The first old photograph I animated was a front-facing portrait with slightly faded edges. I wanted a gentle blink and a small smile, not a cinematic performance. My first attempt moved too much. The face became someone else. That failure taught me the whole rule of this workflow. With old photos, subtlety is not a limitation. It is the point.

Here is the process I use now for family portraits, historical photos, and memorial clips. It works because it starts with restoration and ends with restraint.

Choose the right photo

I look for a face that is large enough to read. Front-facing portraits work best because the model can infer natural motion. Both eyes should be visible, and the face should not be hidden by damage, glare, or a watermark. A high-resolution scan beats a photo of a photo every time.

I avoid group photos as the first test. Multiple faces increase the chance that one person warps or the model chooses the wrong subject. Side profiles and extreme angles are also harder. They can work, but they are not where I start.

Repair before animating

Old photos often need help before AI motion. I clean dust and scratches, fix contrast, and remove yellowing if the original color matters. If the scan is small, I upscale gently. Heavy sharpening can create plastic skin, so I prefer a light touch.

For black-and-white photos, I decide whether to keep the monochrome look or colorize it. These are different creative choices. If the purpose is memory, I usually keep the original tone. If the purpose is social content, subtle color can feel more alive. Either way, I restore the still image first.

Write motion like a memory, not a stunt

My prompts are short because the motion should be believable. I use one action and one camera idea.

Keep the person from the reference image unchanged. She blinks naturally and smiles slightly, soft light, static camera, warm vintage tone

For a male portrait, I might use this.

Keep the face from the reference image unchanged. He slowly turns his head a little, natural blink, static camera, gentle film grain, respectful tone

The phrases that help most are natural blink, slight smile, slow head turn, static camera, and preserve facial features. The phrases I avoid are dramatic, cinematic transformation, smile widely, or look around the room. Those sound exciting, but they increase identity drift.

Pick a model by risk

For a first test, I use a free or low-cost model to check whether the source photo is good enough. If the face remains stable, I move to a production model. Kling v3 is my choice for high-fidelity faces. Kling v3 Omni works well for smooth portrait motion. Seedance 2.5 can be good for more emotional style, but I use it only after I know the likeness holds.

For family portraits, I prefer slower motion. For social content, Agnes Video 2.5 Flash can add more energy. But the more expressive the model becomes, the more carefully I need to check the face.

Aspect ratio and duration

For a profile or memorial clip, 1:1 often preserves the original composition. For TikTok, Reels, or Stories, 9:16 needs a crop that keeps the face centered. For YouTube or a slideshow, 16:9 gives room for text or multiple photos.

I keep most single-photo animations between three and five seconds. That is enough for a blink, a slight smile, or a small head turn. Longer clips increase the chance of drift and often feel less natural.

Inspect frame by frame

I watch the clip once for feeling, then again frame by frame. I look at the eyes, mouth, hairline, ears, and background. If the eyes blink naturally but the mouth changes too much, I remove the smile. If the background shifts, I write static background, subject only moves. If the face looks younger or different, I reduce motion and use a stronger identity-preserving model.

This is the step people skip. A clip can look fine at full speed and still lose the person by the final frame.

Build a memorial version

For a memorial video, I select five to ten photos in chronological or emotional order. I animate each one separately with a similar prompt. Then I connect them in an editor with soft transitions and music. I do not ask one prompt to animate a collage.

I also keep text simple. Names, dates, and short phrases work better in the editor than inside the generated video. This keeps spelling correct and lets the family update the words later.

Respect and disclosure

If the photo is of a real person, especially someone who cannot consent, I keep the motion respectful. I do not make them speak, say things they never said, or place them in invented situations. A blink and a slight smile are honest. A fabricated performance is not.

For public history projects, I label the output as AI-assisted animation. It preserves trust and separates reconstruction from fact.

What I would repeat

Scan at high resolution. Restore first. Crop to one face. Use gentle motion. Keep the camera still. Test with a free model. Check frame by frame. Keep clips short. Save the winning prompt.

That is how an old photo becomes alive without becoming someone else.

How to Animate Old Photos with AI: Complete Guide for 2026