Kling 3.0 Text-to-Video

Kling 3.0 Text-to-Video is the top of the Kling line in this pack: text in, video out, powered by the current Kling 3.0 models through WaveSpeed. It's the node you reach for when a clip actually has to be good - the 3.0 family is a clear step up in motion coherence and prompt fidelity over the 2.6 tier, and you get three variants to pick from: standard, Pro (higher quality), and 4K (highest resolution). It's the full parameter surface too - negative prompts, cfg scale, sound, shot composition, scene-segmented multi-prompts, and element lists for visual consistency.

The honest framing: this is a paid API node, so "the node you reach for" means "the node you pay for." Use it for the keeper takes, not for burning through fifty draft ideas - that's what the cheaper Kling 2.6 node is for. When a shot matters, though, the gap in quality is real and visible.

How it works

You give it a prompt (required unless you provide multi_prompt), it submits the job to WaveSpeed, polls until the render finishes, and returns a video_url STRING. Wire that into Preview Anything to watch or download the clip - every video node in this pack returns a URL, not a video tensor.

The inputs that matter

• prompt - the scene description. Be explicit about subject, motion, camera.
• model - Kling 3.0, Kling 3.0 Pro, or Kling 3.0 4K.
• duration - 3 to 15 seconds, default 5.
• aspect_ratio - 16:9, 9:16, or 1:1.
• cfg_scale - prompt adherence, 0–1, default 0.5.
• sound - synchronized audio generation, off by default.
• negative_prompt - what to exclude.

Then the advanced pair: multi_prompt is a JSON array of scene-segmented prompts (mutually exclusive with prompt - the API wants one or the other, and shot_type "intelligent" requires it), and element_list is a JSON array of pre-generated element IDs for keeping visual consistency across shots. These are the fields you'll reach for once you're doing multi-shot sequences rather than one-offs.

The seed widget is cache control only - it isn't sent to the API. Randomize re-runs (and re-bills) each queue pass; a fixed value lets ComfyUI serve a cached result.

Installing it

Part of the ERPK Collection (eRepublik-Labs/comfyui-nodes-erpk). ComfyUI Manager: search erpk → ERPK Custom Nodes → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt

WaveSpeed API key required (Settings → ERPK → API Keys). No model downloads.

Common issues

• Multi-shot setup rejected - you set shot_type to "intelligent" without providing multi_prompt, or provided both prompt and multi_prompt. Pick one.
• "Prompt cannot be empty" - fill the prompt (or supply a valid multi_prompt).
• Jobs take minutes - long renders are normal; the client polls at ~10s intervals with a generous timeout. Don't cancel a healthy poll.
• Cost creeps up on 4K - 4K is the priciest tier. Reserve it for finals and use standard/Pro for drafts.