Comfyui-Kling-Wrapper Image2Video

This is the node most people install this pack for. Image2Video is the middle of Kling's lineup and the most flexible: one still image in, a motion-filled clip out. Community consensus has treated Kling's image-to-video as the closed-source quality bar for years - even the local Wan crowd benchmarked against "Kling 2.0 master" - and this node is how you reach that bar from inside ComfyUI.

It's an API node, not a local generation node. Nothing renders on your GPU. You send the image to Kling, it renders in their cloud, and you get back a URL. That means no VRAM math, but also no free rides - every run is metered against your Kling API credits.

What you feed it

The image input is a plain ComfyUI IMAGE connection, so anything upstream that produces an image works: a KSampler, an image loader, or this pack's own Image Generator. There's also an image_tail input if you want to specify the ending frame, which is a neat trick for controlled loops - the video starts from image and resolves to image_tail.

The model dropdown is the big decision, and it's well stocked: kling-v1 through kling-v2-6, plus kling-v3, kling-v3-omni, and kling-v3-turbo. Start with kling-v3 for quality; grab kling-v3-turbo when you want cheaper, faster iterations. mode goes std/pro/4k, but 4k only works on models that support it - the node rejects the combo locally if the model can't do it, so you won't waste a submission.

The controls worth knowing

• cfg_scale - defaults to 0.5, range 0–1. This tunes how literally the model follows your prompt. Kling's default 0.5 is a fine starting point; nudge up if the prompt keeps getting ignored.
• camera_control_type + camera_control_config + camera_control_value - the trio for camera moves. camera_control_type picks the move family (simple, down_back, forward_up, right/left_turn_forward), camera_control_config picks pan/tilt/roll/zoom etc., and the value sets how strong the move is. Leave type at None if you'd rather the model find its own camera.
• shot_type - single or intelligence. intelligence lets the model break your prompt into a multi-shot sequence, which is where Kling's cinematic feel comes from.

Sound and voice

This node is also the pack's voice showcase. sound on/off, a voice_preset dropdown with dozens of preset voices (built for kling-v2-6's native audio), and the custom_voice_id / custom_voice_list inputs for cloned voices. The README's recipe for a cloned voice: select kling-v2-6, set mode=pro, sound=on, a 5- or 10-second duration, and reference the voice in the prompt as <<<voice_1>>>. voice_preset, custom_voice_id, and custom_voice_list are mutually exclusive, so pick one path.

You can also attach an element_list here (from the Element List Builder) for subject consistency, and a reference_video URL.

Outputs and install

Outputs are url and video_id - wire url into the pack's Preview Video node to download and view. Install via ComfyUI Manager (search "Comfyui-Kling-Wrapper") or git clone https://github.com/magicwang1111/Comfyui-Kling-Wrapper.git into custom_nodes, pip install -r requirements.txt, restart, and put your Kling API key in config.local.json. No model files, no VRAM - just credits.

Where people get burned: expecting 4k from a model that doesn't support it (the node catches this), and expecting custom voices to work on kling-v3 (the API rejects them; v2.6 is the reliable voice model). Match your model to what you're actually asking for and this node is remarkably painless.