Animate an old family photo with AI, without getting it wrong

Restore first: a scratched or faded photo makes a poor video

A video model treats every pixel of your photo as truth. A scratch becomes a scratch that moves, a fold across a face becomes a crease the model tries to hold in place while the head turns, and faded, low-contrast skin gives it almost nothing to build a face from. The result is the typical failure you see in “bring old photos to life” clips: the background moves, the person stays flat, and the face slides into a generic one halfway through.

So before you think about movement, restore the old photo. In Airpic the Old Photo Restorer runs in stages you can compare one by one against the original (Straighten, Damage, Clean up, Faces, Color). What we set for a typical scanned print from the 1950s to the 1980s:

• Straighten automatically on. Most flatbed scans are tilted by a degree or two, and a tilted frame makes any camera move in the video look wrong, because the model treats the edge of the print as a horizon.
• Repair scratches and stains (paint them) on, and paint over every scratch, fold and stain yourself with the brush. Those areas are filled in before anything else runs. Be generous: the fill comes from the surroundings, and a half-painted scratch stays visible and then moves.
• Enhance faces (AI) on, with Strength between 50 and 60 percent. At 100 percent the rebuilt face is sharp but drifts towards a generic face; at 50 to 60 it still looks like the person. With a badly degraded face, add a reference photo of the same person (one sharp photo where the face is clearly visible is enough, up to three from any period). The main face is then rebuilt guided by the references, and Fidelity sets how much of that rebuilt face blends into the original.
• Upscale ×2 only if the scan is small, under about 1500 px on its long side. The animation tool accepts images from 300 to 5760 px per side, so a 3600 px scan doubled would be too big for it, while a 1000 px snapshot doubled gives the model real detail to work from instead of inventing it.
• Colorize if you want colour. A colourised first frame gives the model skin tones to work with. If you prefer black and white, that is a legitimate choice and the video will stay black and white.

If the photo is grainy rather than scratched (a snapshot on fast film, or a scan pushed too far in the scanner software), run it through the photo denoiser first with the “Scan” preset, then restore. For a sharp black-and-white photo that needs nothing else, the photo colouriser alone is enough. The full procedure, including how to scan the print properly, is in our guide on restoring old photos with AI.

One honest limit: if a face in the photo is under about 100 pixels across (twelve people on a 1200 px scan, say), the restorer warns you that the reconstruction tends towards a generic face, and the animation will inherit that. Crop to one or two people before you restore, or accept that the video will show the scene rather than recognisable expressions. Upscaling a tiny face does not bring back an identity that was never recorded.

How to describe the movement: small gestures make an old photo move, big actions break it

The second thing that decides whether a photo to video AI result looks like your relative or like a puppet is what you ask for. The model has one frame. Everything it cannot see (the back of the head, the teeth behind closed lips, the hand hidden behind a coat) it has to make up, and made-up detail is where the uncanny look comes from.

What works, in order of reliability:

• A slow camera move on a still subject: “the camera slowly pushes in”, “a gentle slow zoom towards her face”. Nothing about the person has to be invented, only a little parallax, and image to video models do that well. It is the safest way to make an old photo move.
• A blink and a slight smile: “she blinks once and smiles slightly, without showing teeth”. The last three words matter. A closed-lip smile stays close to the source; an open smile means inventing a mouth.
• Breathing and small shifts: “his shoulders rise slightly as he breathes, the lamplight flickers softly”. Tiny movement reads as alive without changing the face.
• Environment motion: wind in the hair, a curtain moving, leaves, smoke from a chimney, water. These give the clip life while the person stays almost still.

What goes wrong almost every time:

• Turning the head more than a few degrees. The model invents the far side of the face.
• Walking, standing up, waving, hugging, lifting a child. Hands and legs are the weakest point of every video model, and a 1960s dress or a uniform makes it worse because the model has little idea how that fabric moves.
• Speaking. Lip movement without sound looks like bad dubbing, and the model has no idea what was said.
• Two or more people interacting. Ask for one small gesture per person at most, or for camera movement only.

In Airpic you write the description in your own language in the box labelled “Describe the motion (optional)”, up to 800 characters. Our AI turns it into a cinematic English prompt with camera and motion directions and shows it to you under “Prompt written by our AI (what the model receives)” before anything is generated. Read it. If it says “she turns and walks away” when you wrote “she looks up”, rewrite yours more plainly. If you leave the box empty, the AI proposes a motion from the photo itself; for a portrait, check that proposal just as carefully.

A useful trick: describe the period and the light. “1950s studio portrait, soft lamplight, she blinks and smiles slightly, the camera slowly pushes in” keeps the model inside the look of the photo instead of modernising it.

Fast or Maximum quality: which mode turns a picture of a real person into a video

Airpic offers two quality settings when you animate a photo: Fast · 768p and Maximum quality · 720p. The names suggest an easy decision: pick Maximum for anything that matters. For old family photos that is the wrong decision, and not because of cost.

Maximum quality does not accept photos of real people. The video model behind it rejects any image in which it detects a real person, for privacy reasons, before generating anything. In Airpic the job stops, you see a message saying the photo was rejected, and your credits come back automatically. It is not a fault and there is no way around it: a restored portrait of your grandfather will be rejected every time. Maximum quality is for the house, the street, the family car, the garden, a pet, a hand-drawn illustration: anything without a recognisable person in it.

So for portraits, groups, weddings and children, use Fast. The name refers to generation time, about 1 to 3 minutes for a clip, not to a cut-down result. It produces a 768p MP4 of 5 or 10 seconds that keeps your photo as the first frame.

