Across the portfolio at Inithouse, a studio running parallel product experiments, we handle a lot of user-uploaded images. But one product pushed us into territory we did not expect: scanned archive photos from the 1940s through 1990s, sometimes cracked, faded, or shot on expired film.
Živá Fotka (Alive Photo) is an AI tool that turns a static photo into a short living video. It can also edit and colorize old or black-and-white photos so the result looks natural, not generic. After processing over 10,000 photos, we mapped out the workflow that consistently produces the best results from scanned originals.
Here is that workflow, end to end.
What you are starting with matters
Not all scans are equal. A flatbed scan at 600 DPI with dust removal is a different input than a phone snapshot of a print on a coffee table. The pipeline handles both, but the results shift depending on what you feed it.
Before uploading, three things worth checking:
- Crop tight around the subject. Background clutter confuses facial detection on low-resolution scans. A centered face with minimal margin works best.
- Exposure. If the scan is washed out or too dark, a quick levels adjustment in any photo editor helps. The AI compensates, but starting closer to balanced gives it less guesswork.
- Physical damage. Creases, tears, water stains across the face area reduce detection accuracy. If the damage is outside the face, the tool handles it well. If it cuts across the eyes or mouth, results will vary.
How 68+ facial point detection handles old photos
Modern portrait detection typically maps 68 or more landmarks across a face: jawline, eyebrows, nose bridge, lip contours, eye corners. On a crisp, well-lit selfie, this is trivial. On a grainy 1960s wedding photo, it is a different problem.
What we observed across our processing pipeline:
- Black-and-white photos with clear contrast work surprisingly well. The model does not need color to find facial structure.
- Faded photos (low contrast between skin and background) sometimes need a contrast boost before upload.
- Group shots with multiple faces: the tool detects and animates the primary (largest) face. For group photos, cropping to one person first gives better results.
The detection runs in roughly 18 seconds on average. When it succeeds, the animation maps natural micro-movements (subtle blinks, slight head turns, breathing motion) onto the detected landmarks.
Colorization: before or after animation?
This is the question we get asked most. Živá Fotka supports both colorization and animation, and the order matters.
| Source type | Recommended approach | Why |
|---|---|---|
| Black-and-white, good contrast | Colorize first, then animate | Color gives the animation more visual depth; skin tones make micro-movements look more natural |
| Black-and-white, damaged or faded | Animate first, then colorize the video frame | Animation on B&W is more forgiving of imperfections; colorizing after avoids amplifying artifacts |
| Already color, old/faded | Skip colorization, animate directly | Re-colorizing a color photo can shift tones unpredictably |
| Color photo, modern | Animate directly | No editing needed |
| Sepia-toned prints | Colorize first | Sepia confuses the colorization less than true B&W, and the result is more vivid |
We measured this across hundreds of processing runs. The "colorize first" path produces results rated 0.3 points higher on average (on our internal 1-5 quality scale) when the source is clean B&W. But when the source has damage, animating first and colorizing the output avoids the colorizer "painting over" cracks and stains.
Sharing the result
Once the animation is ready, there are three ways to get it out:
- Direct link. A shareable URL that plays the video in any browser. No app install, no account needed.
- QR code. Printed QR that links to the video. We see this used at family reunions and memorial events, where someone prints the QR next to the original photo.
- Greeting card with custom text. A formatted page with the animated photo, a personal message, and the QR code. Useful for birthdays, anniversaries, or as a gift.
One pattern we did not anticipate: families using the greeting card format for memorial purposes, animating a photo of someone who has passed and sharing it at a gathering. We approach this with care. The tool does not add expressions or emotions that were not in the original photo. It preserves the person as they were.
What we learned building this
This workflow (scan preparation, detection, colorization sequencing, sharing) emerged from watching how real users interacted with the tool. We did not design it up front. We observed 10,000+ processing runs and mapped the paths that produced consistently good results versus the ones that did not.
The pattern is similar to what we see across the Inithouse portfolio. With Pet Imagination, our AI pet portrait generator, upload quality drives output quality in the same way. With Magical Song, the input (occasion details, names, relationship) shapes the output just as much as the model does. The common thread: when the user input is specific and well-prepared, AI tools deliver better results. Not because the model is smarter, but because it has more signal to work with.
If you have a box of old family photos and a flatbed scanner, alivephoto.online is where the workflow starts. No signup required. Photos are deleted after processing.
Top comments (0)