AI Lip Sync: A Talking Avatar from a Photo and Audio

TL;DR. Lip sync is when AI moves the lips of the person in a photo to a given soundtrack. As input you need two things: a portrait image and audio. The output is a video where the person speaks your recording. In Pixyn this is Kling AI Avatar v2 in two variants, Standard and Pro; the clip length is set by the audio length, so the price depends on it too. Cost: from 18 tokens on Standard and from 24 on Pro.

---

Three different tasks called by one word

When people search for "lip sync", they mean at least three dissimilar things. The choice of model depends on what you already have.

What you have as input	What you get	Which tool
A photo + audio	the person from the photo talking	an avatar (this article is about it)
A finished video + audio	a re-voicing that hits the lips	lip sync on a video
A finished video + text	synthesized voice-over and sync	lip sync from text

The second and third options do not create the person anew: they redraw the lips in your clip. This is what you need to translate a finished video into another language: the picture stays yours and the speech changes. In the third option you do not need an audio file at all: the voice is synthesized from the text, and you can set the language and the speed of speech.

Next, the first case, the most common one: one photo and a soundtrack.

What you need as input

Just three fields, two of them required:

Input	Required	What it affects
A photo	yes	who speaks
A soundtrack	yes	what is said and for how long
A prompt	no	a hint about behavior in the frame

There is no separate duration setting: the clip comes out exactly as long as your audio. This matters for the price: the base includes the first 3 seconds, and each next second adds to it. A two-minute recording costs noticeably more than a fifteen-second one, and you see it on the button before the run.

Which photo will do

The model works with the face, so almost everything depends on the source:

• The face large and frontal. A profile or a half-turn gives noticeable distortion when the mouth moves.
• Nothing covers the mouth and chin: a microphone, a hand, a scarf.
• Even light. The model redraws harsh shadows on the face in its own way.
• One face in the frame. With a group photo the result is unpredictable.

Don't do this: take a frame from motion where the face is blurred and expect crisp articulation.

Lip sync does not restore details that are not in the source. A blurred mouth stays blurred, and the motion only emphasizes it. Take a sharp frontal photo: it affects the result more than the choice between Standard and Pro.

How to prepare the audio

Sound matters more here than it seems: the model syncs the lips exactly to it, and anything that makes the speech harder to make out turns into uneven articulation.

• Record in a quiet room. The model takes background hum and echo as part of the track.
• Speak at an even pace. Tongue twisters and sharp pauses sync noticeably worse.
• Remove music from under the voice. If you need a background track, put it on after generation, in editing.
• Trim the silence at the start and the end. It goes into the clip length, and so into the price.

No voice of your own? You can first voice the text with speech synthesis and feed the resulting file here as audio. It is a common two-step combo; see AI text to speech.

Standard or Pro

Pro gives neater facial expressions and holds the face better on long recordings, but it also costs more, both in the base and in the per-second addition.

The practical approach: test on Standard, make the final on Pro. If Standard gave an acceptable result on fifteen seconds and you need a minute, move to Pro. If Standard noticeably distorted the face, the cause is more often the source photo than the model version.

Where to use it

1. A talking head for a training video: without a shoot or a studio.
2. Localization. One photo of the speaker plus tracks in different languages.
3. An avatar for social media. Short vertical clips with one character.
4. A voiced presentation. Slides separately, a talking presenter separately.

---

Try it. Create an account: new accounts get 6 tokens. They are enough for a test picture; a talking video starts from 18 tokens and depends on the audio length. Plans are on the pricing page.

FAQ

How long is the finished video? As long as your audio. There is no separate duration setting.

Can I voice a text first and then do lip sync? Yes, it is a common scenario: speech synthesis first, then the resulting file is fed here as audio.

Why does the price change? It depends on the length of the soundtrack: the base includes the first 3 seconds, and each next second adds to it. How spending works: the card about tokens.

Will a drawing or a 3D character work instead of a photo? Sometimes, but the model is trained on people: the closer the source is to a real portrait, the steadier the result.

Can I translate a finished video into another language? Yes, but that is the second scenario from the table above: you take your clip, not a photo, and feed a new track or text to it. The face and the setting stay yours, only the speech changes. Other Kling models are covered in the Kling guide.

Source: the official description of Kling AI Avatar v2 by Kling AI. The input is exactly an image, audio and an optional prompt, and the video length follows the length of the soundtrack.