What Maximum quality costs in time, measured rather than estimated: in our production test on 15 October 2026, a photo animated into a 5-second video at 1112 × 834 px took 396 seconds with the quality setting at its highest. That is 6 minutes 36 seconds for 5 seconds of video, which is why the tool page tells you to expect 6 to 7 minutes. Plan for it, and do not close the tab while it runs.

Two more practical points about the settings:

• Choose 5 seconds unless you have a reason for 10. A blink and a slow push-in are over in 5 seconds; stretching them to 10 makes the model add a second gesture to fill the time, and the second gesture is usually the one that looks wrong. Ten seconds works well for environment motion (a harbour, a garden, a market) where there is no face to drift.
• Animating works differently from Airpic's image tools. There is no free preview of the clip: AI credits are deducted when you press Generate and refunded automatically if the video fails or the provider rejects the photo, but not if you simply dislike the result. Two descriptions for the same photo are two generations. The current cost of each mode and length is on the tool page.

Consent and respect: bringing an old photo to life when the person has died

A moving image of someone who has died lands differently from a photograph. Some people find it comforting; others find it intrusive, even frightening, and they are not wrong to. A photograph is a record. An animated photograph is a performance the person never gave: that blink and that smile were chosen by you and invented by a model. Treat it that way.

Before you animate a photo of a relative:

• Ask the people closest to them. A daughter, a widow, a brother. Not for permission in a legal sense, but because the clip will reach them and they should see it from you, not in a group chat. Some will say no. That is the end of the discussion.
• Do not publish without asking. A family video on a private phone is one thing; a reel with a trending song is another. If anyone in the photo is alive, they decide about their own face. If they are not, the closest family decides.
• Keep the gesture inside what the person actually did. If your grandmother never smiled in photos, do not make her grin. A blink, a breath, a camera move that brings you closer to her: that is a gentle use. Making her speak, wave or laugh is putting words in her mouth.
• Label it. When you share it, say it is an AI animation of a photo from 1962, not footage. Twenty years from now nobody in the family should mistake the clip for a home movie that never existed.
• Keep the original. Store the restored still and the untouched scan next to the video. The video is an interpretation; the scan is the document.

There is also a privacy point specific to this tool that you should know before uploading. Airpic's image tools (restoring, colourising, denoising) run on our own GPUs in the European Union and never keep your files. Animating a photo is different: the restored image and your description are sent to the video model provider's API outside the EU, and to the AI that writes the prompt. Airpic does not use your files for training and deletes the finished video 1 hour after you download it, 24 hours at most; what the provider does is governed by its own terms. If a photo is one you would not email to a company abroad, do not animate it. Where each Airpic tool runs is explained in our article on AI image processing in the EU.

Step by step in Airpic: from restored photo to video in 5 or 10 seconds, then join the clips

The whole process for one photo, as we do it, takes about fifteen minutes of your time plus the waiting.

1. Scan or photograph the print. 600 dpi on a flatbed for a 10 × 15 cm print gives roughly 2400 × 3600 px, which is plenty; do not tick Upscale ×2 on a scan that size. No scanner? On a phone, the Old Photo Restorer shows a “Photograph the print” button that opens the camera: shoot in daylight, away from the window to avoid reflections, with the print flat. If what you have is slides or negatives, see our article on scanning slides and negatives at home.
2. Restore. Open the old photo restorer, upload the scan (JPG, PNG, WebP or HEIC, up to 20 megapixels), tick Straighten automatically, tick Repair scratches and stains and paint over the damage, keep Enhance faces (AI) on at 50 to 60 percent Strength, tick Upscale ×2 only if the scan is small, and tick Colorize if you want colour. Press Process image. Processing and the preview are free, and the first full download each month is free as well; after that you pay with credits that do not expire. Tap the stage chips to compare each step with the original; if the face has drifted, lower the Strength or add a reference photo and run it again. Download as JPG: the animation tool takes JPG, PNG or WebP up to 10 MB, and a PNG of a large restored scan can exceed that. If it does, the image compressor is free, keeps the ICC profile, and gets it under 10 MB in seconds.
3. Animate. Open Animate a Photo with AI, drop the restored JPG, write the motion in the “Describe the motion (optional)” box (for example “she blinks and smiles slightly, without showing teeth, the camera slowly pushes in”), set Length to 5 s, set Quality to Fast · 768p for any photo with a person in it, and press Generate video. Read the prompt our AI shows you; if it added an action you did not ask for, change your text and generate again. Wait 1 to 3 minutes. Watch the clip at least twice, once at full size, looking at hands, teeth and the outline of the head. If something drifts, use “Generate another variant” with a plainer description, or cut the request down to camera motion only.
4. Download. Download MP4 saves the clip. Store it with the restored still and the original scan, named with the year and the people in it, so the three stay together.
5. Join several photos into one video (optional). If you have restored and animated five or six photos for an anniversary or a memorial, open the AI video editor, upload the MP4 clips and, if you like, a piece of music. Before it edits anything it asks you the format (16:9 for a television, 9:16 for phones), the target length and the pace; choose a slow pace and fades between clips. Add a short title to each photo (a name and a year), check the preview, which plays free in the browser, and only then export the 1080p MP4. One restrained edit of six clips with soft fades is far better than one with wipes and karaoke subtitles: this is a memorial, not a reel.

If you want to see how the family reacts before doing the whole album, animate a photo with AI once, with the gentlest description you can write, and show it to one person first.

FAQ

Keep reading