Go looking for a way to make a photograph talk and the workspace hands you a cast list. There is a panel of people who will say whatever you type, and there are 162 of them, and not one of them is the person in your photo.
This piece takes both routes to a talking clip. It counts what is in that cast list, then takes a single generated portrait and asks the video generator to make him say one line, with a control that asks him to say nothing, and measures the mouth frame by frame against the sound that came back with it.
The cast list you get instead of a tool
The avatars live where you would not think to look: add a text layer in the editor, and the right-hand panel gains a tab called AI avatars next to Presets, Basic, Text to speech and Animation. Whatever the text layer says is what the avatar will say.
Scrolling the panel to the end and reading every label: 162 entries, twenty first names, and pose or wardrobe variants under each.
Most of the variants are framing and wardrobe rather than different people. Seventy-four are labelled Standing, thirty-five Holding phone, thirty Walking, three Sitting, four Christmas, and sixteen are named by what the person is wearing — Dan-Polo, Eve-Hoodie, Kai-Suit. The panel has a checkbox for applying an avatar to all text and an Add button that carries a Pro badge. What it has no control for is bringing in a face of your own.
That is the honest shape of this feature. It is a way to have a presenter deliver a script, which is a real thing to want. It is not a way to make your photograph talk.
The other route: a talking clip from the photograph you already have
The route that does use your own picture is the video generator. You attach a photo, describe what the person does, and it returns five seconds. The photograph used here was generated too, because this piece is not going to animate a real person's face to demonstrate a point.
The starting picture: front on, mouth closed, plain wall, head and shoulders. It was generated for this article, and no real person's photograph was used anywhere in it.
Three requests went out from that one file, differing only in what the man was asked to do.
The man in the photo looks straight into the camera and says the sentence "I told you I could cook." His lips move clearly through the whole sentence and he raises his eyebrows once as he finishes. He stays in the same place, head and shoulders in frame, and nothing else in the room moves. Locked-off tripod shot, no zoom, no push in, no dolly, no pan and no crop, and the last frame has the same framing as the first frame.
What came back: 1280 by 720, 24 frames per second, 5.04 seconds. This clip has sound — press play rather than expecting it to autoplay. Shown here as frames in sequence; the full video and a GIF version are supplied alongside this document.
The clip arrived with a voice already on it
This is the part worth knowing before you plan anything. The file came back with an audio track carrying speech, peaking at 0.62 on a scale where 1.0 is the ceiling. Nothing was asked for except that the man say a line, and no audio was uploaded.
The control request makes the point sharper. The same photograph, told to look at the camera with his mouth closed and blink twice, came back with an audio track that is digitally silent — a peak of 0.000. The voice appears when you ask for speech and not otherwise.
Whether the mouth is moving, measured against a face that is not
The control clip is what gives the other two a scale. Its mouth still moves a little, because the face breathes and blinks.
Measuring the frame-to-frame change inside a box on the mouth, the plain wall behind him reads 0.36 and the not-speaking clip reads 1.07. Setting the line at the control clip's 95th percentile, 1.90, the talking clip crosses it in 69 percent of its frames and the not-speaking clip in 5 percent, which is where that line was defined to put it. The mouth is doing something, and the something is roughly fourteen times the background noise.
The mouth and the voice track each other, until you ask for more
Both panels come from the same three files. No audio was added to any of them.
For the plainly delivered line, mouth movement and audio level rise and fall together, at a correlation of +0.54 across the 120 frame-to-frame steps. The two speech bursts in the sound sit under the two clusters of mouth movement.
The theatrical version breaks that. Asked for widened eyes and a broad open mouth on the last word, the clip came back louder — a peak of 0.89 against 0.62 — with the mouth above the control line in 92 percent of frames, and a correlation of −0.10. The chart shows why: its sound stays near the top of the range through the last full second while the mouth movement has already fallen back to the floor. The performance and the audio come apart at the end, which is exactly where a punchline would sit.
Making the workspace read its own clip back
Knowing there is speech is not the same as knowing what it says. There is a way to find out without leaving the page: drop the clip on a timeline in the editor, open Captions, and run Auto captions on it. The workspace's own speech recognition then writes down what it hears.
The readback, and the two caption clips it produced on the timeline.
It read back two segments. From 0.53 to 2.43 seconds: em>I told you I could cook/em> — the line, word for word. Then from 3.20 to 4.03 seconds: em>told it/em>. The line arrived once and in full, and then a second, partial utterance arrived a second later that nothing in the request asked for.
That second fragment is the thing to plan around. It is the reason the clip is five seconds rather than two and a half, and it is the reason a clip like this wants a trim before it goes anywhere.
What the four requests cost
The cost lines as shown under each request, read on 14 September 2026.
Seventy-six credits, of which fifty went on the two clips that exist to make the comparison. One photograph plus one talking clip is 26. The photograph and the clips were made through the AI image generator and AI video generator flows. The credit balance in the header read 24,091 before the first request and 24,015 after the third clip, which matches the four cost lines; running the captions afterwards did not move it. These figures are for these configurations on this account and this date, and other models, durations and resolutions are charged differently.
Write the line first, then cut to what came out of him
The plan this playbook usually starts from — find a funny audio clip, then find a face to put in front of it — is the wrong way round here. The face arrives with a voice attached, and the mouth is timed to that voice. Drop in a different audio track and the timing you were given goes with it; what you are left with is a face moving to nothing in particular.
So write the line before you generate, put it in the request, and treat the result as footage rather than as a finished post. Then run the captions on it and read what actually came out. In this run that readback is what showed the line was intact for two seconds and that a stray fragment followed it, which turns the edit into a decision about where to cut rather than a question about what to dub. The comedy you were planning to add in post is a sentence you write before you spend the 25 credits.
Everything above was generated, including the man. Putting a photograph of an actual person into a tool that makes them speak is a different act from animating a cat or a toy: it produces a recording of someone saying words they did not say. Use your own face, or a face whose owner has agreed to this specific use, and do not make a real person appear to say something they have not said. Do not do this with photographs of children. The composer's own line asks you to use only media you have rights and consent for, and this is the case where that sentence is doing real work.
The avatars route has its own boundary: those 162 people are supplied under the workspace's terms, and the Add button is marked Pro, so read what the plan covers before you build a posting schedule on them.
Also not covered: adding your own audio track, the Text to speech voices, clips longer than five seconds, resolutions above 720p, vertical framing, and cutting several of these together into one post.
Written 14 September 2026. The portrait and all three clips were generated in CapCut for this article; no real person, photograph or voice recording was used. All three clips are 1280 by 720, 24 frames per second, 121 frames, 5.04 seconds, and carry the audio the generator returned with them. Interface labels, the avatar count and the cost lines reflect a single session on that date and the configurations named above, and may change. Mouth motion is the mean absolute difference between consecutive frames inside a box covering the lower third of the detected face box; the same measurement on a fixed patch of the wall behind him is quoted as the noise floor. Audio levels are per-frame root-mean-square in decibels at 24 frames per second. The correlation is Pearson's, taken over the 120 frame-to-frame steps of each clip. Caption text and timecodes are the workspace's own, converted from minutes:seconds:frames at 30 frames per second.