This page covers what makes lip sync look right, how to prepare audio for the cleanest result and how it handles different languages and singing.
What makes an AI lip sync video look natural
Viewers notice three things: whether consonants like p, b and m close the lips, whether vowels open the mouth to the right shape, and whether the rest of the face reacts. A mouth that moves perfectly inside a frozen face still looks fake. Generating the full head and upper body from the audio is what makes OmniHuman 2 clips feel spoken rather than dubbed.
Timing comes straight from your audio file, so there is no manual alignment step and no drift over long clips.
Preparing audio for clean sync
Small audio fixes have the biggest effect on accuracy:
- Record close to the microphone in a quiet room; avoid echo.
- Remove music and sound effects under the voice — add them back after rendering.
- Trim long silences at the start and end so the video doesn’t sit still.
- Keep levels steady; very quiet passages are harder to read.
- Use MP3 or WAV between 4 seconds and 5 minutes (recordings are converted automatically).
Any spoken language
Lip movement follows the sound itself, so speech in English, Spanish, Turkish, Arabic, Hindi or any other language works the same way. The optional direction prompt has its own list of supported languages, but the audio doesn’t. This makes the model useful for localising one presenter photo into several languages — render the same image with each dubbed track.
Lip sync for songs
Singing is harder than speech: long vowels, breaths and rhythm. Give the model a clean vocal — an a cappella stem or a mix with the instrumental turned down — and describe the performance in the prompt (“eyes closed on long notes, sways to the beat”). The full mix can go back on in your editor.
Common artefacts and how to fix them
Most issues trace back to the inputs and can be fixed without changing the model settings. If the mouth looks blurred, the photo is usually too small around the face or slightly out of focus. If the lips lag, there is often leading silence or a soft fade-in at the start of the file. If teeth flicker, a wide-open smile in the source photo is usually to blame — a relaxed, slightly closed mouth gives the model a cleaner starting point.
- Blurry mouth → use a sharper, larger face crop.
- Lag at the start → trim silence before the first word.
- Mumbled syllables → re-record closer to the mic, without music.
- Flickering teeth → choose a photo with a neutral or closed-mouth expression.
FAQ
Do I need a video to lip sync?
No. OmniHuman 2 generates the video from a single photo; you only need an image and an audio file.
Which languages are supported for lip sync?
Any spoken language. The mouth follows the sound, not a text transcript.
Why does my lip sync look off?
Usually background music, echo or a face partly hidden in the photo. Clean, dry audio fixes most cases.
Can I lip sync a song?
Yes. Use a clean vocal track for best results and add the instrumental back later.
How long can the audio be?
Between 4 seconds and 5 minutes per video.
Can I fix lip sync on footage I already filmed?
This tool generates new video from a photo. For re-timing existing footage, a dedicated video lip-sync tool is a better fit.
Does the video include my original audio?
Yes. The downloaded MP4 contains the exact audio you uploaded, synchronised with the generated performance.
Is there a minimum audio length?
Yes. Audio must be at least 4 seconds long; anything shorter can be padded with a short pause at the end.