How to Animate a Still Photo Without Losing the Face

Animating a still photo is not the same job as generating a video from a text prompt. You already have the look, the framing, and the face. The job is to add movement without letting the model drift, morph, or replace what made the photo worth using. Most failed photo animations fail for the same reason: the request asked for too much motion, too much camera travel, or too many things happening at once. This guide covers the small-motion workflow we use in the photo-to-video workspace — how to set the first frame, pick a modest camera move, write a prompt that describes movement rather than content, and review the output for identity and anatomy before you commit to a final render.

Prova questa ricetta

Da foto a video AI

Apri lo strumento completo

Caricamento dello studio...

Start with the first frame, not the story

The uploaded image is not an inspiration for the clip; it is the first frame. That distinction changes how you choose the source. A portrait with a clean background, even front lighting, and a fully visible head and shoulders will animate more predictably than a cropped party photo with half a face at the edge of frame. The model has to extrapolate everything outside the borders and everything that was never recorded, and the more it has to invent, the more likely it invents something you did not want.

Before uploading, check three things. First, resolution: a source image that is too small will produce a soft, blocky clip, and upscaling it beforehand usually gives a better result than relying on the video model to sharpen it. Second, orientation: pick a crop that matches the output aspect ratio you plan to render, because letterboxing or a forced reframe at generation time introduces artifacts. Third, and most important, the face itself: if it is turned more than about three-quarters away, partly obscured, or heavily filtered, identity preservation becomes much harder to control and you should expect more review passes.

Treat the source as a contract. Whatever expression, lighting direction, and head angle it shows is the state the clip begins from. If you want a smile, start with a smile. If you want the subject looking left, start with a photo looking left. Asking the model to arrive at a different face partway through the clip is exactly where identity drift begins.

Why small motion is more reliable

Model-generated motion accumulates error. Each frame is conditioned on the previous one, so a small mistake in frame twenty becomes a larger mistake in frame eighty. When you ask for a dramatic turn, a fast gesture, or a full-body walk, you give the model more opportunities to make those small mistakes — and once the jawline or the hands start to drift, the clip rarely recovers. Small motion keeps the error budget low. A slight head tilt, a blink, a breath, a drifting highlight: these are changes the model can carry across five seconds without losing the underlying structure of the image.

This is also a perceptual point, not just a technical one. Real photos, when they come alive for a second or two, tend to move very little — a subtle shift in expression, a breeze in the hair, a gentle push-in. A still that suddenly swings its head and gestures broadly reads as animation, and usually as artificial animation. Restraint makes the clip feel photographed rather than generated.

Practically, you can think of motion in three tiers. Tier one is micro-motion: breathing, blinking, a slow head tilt of a few degrees, hair and fabric settling. Tier two is a single camera move with a static or near-static subject. Tier three is subject action plus camera movement plus environmental motion, all at once. Tier one and tier two are where photo animation works well. Tier three is where it usually falls apart, and it is often better handled by text-to-video, where the model has no pre-existing face to contradict.

Choosing a camera move for a still

A camera move gives the clip a sense of shot design without asking the subject to perform. It is also the safest kind of motion, because the face can remain essentially unchanged while the frame does the work. The trade-off is that camera moves must be described precisely; vague requests like "cinematic camera" give the model room to invent movement that fights the source image.

For a single still, a slow push-in is the most forgiving option. It keeps the subject centered, increases the importance of the face, and hides small inconsistencies at the frame edges because they drift out of view. A slow pull-back works if the source has enough environment around the subject; if you cropped tight, pulling back forces the model to invent background, and invented background is where limbs and shoulders start to distort. Lateral tracking and parallax moves are riskier on a portrait because the subject is usually fixed and the model has to reconstruct what is behind them. Orbit moves are riskiest of all, since even a small rotation requires the model to imagine the side of the face it has never seen.

Pay attention to direction as well as type. A push-in toward the face heightens attention; a push-in toward the chest or hands can de-emphasize the identity you are trying to protect. If the source has a strong leading line — a corridor, a shoreline, a row of windows — align the camera move with it. For a deeper treatment of how camera language translates into prompt language, see the guide on text-to-video camera movement; the same vocabulary applies here, just constrained to a single source frame.

