Bring an old photograph
back to life

Before you start

What it does and does not do

It turns one still photograph into a 6-second seamless loop. The person blinks, smiles, nods or waves, while the background, clothing and framing stay put.

It does not restore photographs. Creases, mould marks and missing corners are carried into the video exactly as they are, and they become more obvious once they move with the face. Clean those up in an editor first if you want them gone. It does not colourise either: a monochrome original stays monochrome.

Which photographs work

A clear face, front-facing or close to it, even light, and one subject in frame. Those four decide most of the outcome. When a mask, a microphone or a hand covers the face, the model invents whatever is hidden, and the result is no longer the same person.

Resolution matters less than people expect. We shrank a portrait to 96 pixels wide, scaled it back up, and the result was almost indistinguishable from the full-resolution original. What really breaks it is the wrong subject: pick a portrait motion for a photograph of a bird and the model grows a human out of nowhere. That happened for real, and the comparison is on the failure cases page.

Match the motion to the photograph

Motions are split into "portrait" and "animals and other", and the split matters more than it looks. A face detector runs on upload, and choosing a portrait motion when it finds no face brings up a warning, though you can still go ahead.

The size of the movement matters too. "Wave" needs room below the shoulders, so a half-length shot suits it better than a head-and-shoulders one. "Turn to look at you" has the most effect on a photograph already shot in profile or three-quarter view. When in doubt use "quiet breathing", which is the most conservative and the hardest to get wrong. Every motion is shown running on the examples page.

Why six seconds, and why the loop is seamless

The model produces 73 frames, which at 24 fps is 3.04 seconds. The start and end of a generated three-second clip rarely line up, so looping it directly shows an obvious jump. Playing it forward and then backward makes the two ends identical, giving a 6.08-second loop with no visible seam. The cost is a slight back-and-forth feel, which is why the motions avoid anything with a clear direction. The implementation is written up in how it works.

How long it takes

A steady 42 to 45 seconds. Most of that is not the computation. Encoding the photograph and the instruction into something the model can read takes nearly 20 seconds on its own. The first clip after an idle spell is noticeably slower, because the model has been swapped out of video memory and has to be moved back.

There is one GPU and it handles one job at a time, so at busy moments you queue. The position on screen is read from the real queue rather than being a timer animation. You can close the page or switch apps while it runs; it picks the progress back up when you return.

What happens to your photograph

The whole generation happens on a GPU in a room at home, with no third-party AI API involved. Your upload and the finished video are both deleted after seven days. No training or fine-tuning step reads your files: the model is downloaded and used as it is, and the weights never change.

Ownership of a result is tied to your account or your browser, not to your network address, because two people behind the same connection would otherwise look like one person. The full description is in the privacy policy.

No account needed

There is a free allowance every day and it asks for nothing. If you want to keep going after that, an account is free, uses an email address as the name, and sends no marketing. Everything is free at the moment.

More questions

How long it takes, what it costs, whether you can use the result commercially, whether you can upload somebody else's photograph, and what to do when the result is not right: those are all in the FAQ. For getting a good copy of a print in the first place, see old photographs.