From Static Photos to Moving Portraits

Of everything these tools do, animating a photograph of someone who is gone is the one that lands hardest. It is also the one where the difference between what the technology does and what it appears to do matters most.

What is actually happening

The model has never seen that person move. It has no access to how they held their head, how they blinked, what their smile did on the way to being a smile.

What it has is an enormous amount of footage of people in general. From that it learned what faces tend to do — the small settling movements, the way hair drifts, the rhythm of a blink — and it applies that general knowledge to the specific face in your photograph.

So the output is not a recovered moment. It is a plausible interpretation of a still, rendered convincingly. That distinction is easy to lose precisely because the result is affecting.

Why it affects people so strongly

Motion is one of the strongest cues we have for presence. A still photograph reads as a record of something past. The same face blinking reads, at some level below thought, as someone being there.

That response is genuine and there is nothing wrong with wanting it. The reason to keep the distinction clear is not to spoil the effect — it is that families share these, and a clip presented as footage is a different claim from a clip presented as an animation.

Getting a good result

The technical requirements are unforgiving in a specific way: the model can only propagate what is already in the photograph.

Scan, do not photograph. A flatbed scan at 600 dpi gives far more to work with than a phone photo of a print, and avoids the perspective distortion and uneven lighting a handheld shot introduces.

Restore first, as a separate step. Repair damage, then upscale, then animate. Three operations, checked individually. Bundling them makes it impossible to tell which one invented the thing that looks wrong.

Choose a photo where the face is large. Group photographs animate badly — a small face has too little structure for the model to hold coherent, and it drifts into someone else.

Front-facing or slightly turned. Profiles and extreme angles fall outside what the models handle.

Leave some background. Camera movement needs somewhere to go.

Ask for one small motion. "A slight turn of the head and a blink" produces something believable. Anything more ambitious produces something that announces itself.

The restraint matters more here than anywhere else. A small, quiet movement reads as the photograph breathing. A large one reads as a puppet.

What to check

• First frame against the original. It should be recognisably the same person.
• Last frame against the original. This is where drift shows.
• The hairline and the boundary with the background, which is where the model's uncertainty collects.
• The eyes. If they change shape or spacing during the clip, the identity has moved.

Generate two or three times. Variation between runs is significant, and with a face that matters to you, it is worth choosing rather than accepting.

Keep the original

The one piece of advice specific to archive material.

Keep the unmodified scan, unchanged, alongside anything you make from it. Label which is which. An animation is an interpretation and a restoration is a repair — both are worth having, and neither should quietly replace the record.

This is the same principle that governs upscaling a family photo: the tool invents detail in a face that was never captured, which is fine for a print to hang and not fine as a historical document.

The wider question

Whether to animate a photograph of someone who has died is a question for the people who loved them, and people answer it very differently. Some find it a gift; others find it wrong in a way they struggle to articulate.

What is worth separating from that is the factual question, which has a clear answer: the clip is not footage, it does not show how they moved, and presenting it as though it does is the part that causes harm. The general framing is in the ethics of AI-generated photos.

Photo to Video runs in the Kitana apps, alongside the restoration tool among the thirteen photo tools. The technical mechanics are in how photo to video works.