AI Lip Sync Video Generator

Your photograph

002003004005

Generated — picture and sound together

Four to fifteen whole seconds at 24 fps, 768P or 2K, in eleven languages we have run end to end. JPG, PNG, WEBP, HEIC and HEIF all go straight in, which is worth saying because most photographs of people are phone photographs and most tools quietly make you convert first. A phone snapshot often animates better than a polished headshot — retouching removes the texture that motion needs. Teeth, if you have the choice: a closed mouth holds no information about what is behind the lips, so the model invents an interior, and invented interiors are what make a result read as wrong.

Your photo is frame one

001

Your photo is frame one

Any whole second

4–15 s

Any whole second

Languages run end to end

11

Languages run end to end

Sound from the same pass

Stereo

Sound from the same pass

• No audio file to make first
• No frame-by-frame timing
• No separate lip-sync pass

The finished clip

What AI lip sync looks like when it comes out right

16:9 · 6s

The full prompt, not a summary

Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3.

6s · 16:9

The photograph above, and the six seconds it became

One run, one prompt, exactly as the model returned it

• 16:9 · 15s

• 16:9 · 15s

• 16:9 · 15s

• 16:9 · 15s

1. 

   01

   Drop in a face

   A photo works, and so does two to fifteen seconds of footage. One face, looking at the camera, nothing across the mouth. iPhone HEIC files go straight in, and the box reads the file back to you — size, shape, weight — before you have spent anything.

   One faceJPG · PNG · WEBP · HEIC
2. 

   02

   Say what they say

   Type the sentence and a language tag and the voice arrives with it — there is no audio file to make first. The line in the box is seventy-three characters of the 7,000 you are given; everything else is for describing the shot.

   73 of 7,000 charactersNo voice to cast
3. 

   Audio track · 32 kHzLIP-SYNC.MP4

   03

   Play it, then change one word

   The clip arrives with the sound already in the file. Most people keep the second or third one, not the first — every attempt is saved, so you are editing rather than starting over.

   4–15s · up to 2KSound already muxed

The first two cards are screenshots of the box at the top of this page, not a drawing of one. The photograph you upload is never modified — what comes back is a new file. The clip on the third card is MiniMax’s own, published with its prompt in the H3 launch post: the still set the face, a vocal track set the singing, a film clip set the camera move. That run took the audio route; the box above takes either.

• 

  16:9 · 8s

  A line straight to camera

  The whole clip is one person and one sentence. Keep it under about twelve words at six seconds — long sentences at short durations are the usual reason a first attempt disappoints.

  Make this one
• 

  9:16 · 9s

  One line of dialogue for a scene

  A character says the line in the shot you already imagined. Describe the person as well as the words — clothing, hair, one distinguishing feature — and the face stays put across the clip.

  Make this one
• 

  9:16 · 15s

  Someone talking about a product

  Person still plus product still, and the spoken lines written shot by shot. This is the one that runs the full fifteen seconds, so it is also the one that shows you where the ceiling is.

  Make this one
• 

  16:9 · 15s

  A vocal, not a spoken line

  Set the tone to sings, or upload the vocal and let it drive the mouth. The uploaded track becomes the finished soundtrack rather than a reference for one, and uploading it is not billed. Keep the clip at least as long as the track.

  Make this one

Starting from a video and the original pixels have to survive? That is dubbing, and a crop-and-paste tool does it better than this page — everything here rebuilds the shot rather than editing it.

For scale: Google's own documentation for Veo 3.1 says English is fully supported and other languages have not been evaluated. Need the same face or the same voice across several clips? Use reference to video.

What a run costs

• The tell

  A mouth that does not match the skin

  A patch at a different resolution from the face around it, and a line along the jaw that crawls when the head turns. Once you have seen it you cannot stop seeing it — and it was never your photograph’s fault.

• The cause

  A 96-pixel square, pasted back

  The usual method finds the mouth, crops a small square around it — often 96 by 96 — generates new mouth shapes at that size and pastes them onto your frame. Everything outside the square never went through the model.

• How to tell first

  Read the verb a tool uses about itself

  Maps mouth shapes onto the face in your video. Blends them into the original image. Onto, into — either way something is being laid over a face that is already there, and the jaw line is where it ends.

• What happens here

  A new frame, sound and all, in one pass

  MiniMax H3 never finds your mouth, because it is not editing your frame. No crop, no paste, no edge to track: the mouth carries the resolution of the cheeks because there is only one resolution in the picture. The voice comes out of that same pass, so you never get an excited soundtrack over a flat face.

1. 01Length is a budget

   Output is a whole number of seconds from 4 to 15. At an ordinary speaking pace that is roughly ten to forty words. Write the line first, count it, then pick the duration — doing it the other way round is how you get a delivery racing a clock it cannot beat.

   4–15s, whole seconds

2. 02One tag is load-bearing

   [Language] selects the phoneme set the mouth is shaped from, so it is the tag that changes the picture. (S1) only earns its place when the frame holds more than one candidate speaker — against a single face it changes nothing you can see.

   (S1) … [English] …

3. 03The character ceiling is not for the line

   7,000 characters covers the whole prompt. A spoken line is tens of characters; everything else is the shot — who, where, lit how, framed how. Running out of room is a description problem, never a dialogue one.

   7,000 characters

4. 04Seconds are the invoice

   $0.08 per output second at 768P, $0.13 at 2K. The line sets the length and the length sets the bill, so a fifteen-second take costs nearly four times a four-second one. Most lines worth hearing are short.

   $0.08 / s · 768P

None of this is guesswork about the model — the durations and the character ceiling are the published limits, and the per-second prices are the ones on pricing. Want the shot described for you? Prompt generator writes the other 6,900 characters.