How to Make an Old Photo Talk
A scanned photograph of someone who is gone, given a voice and a moving mouth. What works, what fails, and the part worth thinking about before you do it.
LivePhotoLabThe request usually arrives the same way: someone finds a photograph while clearing out a house, and wants to hear the person in it say something again. That is a different job from animating a modern selfie, and old pictures fail in their own specific ways.
Start with the scan, not the tool
Most disappointing results are decided before any AI is involved.
Photograph the print flat, in daylight. Put it on a table near a window, not under a ceiling light, and hold the camera parallel to the print. A phone camera is fine. Angled shots make the face lean, and the model will animate the lean.
Fill the frame with the face. A face that occupies 200 pixels in a 4000-pixel scan of the whole album page gives the model almost nothing. Crop to head and shoulders before uploading.
Leave the damage alone for now. Cracks and spots matter less than sharpness. A slightly damaged but sharp face animates better than a clean but blurry one.
What decides whether the mouth works
Lip sync needs to find a mouth. Three things break that:
- A hard side profile. Anything past a three-quarter turn and there is no mouth shape to move.
- A hand, pipe, or moustache covering the lips. Common in portraits from the 1940s and 50s, and it is the single most frequent failure we see.
- Very low contrast. Faded prints where the lips and the skin are nearly the same tone leave the model guessing.
Eyes closed is fine. Hats are fine. Black and white is fine — the output stays black and white, which usually looks better than a colourised guess.
Write the line the way they spoke
A generated voice reads what you type, and typed sentences are stiffer than spoken ones. Two habits help.
Keep it under about twelve seconds — that is roughly 180 characters, which is also the limit here. One sentence, maybe two.
Write the contractions. "I'm going to the shop" sounds like a person; "I am going to the shop" sounds like a form letter. If they had a phrase they always used, use that instead of something grand.
Four voices are available, two female and two male. None of them will be their voice. That is worth accepting up front rather than being disappointed by.
The part worth sitting with
There is a real difference between a memorial video that says this is my grandmother, and here is her photograph moving and one that puts new words in a dead person's mouth and shows it to relatives without saying it was generated.
The first is a keepsake. The second tends to upset the people it is shown to, especially the ones who were closest. If you are making this for a funeral or an anniversary, tell the family what it is before you play it.
And the rule that does not bend: photographs of living people need their permission. A tool that makes anyone say anything is a tool for making anyone say anything.
What it costs
Talking photos run on purchased credits, because each one pays for a generated voice and a generated video. The $4.99 Mini Pack covers two of them, with no subscription, and credits stay valid for a year. A run that fails returns its credits to your balance automatically, so a photo that turns out to be unusable does not cost you anything.
If you only want the picture to move — a breath, a small smile, no speech — the old-photo animator does that, and the first video there is free.