Writing motion into the prompt

The prompt for an image-to-video run should describe how the scene changes, not what the scene contains. You have already told the model everything about content by uploading the image. Repeating the subject description tends to pull the model toward its own idea of that subject and away from the pixels you provided.

A workable structure is: camera behavior, then subject micro-motion, then atmosphere, then what to hold constant. Keep it short. Here is an example prompt for a portrait:

Slow push-in on the subject, subtle natural head tilt, gentle breathing motion, a single blink, soft light shift across the face. Keep the face shape, expression and features unchanged. Static background, no camera shake, no zoom beyond a slight push-in.

Note what is missing: no description of the person, no mention of clothing colors, no adjectives about beauty or mood. Note also the negative clauses at the end — they are not guarantees, but they make the intended bounds explicit. "No camera shake" and "keep the face unchanged" are instructions the model can weigh against the rest of the prompt.

If you want a specific action, keep it singular and small: "hair moves slightly in the breeze," "eyes glance to the side and back," "shoulders lift in a slow breath." Avoid compound requests such as "turns, smiles, and lifts a hand," because each element has its own failure mode and they will interact.

Checking identity and anatomy

Review the clip before you commit to it. The checks below are ordered from most to least common failure. Watch the clip at normal speed first, once, without pausing — if the clip reads wrong at speed, no amount of frame-by-frame analysis will save it. Then scrub through slowly or download the frames and inspect the problem areas.

Identity preservation is not something the model can guarantee. What it can do is stay close to the source when the requested motion is small, the face is clearly visible, and the prompt does not ask for expression changes. If a pass drifts, the fix is almost always to reduce motion, not to add more instructions. Adding "do not change the face" three times does not recover a clip that has already lost the jawline in the second second.

For anatomy, look at hands and wrists first; they are the most reliable indicator of whether the model is holding together. If the hands are out of frame in the source, keep them out of frame in the prompt too, or you are asking the model to invent them.

Duration, resolution and cost

Five seconds is a good default for a photo animation. It is long enough for a slow push-in and a breath, short enough that small errors have less room to compound, and it loops acceptably. Going longer is possible, but each extra second increases the chance of drift and the cost of re-running. If you need a ten-second clip, consider generating two five-second clips from the same source with a slower move in each, rather than one long generation.

Resolution should match your delivery target. Higher resolution does not improve motion; it improves detail. If the final use is a social post or a website hero, rendering at the platform's display size and letting the browser scale down usually looks better than rendering oversize and letting the platform compress. Match the source image's aspect ratio and keep the frame size reasonable relative to the source resolution; upscaling a small photo into a large render tends to produce soft skin and smeared edges.

Cost is shown in the studio before you submit, based on duration and resolution, and the exact credit cost is displayed in the confirmation step. That is the number to check — there is no fixed per-second rate to quote here. Re-runs are the main cost driver in practice, so budget for two or three attempts per usable clip and make each attempt count by changing one variable at a time.

Turning the clip into a usable loop

A five-second clip is rarely used at full length. Most uses are shorter, looped, or slowed. Plan for the loop when you set the motion. A clip that starts and ends in roughly the same pose loops seamlessly; a clip that ends with the head turned or the camera pushed in will jump when it repeats. A slow push-in is the worst loop of all, because the frame composition has changed by the end. The fix is to make the motion cyclic: a breath in and out, a head tilt that returns, a drift of light that returns. If you need a push-in, use it as an opening shot that cuts away, not as a loop.

Two small production tricks help. First, export at a frame rate that matches your delivery and avoid resampling after the fact; resampling introduces its own artifacts. Second, when a loop has a visible seam, a short crossfade of a few frames at the join hides the discontinuity without requiring a new render.

Finally, keep a written record per clip: the source image, the prompt, the duration, the resolution, and what you changed between attempts. Photo animation is a small-motion craft, and small-motion craft is mostly iteration. A one-page log of what worked on a specific face saves more time than any prompt template. When the workflow is stable, check the current plan options and production volume against your actual render count rather than a guess.