Lip-sync models

Lip-sync (Липсинк) takes an existing video and a replacement audio track, then changes the visible mouth movement. It does not write dialogue, translate a script, choose a voice or guarantee the speaker remains unchanged.

lipsync-2

The default Lip-sync choice. Start with a video showing a visible person naturally speaking.

Setting	Value
Name	lipsync-2 (баланс)
Availability	available
Default for new nodes	yes

lipsync-2-pro

An alternative focused on face detail. Review teeth, beard and mouth edges in motion; a more detailed face is not proof of timing accuracy.

Setting	Value
Name	lipsync-2-pro (детали лица/зубы)
Availability	available
Default for new nodes	no

sync-3

A distinct newer architecture offered in the same selector. The author describes improved handling of silent lips and difficult angles; this is not a comparative Vini benchmark.

Setting	Value
Name	sync-3 (максимум, 4K, экстрим-ракурсы)
Availability	available
Default for new nodes	no

Kling LipSync

The choice for 3D and 2D characters. Kling states lip-sync for human characters that are real, 3D or 2D, and not for animal characters; the three Sync Labs models are trained on real human faces, and their author lists cartoon faces among unsupported cases. The provider accepts a 2–10 second clip with sides between 720 and 1920 pixels, and 2–60 seconds of audio up to 5 MB. Vini refuses a clip outside 2–10 seconds before any charge and names its length, scales a smaller frame up to 720 pixels without cropping, and re-encodes audio heavier than 5 MB to MP3 of the same length. There is no duration-mismatch setting: when the audio is longer than the clip, the provider trims it. The charge counts input clip seconds rounded up to 5 seconds, so a 3-second clip is charged as 5 seconds and a 7-second clip as 10.

Setting	Value
Name	Kling LipSync (3D и мультяшки, ролик 2–10 с)
Availability	available
Default for new nodes	no

Prepare the inputs

Use a short clip with a clearly visible face and natural speaking movement. Prepare and approve the replacement speech first. Run Lip-sync on one shot at a time rather than on a whole Montage: a wide shot, two characters in frame or a full music mix as the audio leaves the model little to work with. Feed the isolated voice (Voice isolation) and add music in the final edit. Connect video and audio to Lip-sync, choose the model and explicitly review what happens when the durations differ.

For the Sync Labs models, the current default mismatch mode is cut_off, which cuts the longer input. Other choices include looping, bouncing, padding audio with silence and remapping. Choose deliberately: a successful render can still omit the end of a line if the selected duration policy calls for that. Prefer matching input lengths when the entire line and action must be retained.

Studio example and verification

Start with a close or medium shot of one person speaking and a short approved Voiceover sentence. Run Lip-sync, then listen and watch together. Check the first and last phoneme, mouth edges, expression, head turns, and the end of the clip. Compare the original face to the output.

According to the model author, still segments are a limitation for lipsync-2 and lipsync-2-pro; sync-3 can animate silent lips, with its own performance limits. Small, obscured or non-human faces still require care. Author-side advanced options are not implied to be exposed in Vini.

On 22 October 2026 Vini ran one paid comparison on a close, front-facing 3D tomato character with an isolated Russian voice: Kling LipSync, sync-3 and the Sync react-1 face mode all redrew the mouth on that shot. It was a single 5-second clip, not a benchmark; on the same project an 80-second Montage with wide shots, two characters and the full song as audio did not get usable mouths from sync-3 or lipsync-2-pro.

Reviewed 22 October 2026 against the current node and server contract. Current MCP execution support is narrower than UI availability.