Kling 2.6 Image-to-Video

Kling 2.6 sits a generation above 2.5 Turbo, and its headline trick is joint audio-video generation - the model produces synchronized sound with the footage. That's why the very first thing to know about this node is in the default settings: sound is true out of the box, and the tooltip is blunt about what that means - it doubles the cost. Read that again before you queue your first clip.

It's part of the ERPK Collection, the community pack wrapping the WaveSpeed AI API. No local model, cloud-rendered, billed per clip.

How it works

You connect a starting image (or a URL), describe the motion, camera moves, and audio, and the node uploads the frame to api.wavespeed.ai. It polls every 10 seconds up to a 15-minute timeout, then returns a video_url STRING. Wire that into the pack's Preview Anything node to preview and download. As with the other Kling nodes, the seed widget is cache control - no real seed parameter on this endpoint, so fixed seed = cached clip, randomize = fresh billed run.

The inputs that matter

• model - Kling 2.6 or Kling 2.6 Pro. Everything interesting is gated on Pro: cfg_scale, end_image, and sound.
• image - your start frame as an IMAGE tensor; preferred over image_url (URL fallback, max 10MB, min 300px per side, aspect 1:2.5 to 2.5:1).
• prompt - scene motion, camera moves, and audio all in one box. Since the model generates audio too, writing "wind, birdsong" style cues actually lands.
• sound - defaults to true and doubles cost. Flip it off when you're iterating on visuals only. Also: it cannot be combined with end_image - the API refuses the combination, so you pick one.
• end_image - Pro only, an end frame for guided transitions. If you want sound, this is off the table.
• cfg_scale - Pro only, 0.3–0.8, default 0.5. Higher follows the prompt more closely. Same idea as guidance_scale on 2.5 Turbo, different name.
• negative_prompt - what to exclude, from both visuals and audio. "No watermark", "no sirens" - it works.

Install

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

Or ComfyUI Manager → search erpk → ERPK Custom Nodes. No model files to fetch. You need a WaveSpeed API key with credits, set under right-click canvas > ERPK Settings (Settings > ERPK > API Keys).

Common issues

• "Cannot be combined with sound" errors - you connected end_image (or end_image_url) while sound is on. They're mutually exclusive on this API; disable one.
• Unexpectedly big bills - sound defaults on, and it doubles the per-clip cost. If you didn't intend it, flip sound to false.
• Fixed seed clip never changes - cache control, as on the other Kling endpoints. Randomize for a new generation.
• Clip ignored your audio cues - the model treats audio as a prompt-conditioned element, not a guarantee. Keep cues short and physical.

The play: leave sound on for the shots that will actually be heard, turn it off for the rest, and remember 2.6's sound is the differentiator over the 2.5 Turbo nodes - which is also why they're the cheap option.