KIE Kling 2.5 I2V Pro

Kling 2.5 I2V Pro is the pack's "control the ending" video node. Most image-to-video gives you a start frame and hopes. This one takes a required first frame, an optional last frame, and generates the clip between them - so when you need the shot to land on a specific final pose, you can say so instead of rerolling until the model happens to agree. It's Kling 2.5 Turbo's I2V Pro model, served through Kie.ai.

The last-frame input is the reason to reach for this over the pack's plain Kling 2.6 I2V node. If your pipeline needs a beat-accurate ending (a loop, a title card reveal, a character landing a pose), that single input is worth the extra setup. If you don't need the tail frame, the simpler node does the job with fewer inputs.

How it works

Standard KIE video job lifecycle with the usual resilience: upload the frames, POST a createTask with the Kling 2.5 I2V Pro model id, poll recordInfo every 10 seconds up to the 2000-second video timeout, retrying automatically on transient failures. The result downloads as a ComfyUI VIDEO.

The inputs that matter

• first_frame - required. The opening image, and the one that defines your subject. Feed it a clean, well-framed still.
• prompt - required field, though it can be empty; it's your motion description. Give it the action: "the subject stands and walks toward camera."
• last_frame - optional. The ending image. Both frame inputs take a batch, but the first image in each is what's used. This is the differentiator - connect it when the ending matters.
• negative_prompt - optional, and a nice luxury to have on a video node. Use it to block artifacts ("morphing, extra limbs, warping") that plague short clips.
• duration - 5 or 10 seconds (default 5).
• cfg_scale - 0.0–1.0, default 0.5. This is the prompt-adherence dial. Higher = tighter to the prompt; the sweet spot for Kling tends to be around 0.5. If motion comes out ignoring your description, nudge it up.

Output: video, ready for preview or a saver.

Installing

Installs as part of the whole pack. ComfyUI Manager: search "ComfyUI Kie API". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/gateway/ComfyUI-Kie-API
# restart ComfyUI

No model files. API key at config/kie_key.txt in the pack folder (copy from kie_key.example.txt).

Where people get burned

The frame-batch behavior trips people up: both inputs take batches, but only the first frame of each is used - wire a whole batch in and the extras are silently ignored. The other classic is expecting the last frame to be literal; it's a strong prior, not a promise, so leave slack in your ending. Long runs are normal - the 2000-second timeout exists for a reason, and killing a job at two minutes wastes the credits you already spent. If something fails, the console state log and kie.ai/logs will tell you why, and the usual first suspect is a drained balance.