KlingAI Lip-Sync

Dubbing done right is a pain in the neck. This node is the shortcut: you give it an existing video and an audio track, and Kling re-animates the mouth to match the new audio. No required prompt, no reference images - just video and audio in, a re-synced VIDEO out. It's Kling AI Lip-Sync, and it's the node that turns "I need to change what this person says" from a reshoot into a single cloud call.

Why you'd reach for it

Local lip-sync tooling exists but it's fiddly - face mesh, audio conditioning, and a lot of nodes that occasionally melt down. This is one node doing the whole job on Runware's GPUs: keep the performance, replace the words. That covers localization, fixing a botched line, ADR, or just swapping in a different take of dialogue. The providerSettings give you real control over the mix, which is the part people usually miss.

How it works

Both required inputs are strings: video and audio each take a URL, a file path, or a Runware media UUID. They go to the Runware REST API via the runware SDK, Kling regenerates the mouth/lip motion to the new track, and the result returns as a native ComfyUI VIDEO. Because the inputs are references rather than tensors, you can point it at anything Runware can fetch - including clips you've already generated in the same graph.

Inputs that matter

• video (STRING) - required. The source clip (URL/path/UUID).
• audio (STRING) - required. The new audio track (URL/path/UUID).
• providerSettings.klingai.originalAudioVolume (0–2, default 1) - how much of the source clip's original sound survives. Set 0 to kill it completely.
• providerSettings.klingai.soundVolume (0–2, default 1) - the volume of the newly generated audio. These two are how you balance old ambience against new speech.
• numberResults (1–4) - variations, each a separate cost.
• outputFormat (MP4/WEBM/MOV), outputQuality (95).
• safety.mode - none / fast / full content checking.

Output: one video socket (VIDEO type).

Install and the API key

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

Restart ComfyUI (or Runware via ComfyUI Manager). Key from the Runware dashboard → Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

• The inputs are strings, not VIDEO/AUDIO tensors. Trying to wire a video node's output directly in will fail - export/reference it by path or URL first.
• Head-turns and profile shots are where lip-sync degrades first; straight-on, talking-to-camera footage gives the cleanest result.
• It's closed Kuaishou tech, paid per run, with the cost printed on the title bar. Loop with numberResults at 1.
• If your source audio has useful ambience you want to keep, raise originalAudioVolume - the default blend is often better than going full mute.

For anyone who's ever re-recorded a line and wished the mouth would just follow, this node is the fix that's actually good.