Bringing old family photos to life
The most common use of this tool isn't selfies — it's the faded print from the back of a drawer, and seeing someone blink again decades later. This page covers what's specific to old photos, and a few things worth thinking about first.
Photographing the print well
With old photos, almost everything depends on how you capture the print. If you're using a phone:
- Diffuse light, no flash. Flash leaves a bright patch on the paper, and that patch becomes a moving highlight in the video. An overcast window, or two lamps from opposite sides, works best
- Shoot square-on. An angled shot distorts the face, and the model animates the distortion
- Crop out the surroundings. Album sleeves, frame edges and the table underneath all become part of the picture as far as the model is concerned
- Fill the frame. Don't stand three metres back
- A flatbed scan at 600 dpi always beats a phone capture
One caveat worth stating: we tested this, and resolution matters less than expected — a portrait crushed to 96 pixels wide and blown back up produced a near-identical result (see failure cases). What matters when copying prints is glare, distortion and clutter, not pixel count.
Creases, mould spots, fading
This tool does not restore photos. Damage is carried straight into the video and animated along with the face, which makes it more obvious, not less. The sensible order is:
- Repair obvious creases, spots and scratches in an editor first
- If the face itself is damaged, patch it until the features are complete — the model needs outlines, not perfection
- Don't bother correcting fading or a yellow cast: the model doesn't care about colour, and the warmth usually looks good
- Then upload
Black and white photos
They work, and usually work well. Without skin tones there's nothing for the model to get wrong about skin tones. Lincoln, Mark Twain and Tesla in our examples are all 19th and early 20th century black and white photographs, and none of them look odd in motion.
The output stays black and white — this tool does not colourise. Run a colourisation tool first if you want that.
Group photos
With several people in the frame the model doesn't know who should move; you tend to get the whole row twitching, or features bleeding between adjacent faces. Crop out the one person you want and upload that. Do them one at a time if you want everyone.
Which motion to choose
| Photo | Motion | Why |
|---|---|---|
| Old, mediocre quality | Quiet breathing | Smallest motion, least chance of the face drifting |
| Formal, unsmiling portrait | Smile and nod | Stern to warm is the most striking change |
| Already smiling snapshot | Burst out laughing | Extends the existing expression, most natural |
| Three-quarter view | Turn and look | Exactly what that motion was built for |
| Half-body with hands visible | Wave hello | Hands need to already be there |
Photos of people who have died
This is the most common use of a tool like this, and the one that deserves the most care.
- It is the model's guess, not them. The way the smile forms, the rhythm of the blink, the size of the nod — all of it is an average learned from hundreds of thousands of faces, not their actual mannerisms. Treat it as a photograph that moves, not as "they're back"
- Think about the rest of the family first. The same clip reads as comforting to one person and deeply unsettling to another. Ask before you send it to an elderly relative
- Don't put words in their mouth. Using this to fabricate something they said is explicitly forbidden — see the terms
- Your photo and the result are deleted from our machine after 7 days, so download what you want to keep
Other things people use it for
- Pets. Use the animal motions — cats, dogs and birds all work. Photos of pets that have died are a very common use
- Profile pictures. A moving avatar is much more recognisable than a static one
- Invitations. The host waving hello on an event page
- Exhibits and teaching. A historical portrait in motion lands very differently from a still