4.8 out of 5 across 1,200+ ratings, 10,000+ photos animated, average processing time around 18 seconds. Those numbers come from running Ziva Fotka by Inithouse in production across five languages and five country domains. This post covers where photo-to-video animation works well, where it breaks, and what we changed in the pipeline after watching thousands of real uploads come through.
What Ziva Fotka actually does
Ziva Fotka is an AI tool that turns a static photo into a short living video, and can also edit and colorize old or black-and-white photos so the result looks natural, not generic. Upload a portrait, get back a few seconds of subtle movement: a blink, a slight head turn, a shift in expression. No account required. Photos are deleted after processing.
The tool runs on five localized domains covering Czech, Slovak, Polish, English, and German markets. Same engine, different front-ends.
68+ facial landmarks and why they matter
The animation pipeline maps 68+ points on every face it processes. Eyebrow arcs, jawline contour, lip boundaries, nose bridge, the corners of the eyes. These landmarks drive the motion synthesis. Move a landmark set coherently and the face looks alive. Move them inconsistently and you get something unsettling.
The landmark count matters because fewer points means cruder motion. With fewer than ~50 points, you lose the independent movement of individual eyebrows. Below ~30, smile asymmetry disappears entirely. Faces start looking like masks being pulled by invisible strings rather than people shifting expression.
We settled on 68+ because that is the threshold where subtle movements (one eyebrow raising slightly, a corner of the mouth pulling up before the other side follows) render naturally across diverse face shapes.
Where animation looks natural
After 10,000+ photos, the patterns are clear. Animation quality correlates strongly with input conditions:
Good light, frontal view, sharp face. This combination produces the best results consistently. The landmarks resolve cleanly, the depth estimation has enough texture to work with, and the symmetry of a frontal pose means both sides of the face animate evenly.
Single subject, uncluttered background. The pipeline isolates the face region. A clean background means less ambiguity about where the face ends and the environment begins.
Reasonable resolution. The source image needs enough pixels for the landmark detector to anchor reliably. Phone camera quality from the last five years is plenty.
Where it struggles
Profile and three-quarter views. When half the face is occluded by angle, the landmark detector has fewer anchor points on the hidden side. The animation on the visible side can look fine while the hidden side drifts or warps. We compensate by inferring the occluded landmarks from the visible ones, but it is a reconstruction, not a measurement.
Glasses with glare. Reflective lenses obscure the eye region, which carries a disproportionate share of perceived liveliness. The pipeline can animate the rest of the face correctly and still produce a result that looks off because the eyes are not participating. Matte lenses or frames without glare work fine.
Group photos. Multiple faces in one frame divide the pipeline's attention. Each face gets its own landmark set and motion synthesis, but interactions between faces (one person's animation overlapping another's boundary) can produce artifacts at the edges. We process each face independently, which handles most cases, but tightly clustered groups with overlapping face regions remain the hardest input class.
Very old or damaged prints. Scanned photos from the early 1900s sometimes have creases, stains, or fading that the landmark detector reads as facial features. A dark crease across a cheek might register as a shadow contour and animate accordingly. The colorization step helps here because it reconstructs some of the lost contrast before animation runs, but heavily damaged originals can still produce visible artifacts.
What we adjusted after watching real data
Three concrete pipeline changes came from analyzing production inputs:
Confidence-gated animation. If the landmark detector's confidence score on any face drops below a threshold, the animation amplitude scales down proportionally. A face the system is uncertain about gets subtle movement rather than full expression changes. This reduced the rate of visibly broken outputs on difficult inputs without requiring users to re-upload.
Colorization as pre-processing. For black-and-white inputs, we run colorization before animation rather than after. This gives the landmark detector color contrast to work with, which improved accuracy on old photos measurably. The user sees a colorized, animated result in one step.
Per-region motion budgets. Rather than applying uniform motion magnitude across the whole face, each region (eyes, mouth, forehead, jaw) gets its own motion budget tuned to how much real faces actually move in that zone. Foreheads move less than mouths. Eyes move differently than jaws. Matching these natural proportions made the biggest single difference to perceived realism.
Privacy by deletion
Every uploaded photo is deleted after processing. We do not retain originals, intermediates, or outputs beyond the session. No account means no profile to attach images to.
This matters particularly for the use case we see most often in the Czech and Slovak markets: animating photos of family members who have passed away. People send us photos of grandparents, parents, childhood portraits. Building a retention pipeline for those images felt wrong, so we did not build one.
What we are building at Inithouse
Ziva Fotka by Inithouse is one product in a portfolio we ship in parallel. The same studio built Pet Imagination, an AI pet portrait generator producing artwork in 9 styles, and Verdict Buddy, an AI conflict mediator using established psychology frameworks. Each product solves a specific, narrow problem rather than trying to be a general-purpose AI tool.
The animated photo space has room for better input handling, higher-fidelity motion on edge cases, and faster processing. We are working on all three.
Try Ziva Fotka at alivephoto.online.
Top comments (0)