Kling Image to Video

The classic image-to-video move: feed Kling a still and a short prompt, get back a 5–15 second clip where the image comes alive. This is the node most people mean when they say "I want to animate this picture," and it's where the whole API-toolkit deal finally pays off - no local video model eating your VRAM, no ComfyUI video-pipeline debugging, just an image, a prompt, and a clip on the other side. If you've done any of this locally you'll recognize the shape instantly: it's the hosted equivalent of Wan or LTX image-to-video, except the model is kling-v3 and the compute is Kling's.

Mechanically: the node uploads your image (base64), fires off a generation task, and polls until the clip is done. Because every node here implements IS_CHANGED, re-queueing always re-runs. The interesting design bit is the prompt field - like the pack's Text to Video, it supports @image1 references, so you can point at your input image explicitly ("@image1 walks toward the camera") instead of hoping the model infers the subject.

Inputs worth your attention:

• image (required) - your starting still. There's also an optional image_tail for a second reference image (useful for a defined end state).
• model_name - kling-v3 (default), kling-v2-6, kling-v2-master, or kling-v1-6. Newer isn't always better for your exact subject; if v3 over-animates, try v2-master.
• duration - 5, 10, or 15 seconds. Longer = more credits = more drift risk.
• mode - pro (quality) vs std (faster/cheaper). Start pro; std is for drafts.
• sound - toggle AI-generated audio (Kling v3 generates ambient sound + effects natively). Defaults on, and it's kind of the selling point of this generation.
• cfg_scale - 0–1, prompt adherence vs creativity. 0.5 default.
• camera_control - optional KLING_CAMERA from the Kling Camera Control node. This is how you force a dolly or zoom instead of praying.

Outputs: video (IMAGE tensor frames), video_file (path to the saved clip), audio (AUDIO), url (hosted), task_id.

Installing it

Part of ComfyUI-API-Toolkit. Manager: search "API Toolkit". Manual:

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

Restart ComfyUI, wire in the Kling AI Authentication node, and fund your account.

Gotchas

Cost is the headline. Each clip is a billed generation and video seconds are the priciest thing on Kling's menu, so 15s pro clips with sound will drain credits shockingly fast - draft in std/5s. Two more: the video tensor output holds all frames in RAM, so for long clips prefer the video_file/url outputs (or the Fast Video Saver) or you'll OOM on a modest card. And Kling's motion quality is genuinely good but not infinite - limbs and faces still warp, especially at 15s; the KB's video-ecosystem notes that even the top hosted models get compared against local Wan in the arena, so keep expectations sane.