imagery, and re-generates the clip guided by both. The headline feature is the "reference element" system: you can supply up to four distinct elements (a character, an object, whatever) each with a frontal image and extra reference images, and have Kling carry them into the output video. Think character replacement, consistent subjects across a restyle, or dropping specific people into an existing motion.

What it actually does

Nothing runs locally. Your driving video, prompt, and reference images go to fal.ai, fal runs Kling's Omni video-to-video endpoint, and you get a new clip back - no VRAM, no downloads, billed per generation. Because you're uploading a whole video and multiple images, and video jobs are pricey, this is a deliberate, per-shot tool rather than something to spam.

The inputs that matter

• video (required): the source clip that provides the motion and structure.
• prompt (required): describe what you want the output to look like - the restyle, the scene, the action.
• reference_images (optional): general reference imagery for the look and subjects.
• element_1…4 frontal_image / reference_images (optional): the structured part. Each "element" is a subject you want the model to hold consistent - give it a clean frontal image plus supporting references. The pack's typed builder nodes (added in 2.5) are meant to make assembling these configs less painful than hand-typing JSON.

Rounding it out: keep_audio (carry the source video's audio through, off by default), aspect_ratio (auto/16:9/9:16/1:1), duration (5 or 10), and variations (1–10, each billed). The output is a STRING list - video URLs, named video_url. Feed it into the pack's Fal Save Media from URL node to download the file.

Installing it

Easiest: 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 file needed; the dependencies are fal-client and a fal API key (via config.ini under [API], or the FAL_KEY env var). Note that a VIDEO input and output require a reasonably recent ComfyUI - if the video socket is missing or reads None, update ComfyUI first.

Where people get burned

Two node-specific traps. First, the elements only help if the frontal images are clean and unambiguous - a cluttered or low-res frontal gives the model a weak anchor and consistency suffers. Feed it good references. Second, this is the most expensive shape of call in the pack: a full video plus multiple images plus variations can rack up cost quickly, so keep variations at 1 while you're finding the right prompt and element set.

The rest are the usual API realities: content-policy refusals on people come through as readable fal errors (read them), and on ComfyUI Windows Portable an import error is fixed with .\python_embeded\python.exe -m pip install fal-client.