Back to Blog

Why AI Photo Animations Look Uncanny (and How to Fix It)

Blinking too fast, rubbery mouths, drifting eyes. Here is what actually causes the uncanny look in AI photo animation, and the five settings that fix most of it.

Posted by

The uncanny valley is a timing problem

Most people blame AI photo animation for looking fake because of resolution or detail. In practice, resolution is rarely the issue. Motion timing is. A portrait that moves at the wrong speed reads as wrong long before anyone notices a soft edge or a missing eyelash.

Human faces have a narrow band of believable movement. A real head drifts a few millimetres per second while breathing. A blink takes roughly 100 to 400 milliseconds. When a model compresses a full breath into one second, or makes an eyelid snap shut in 40 milliseconds, viewers register it as a puppet even if they cannot name what is off.

Four failure modes you can spot in one pass

  • Eye drift. Pupils slide sideways because the model optimises for face similarity, not gaze stability.
  • Mouth rubber. Lips stretch beyond their original width when a talking or singing prompt is applied to a closed-mouth photo.
  • Background wobble. The subject moves and the wall behind them breathes with it, which no camera does.
  • Hair freeze. Strands stay rigid while the head turns, so the silhouette looks cut out.

The five settings that fix most of it

These are the levers that matter, in the order we change them when a clip feels wrong.

1. Keep the clip short

Five seconds is the sweet spot for a still image. Longer clips force the model to invent motion it has no evidence for, and invention is where artefacts start. If you need a longer video, cut three short clips and join them instead of asking for one 15-second take.

2. Describe motion, not appearance

The prompt should say what moves and how fast. Words like gently, slowly, subtle head turn and natural blink do more than any quality keyword. Describing the person again (age, clothes, lighting) pushes the model toward re-rendering the face, which is what you do not want.

3. Protect the eyes

If the platform exposes a motion strength or expression control, keep it low on portraits. Eyes are the first thing viewers check and the easiest thing to break.

4. Start from a sharp, front-lit photo

A soft or noisy source forces the model to hallucinate detail, and hallucinated detail flickers frame to frame. Run a restore or upscale pass first when the source is a scan from the 1970s.

5. Accept less motion

The most convincing animations we have produced move very little: a breath, a blink, a two-degree head turn. Restraint reads as authenticity because real photographs of real people mostly hold still.

When uncanny is the point

Trend effects such as dance templates, squish and hug animations are not trying to pass as documentary footage. They are memes, and a slightly plastic look is part of the format. Judge those against a different standard: does the motion land the joke in the first two seconds? If yes, artefacts are acceptable.

The uncanny problem only really matters for memorial, family archive and portrait work, where the goal is a believable continuation of a real moment. That is exactly where the five settings above earn their keep.