The Photo Looked Right. Why Did the AI Video Change the Composition?

1. MartynFoster30 September 2026, 3:56 GMT +02:00

A still photograph can feel completely resolved before animation begins. The crop is deliberate, the subject sits at the right distance from the edge, vertical lines are straight, and the background supports rather than competes with the main subject. Then motion is added. The camera appears to push forward, the subject shifts slightly, or a hand begins to move—and the original balance changes.

The person may grow too large in the frame. A building edge may lean. Empty space may disappear, and a background object may slide behind the subject. Together, these changes can make the video feel as if it came from a different photograph.

This is the central risk when a finished still becomes an image-to-video reference: the source image defines one arrangement, while the generated motion has to invent what happens before, after, and around that arrangement.

Five things to lock before adding motion
Before writing a motion prompt, identify what must remain stable. I would separate the photograph into five groups.

1. Crop boundaries and negative space

Note how much space exists above the head, beside the product, or in the direction the subject faces. If that space carries the composition, it should not vanish during a push-in or pan.

2. Horizon and vertical lines

Architecture, interiors, streets, and seascapes expose small perspective changes quickly. Record where the horizon crosses the frame and which walls, doors, poles, or windows need to remain vertical. A subtle rotation can make a carefully corrected photograph look unbalanced.

3. Subject scale and silhouette

Check the subject’s size relative to the frame and the outline that makes it recognizable. A shoulder, chair back, or bottle neck may change when an unseen angle is inferred. A changing silhouette alters the visual weight even when the subject remains recognizable.

4. Foreground, middle-ground, and background relationships

Write down which elements overlap and which must stay separated. A foreground branch may frame a face in the still but cover it after a pan. A distant lamp may appear to move toward the subject if depth is interpreted inconsistently.

5. Light and fixed background details

Identify the key-light direction, major shadows, and background details that anchor the scene. Reflections, window frames, signs, shelves, and repeated patterns are useful checkpoints.

Why a small movement can change the whole frame

A still image does not contain the hidden side of a person, the space beyond the crop, or the exact depth between every object. When a prompt asks for a pan, a dolly move, or even a slight body turn, an image-to-video workflow must create new views from incomplete information.

A push-in can become a simple enlargement, a physical camera move, or a combination of camera and subject motion. A pan may reveal content outside the original border. A head turn may require the shoulders, hair, background separation, and shadows to be reconstructed together.

This does not mean that motion should be avoided. It means the movement request should respect the information available in the photograph. The more the requested shot reveals beyond the original view, the more opportunities there are for scale, perspective, edges, and background details to drift.

Check the first, middle, and final frames

Do not judge the clip only at normal speed. Pause on the first, middle, and final frames, then compare each one with the original photograph.

At the first frame, check whether the generation already changed the crop, aspect relationship, horizon, or subject size. If the first frame is wrong, the problem begins before the motion develops.

At the middle frame, inspect overlaps and perspective. Has the foreground crossed into the subject? Are straight lines bending? Has the subject become wider or narrower? Does a fixed background object suggest a changed depth relationship?

At the final frame, look for accumulated drift. Compare the remaining negative space, the subject’s distance from each border, the horizon, vertical lines, shadows, and repeated background details.

For a short image-to-video draft, a tool such as Vidu Q3 (https://viduq3.com/) can be used after these fixed elements are identified. Keep the first attempt simple: one main movement, a clear camera direction, and an explicit note about what must stay unchanged. The generated clip still needs the same three-frame review; the reference image guides the result, but it does not prove that every compositional relationship will remain fixed.

Decide what to change next

Reduce the movement when the first frame is correct but the middle and final frames drift. A shorter push, a slower pan, or one small subject action usually asks the workflow to invent less unseen information.

Replace or prepare the reference image when the source creates ambiguity. Tight crops, foreground obstruction, weak depth cues, cut-off limbs, reflections, or repeated backgrounds can make spatial relationships difficult to infer. A wider crop or cleaner version may provide better boundaries.

Regenerate with clearer constraints when the movement is reasonable but important fixed elements change. Name the stable feature directly: locked framing, fixed horizon, unchanged subject scale, stationary background, or no camera rotation. Change one major instruction at a time so the effect of the revision remains visible.

The goal is not to force every pixel to stay frozen. It is to preserve the visual decisions that made the original photograph work. Motion should extend the composition, not quietly replace it.

For photographers using image-to-video tools, which type of photograph gives you the most trouble: architecture with straight lines, portraits near the frame edge, product shots with reflections, or scenes with foreground objects crossing the subject?