Wan 3.0 Lip Sync Video

One prompt, one spoken line

Medium A-roll. The woman stands behind the kitchen counter
holding the pink portable blender, speaking naturally to camera.
She says: "I get asked how I make smoothies so quickly."
No subtitles, no text overlays, no UI, no watermark.

• Shot, camera, and who is talking
• The quoted line, and what not to print

2484 / 20,000

One pass out

9:16 · 720P · 30S

Picture onlyWith the spoken line

wan3.0-video · 16:9 · 720P · 14s · speech and mouth in the same pass

Those four lines are lifted from Alibaba's own thirty-second demo, which runs to 2,484 characters against a 20,000-character ceiling. The room is for pinning the speaker, the product and the exclusions — not for writing more. Uploading a portrait instead of describing one changes nothing about how the line itself is written.

The prompt generator takes the line you want said and returns a full Wan 3.0 prompt around it — speaker pinned, one camera move named, the audio layers ranked, and the line delimited so it is performed rather than narrated. It checks the result against the rules this model actually enforces, so the clause people forget is added before the credit is spent rather than after.

1. 01

   Type the line you need said

   One sentence is enough to start. Who says it and where can be a fragment.

2. 02

   Get a prompt built around it

   Speaker, camera, sound layers and the delimited line, in the order Wan 3.0 reads them.

3. 03

   Send it here and add the face

   The prompt arrives in the box at the top of this page with the length and ratio already set.

Open the prompt generator

Writing the line for a paid ad rather than a piece to camera? The UGC ad generator has the word count fifteen seconds actually holds, and the disclosure rules that decide whether it runs.
Already have a clip whose delivery you liked? Video to prompt transcribes the spoken line and hands it back inside braces — the form this model performs rather than narrates.

• 
  01

  Script → prompt

  Hand over the line and get a full shot description built around it, with the speaker pinned and the line already attributed to them.

  Free · no credits
• 
  02

  Deck or page → script

  A PDF or a public URL becomes the story the shot is built from, rather than slides with an avatar reading them.

  50 pages · 100 MB
• 
  03

  Portrait + line → talking take

  The face goes in as frame one, the line goes in quoted and attributed, and the mouth follows it.

  2–30 s · up to 1080P
• 
  04

  New line, same take

  Change what an existing clip says without reshooting it: the speaker, the timing and the framing stay put and the lip movement is regenerated.

  Footage you already have

Everything above lands in the same place — my creations keeps the MP4, the prompt and the task id for every run, so a take you liked can be reopened rather than reconstructed.

1. 1

   Drop in the face

   The check runs instantly, in your browser. A good file shows a green row; a bad one names the fix before you spend anything.

   Checked free
2. 2

   Quote the line, and say who says it

   The exact words in quotation marks, attributed to the person on camera. A paraphrase — "she says something reassuring" — gets you a line the model invented.

   Exact words only
3. 3

   Close with "no subtitles"

   There is no negative prompt field, so the instruction not to print the line is a sentence like the others. It is one clause and it saves a whole run.

   One clause
4. 4

   Set length and resolution, then generate

   Buy the seconds the line needs first. One MP4 comes back with the speech, the room tone and the mouth movement already in it.

   Up to 1080P

Around ninety seconds for a short clip, and you can close the tab while it runs. The uploaded photograph is never modified — the output is a new file.

Frame 001 · your face

002003004005

Spoken, and synced, from here on

Wan 3.0 lip sync puts your photograph in as the literal first frame and generates 2 to 30 seconds forward from it, at 30 fps in 480P, 720P or 1080P. The words you quoted come back as speech in that clip's own audio track, with the mouth moving to them, because the picture and the sound are rendered in the same pass rather than stitched together in two.

Your photo is frame one

001

Your photo is frame one

Audio files you have to supply

0

Audio files you have to supply

Any whole second

2–30 s

Any whole second

MP4 back, sound already inside it

1

MP4 back, sound already inside it

• No text-to-speech step before the render
• No dubbing pass after it
• No second model to license for the mouth

Ten rounds · 720P · 30s

What the mouth needs to land

3

• The exact words, quoted rather than summarised
• A speaker described closely enough to pin
• The camera on their face while they talk

The rule people miss

There is no negative prompt field, so "do not print this" is a sentence in the prompt like any other. Leave it out and the model is free to treat your dialogue as something to render on screen as well as speak — which is why Alibaba's own thirty-second demo, printed above, ends on that exact instruction.

• no subtitles, no text overlaysRight
• Negative prompt: subtitlesWrong

Anything shaped like a negative prompt is read as more words to render, and the request still succeeds — so the failure arrives as a finished clip with the label "Negative prompt" somewhere in the frame.

Braces work too — {like this} is what the prompt tools on this site emit. What the model is actually looking for is a delimited line with a visible speaker attached to it, and Alibaba's own published prompts use quotation marks for it. Text to video covers the four sound layers and the order to name them in.

The line itself has no separate field and no separate limit: it lives in the prompt with everything else, and the prompt holds 20,000 characters. What constrains a spoken line is the clock, not the character count — around two to three seconds of screen time per short sentence, so a fifteen-word line in a four-second clip either gets rushed or gets cut off. Buy the seconds the line needs before you buy the resolution.