A talking photo: making a still portrait say a lineAnimate a photo with AI

A talking photo: making a still portrait say a line

Animating a photo and making it speak seem like neighbouring jobs, but in fact they take different routes. That is why most “animate” buttons have no speech at all: there the portrait goes into the model as the first frame, and for talking it has to be sent differently. This chapter covers how exactly, what to write in the request and how to supply a particular voice.

In one paragraph

The portrait is sent as a reference, not as the first frame. The line is written in words, in quotes, right in the request. The lip movement is computed together with the picture in a single pass, which is why the lips match the sound instead of roughly landing on it.

If the voice has to be a particular person’s, you attach 2–15 seconds of their recording to the request. That reference is not billed separately.

Your photoPhoto as the first frameMotion, wind, breathing, a camerapush-in. Sound is the ambience.You can set the last frame too,and the model invents the movebetween your two photos.Photo as a referenceThe person says your line; lipmovement is computed with thepicture, in a single pass.A 2–15 second recording canset the voice. It is free.One request, one route. “Wink and then say a line” has to be split into two jobs.
Motion and speech are different routes. A portrait as the “first frame” gives motion; a portrait as a “reference” plus a voice recording gives a person who speaks. They cannot be mixed in one request: that is a limit of the engine itself.

Why you cannot just “animate it and add speech”

This is a limit of the engine itself, not of the interface: the “first frame” and “reference” roles cannot be combined in one request, and a request that mixes them is simply rejected. So “let them wink and then say something” is not one complicated request but two jobs, and they have to be done one after the other.

There is also a fallback worth knowing about so that you do not use it: generating a voice-over separately and laying it under a silent clip. Formally you get a “talking photo”. In practice it shows in the very first second: the mouth lives its own life and the sound lives its own. The difference is exactly that in the first case the model solves one problem (frames and sound together), and in the second, two problems that do not know about each other.

The line has to fit in the frame

The most common way the result goes wrong is not the prompt or the photo but the length of the line. A clip is billed in whole seconds, speech does not compress, and a two-sentence greeting will not fit in a four-second frame: the person gets cut off mid-word, and the money for the frame is already spent.

Will your line fit

A clip is billed in whole seconds, and speech does not compress. Write the line and see how long a frame it needs.

3.8 seconds of speech6 seconds of clip42 characters

It will fit

You will need a 6-second clip. It costs more than a four-second one exactly in proportion to its length. A shorter line means a cheaper frame.

Finished request

Use this portrait as a reference and make a clip in which the person says one line:
“Happy birthday, Mum. I love you very much.”
Appearance, age, hairstyle and clothing exactly as in the photo, unchanged. Static camera, minimal motion: the person speaks and moves slightly, like a living person.
Must not appear: face distortion, background music, readable text in the frame.
Length: 6 seconds.
Open the chat and attach the portrait

The pace used here is about 11.1 characters per second, measured on our own Russian narration. English usually reads a little faster per character, so treat the seconds as an upper estimate: a clip picked from it leaves room to spare rather than cutting the line off.

The one-line rule

One line per clip. If there are two lines, that is two clips joined into one file afterwards. Joining costs nothing, while a fifteen-second frame costs almost four times as much as a four-second one.

Your own voice: what exactly to attach

The voice recording is a separate input in the same request. The requirements are simple, and almost every mistake falls into one of four:

  • Between 2 and 15 seconds. Shorter is not accepted; longer is not needed.
  • One speaker, no music and no second voice in the background. The model takes the timbre from the recording, and it treats everything else in it as part of the voice too.
  • Ordinary speech, not a whisper or a shout. Extremes are reproduced worse than the middle of the range.
  • Do not buy a synthesised voice-over for this. The engine needs a recording of a real voice, and it takes the words from the request. Paid synthesis adds a bill and adds nothing to the clip.

And one subtlety that is more honest to state plainly: write the line in the request even when you attach a recording. That is how the engine developer’s own example is written, with the words in the prompt and the audio cited as the voice, and this wording gives the right clip whether the model reads the words from the recording or from the text.

2–15 sVoice reference recordingNot billed separately.
1 linePer clipTwo lines means two frames and a free join.
1 passPicture and sound togetherWhich is why the lips match the sound.

What this page should not help you do

A talking photo is someone’s face saying words they never said. The rule is simple and not technical: a face and a voice belong to a person, not to whoever has the file.

  • Your own portrait, a portrait of someone close with their consent, an invented character, a historical illustration: fine.
  • Someone else’s face without asking, a public figure with a statement they never made, someone’s voice lifted from their own video: not fine, and this is the case where “technically possible” and “okay” part ways completely.
  • Label the finished clip as generated if it goes beyond your private chat. That is cheaper than any explanation after the fact.

Attach a portrait and a voice recording, and describe what the person says.

Loading…

What to read next

How animation works in general and why the source decides the outcome: animate a photo with AI. What the voice-assistant button can and cannot do: “Alice, animate my photo”. If you need not a talking portrait but a full presenter in the frame, that is a neighbouring job: a talking avatar.

Frequently asked questions

How do I make a photo talk?
Attach the portrait as a reference, write the line in quotes right in the request and order the clip. The lip movement is computed together with the picture in a single pass, so the lips match the sound. Important: this is a different route from ordinary animation, where the photo goes in as the first frame, and the two cannot be combined in one request. You can try it in the Twelver chat.
Can a photo speak in my voice?
Yes. Attach a recording of your voice, 2 to 15 seconds long, to the same request (one speaker, no music, no second voice in the background) and say that the voice should follow this recording. That reference is not billed separately. You should still write the line in words in the request.
How long should the line be?
One line per clip. Measured on our Russian narration, speech runs at about 11 characters per second; English is usually a little faster, so treat that as an upper estimate. A clip is billed in whole seconds from 4 to 15, so a long greeting will not fit in a short frame and the person gets cut off mid-word. Two lines are better shot as two frames and joined: joining costs nothing.
How is this different from “adding a voice-over to a video”?
In one case the model solves one problem, in the other two unrelated ones. When speech and frames are computed together, the lips land on the sound. When a finished voice-over is laid under a silent clip, the mouth lives its own life, and that shows in the very first second. Laying audio underneath remains a fallback, but it is more honest to say up front that the lips will not match.
Will an ordinary animated photo, with no line, have sound?
Yes. Sound is created together with the picture in the same pass and is not billed separately: wind, footsteps, the hum of a room. The side effect is that if you say nothing about sound, the engine composes a soundtrack itself, so silence has to be asked for in words.
Can I animate a photo of a celebrity and make it say something?
Technically the engine does not ask who is in the photo, but the consequences do. A face and a voice belong to a person, not to whoever has the file: someone else’s face with a statement they never made is exactly the case where “possible” and “okay” part ways completely. Fitting uses are your own portrait, a portrait of someone close with their consent, an invented character or a historical illustration; and label the result as generated if it goes beyond your private chat.

In the Twelver chat a portrait can say your line, even in your own voice if you attach a short recording. Your first clip is earnable without paying: sign up, complete five first tasks and come back on the second and third day.

Try it yourself

Everything in this guide runs inside Twelver

One chat for text, images, video, music and voice — no separate services or subscriptions.

Open Twelver chat
Related pageVideo Generation
Rate your experience