Kling Lip Sync Video with Audio

Kling Lip Sync Video with Audio does exactly what it says on the tin: you hand it a video of a face and an audio file, and it resyncs the mouth to match the speech. This is the Kling version of the "talking head" trick, done in the cloud - a built-in partner node that sends your clip and audio to Kling through Comfy's API and returns a VIDEO where the lips actually track the words.

The context you should know: this is the paid-API route in a landscape full of free local options. LivePortrait and ByteDance's open-source LatentSync have covered lipsync locally for a while, and people genuinely debate whether the API versions are worth it. What Kling brings is robustness - the model is built around clean speaker-video generation, and the constraints below are the price of admission. If you just want a quick, reliable result and you already have Comfy credits, this node wins on convenience.

The inputs

• video - the clip with a face to animate. Hard limits from the description: max 100MB, height and width between 720px and 1920px, and 2–10 seconds long. These are enforced before billing, so trim and resize first.
• audio - the track to sync against. Max 5MB, and it must contain clearly distinguishable vocals. Muffled or heavily-mixed audio produces mush.
• voice_language - en (default) or zh. Match this to the speech in your audio; it steers how the model maps phonemes to mouth shapes.

The outputs

Three things, and the extra two matter more than you'd think:

• VIDEO - the synced clip.
• video_id (STRING) - the Kling task's video ID, useful for tracking/API workflows.
• duration (STRING) - the clip's duration, handy if a downstream node needs to know the length without re-probing the video.

Gotchas

• The constraints are the node's whole personality. 720–1920px, 2–10s, 100MB video, 5MB audio - you will hit one of these with a normal phone clip. Pre-process with a resize/trim step in the same graph.
• "Clearly distinguishable vocals" is load-bearing. Background music or two people talking at once wrecks the sync. Feed it a clean vocal stem if you can.
• Video must have a distinct face. Profile shots, occlusion, heavy makeup - all make the model guess.
• Seed isn't part of the schema here, so don't go hunting for reproducibility knobs; this is a straight per-call service, billed through your Comfy account like the rest of the Kling family.

If the local lipsync tools frustrate you with setup and waiting, this is the zero-fuss alternative: upload, wait, download a talking head that matches your audio.