How to Animate a Photo With AI: What Happens Between the Still and the Video

Animating a photograph looks like magic: a still image starts breathing, blinking, turning its head. The mechanics are fairly clear, and knowing them immediately explains why some photos animate convincingly while others turn into something unpleasant to watch.

What the model gets and what it does

The input is one image plus a description of the desired motion. The output is a sequence of frames where the first matches the source and each subsequent one differs from the previous just enough for the movement to read as continuous.

The key point: every frame except the first is invented. The model does not find hidden information in the still about what lay beyond the turn of the head — it fills that in statistically: how a human ear usually looks when the head has turned fifteen degrees, how hair falls, how the shadow on a cheek changes.

Why this is fundamentally harder than generating a picture

An image has one job: to be plausible. A video has two: every frame must be plausible and all frames must be consistent with each other. Models find the second job much harder.

Inconsistency shows up as flicker: features shift slightly frame to frame, fabric patterns jump, the background trembles. The human eye is acutely sensitive to exactly this kind of inconsistency — the same error in a still would go unnoticed.

Where the motion comes from

There are two different approaches, and they give different results.

• Motion from description: you say what should happen ("a slight head turn", "wind in the hair") and the model builds a matching trajectory. Flexible, but unpredictable in detail.
• Motion from a reference: the model takes a finished clip as the source of the trajectory and transfers it onto your image. Predictable dynamics, but limited to the references available.

The first suits situations where the scenario matters; the second, where the motion must land on a precise rhythm.

Movements that work well

Anything the model has seen many times in training and that does not require revealing hidden parts:

• Facial micro-motion: blinking, a slight change of expression, breathing.
• A small turn or tilt of the head within ten to fifteen degrees.
• Hair and light fabric moving in the wind.
• A slow push-in or pull-out without changing the angle.
• Background motion: clouds, water, foliage.

Movements that break

• Large turns: past ninety degrees the model has to draw half a face that was never in the still. The result is a different person in profile.
• Hand movement with changing finger positions — the same defect as in stills, multiplied by the number of frames.
• Interaction with objects: picking up a cup, opening a door. Contact physics is not respected.
• Walking and any whole-body movement: legs are a known weak zone.
• Speech with articulation and no audio reference: lips move unnaturally.

What to do with the source

The requirements match those for image generation, plus two specific ones. First, margin at the edges: if the head touches the border, movement immediately runs past it and the model starts inventing the cropped part.

Second, a simple background. Detailed backgrounds tremble the most: each frame redraws the fine detail slightly differently, and the eye picks it up as noise across the whole picture.

Realistic expectations

Animating a photograph produces a short expressive fragment, not a video shoot. Five to ten seconds of micro-motion that turn a static portrait into a living frame for a story, a cover or a greeting.

Expecting a scene with plot, action and dialogue guarantees disappointment: that is a different task, and it is solved by filming.