Composing a Still Image So an Image-to-Video Model Has Room to Generate Motion

You feed a single still image into an image-to-video model and expect a smooth dolly, a slow pan, or a subject walking across the frame. More often the output is a near-static clip, motion collapse, because the frame gave the model nowhere to go. The structural constraint is that every current model predicts pixel displacement from that single reference frame; if the image is tightly cropped, has no depth cues, and leaves no headroom, the model defaults to the safest prediction, which is no movement at all. The principle that fixes this is image-to-video composition for motion headroom and depth. You compose the still so the model has data it can extrapolate into displacement vectors. This page explains exactly which compositional elements, foreground, midground, background separation, subject headroom, limb clearance, prevent motion collapse and give a camera-control system like Runway Gen-3 Alpha's director-mode the axes it needs to work.

Every image-to-video failure mode maps to a missing compositional element. Background sliding, where the whole scene moves as a flat plane, happens because the image lacks a foreground element close enough to the camera to create parallax separation from the midground and background. Limb mutation happens because the subject's arms or legs touch the body or the frame edge, so the model has no boundary information to keep them separate. Identity shift happens because the face occupies too much of the frame and the model, starved of context, interpolates toward the average face in its training data.

The rule for camera motion composition is: the still must contain at least three visible depth layers, foreground, midground, background, with clear spatial gaps between them. A car parked 2 metres from the lens, a person standing 8 metres behind it, and a building 50 metres back gives the model parallax information to dolly or pan. Without those gaps, the model treats the whole image as one plane and slides it sideways, background sliding, not a camera move.

The rule for subject motion composition is: the subject must have headroom in the direction of intended movement equal to at least one-third of the subject's height, and every limb that will move must be separated from the body and from other limbs by at least a pixel gap the model can read as a boundary. A person with arms crossed over the chest will produce limb mutation on the first frame of motion; a person with arms held 10 cm away from the torso gives the model the boundary it needs to keep the arm as a separate object.

Camera motion, dolly in, dolly out, pan left, tilt up, is the most reliable kind of movement generated by current models because the camera-control systems in tools like Runway Gen-3 Alpha director-mode and Kling 2.6 have dedicated axes for it. But those axes are useless if the source image does not contain the spatial information the model needs to calculate the displacement. A pan requires horizontal parallax between foreground and background; a dolly requires scale change in objects at different distances. Without those cues, the model falls back to background sliding or, in the worst case, motion collapse.

What a Camera-Motion Source Image Must Contain

The source image needs a foreground object that overlaps the midground and background. A tree branch in the top-left corner, a fence post at the right edge, a person standing 3 metres from the lens, any object that creates a near-far relationship. The subject of the shot does not need to be the foreground object; it can be in the midground, but the foreground must be present. Kling 2.6's camera control (pan, tilt, zoom, roll) and Luma Dream Machine Photon's controls (pan, tilt, zoom, dolly) both rely on these depth cues to calculate the transformation.

Why Leading Lines Work

A road, a railway track, a corridor, a row of pillars, any strong diagonal or converging line that recedes into the frame gives the model a depth gradient it can use to calculate forward or backward camera movement. The leading line defines a z-axis. Without it, a dolly-in request produces a flat zoom that looks like the image was scaled, not that the camera moved through space. Runway Gen-3 Alpha's director-mode documentation explicitly recommends leading lines as a compositional element that makes its camera-control panel produce physically plausible movement rather than background sliding.

The Parallax Separation Minimum

For a pan or dolly to read as a camera move and not as background sliding, the distance between the nearest foreground object and the farthest background object must occupy at least 30 percent of the frame's depth. You measure this by estimating how many metres or feet the scene represents from the closest object to the horizon line. A close-up of a face against a wall 1 metre behind it has no usable parallax. A person standing 5 metres from the camera with a building 50 metres behind them has a depth ratio of 10:1, which gives the model strong parallax data.

Motion collapse is not a random failure; it is the model's safe default when it cannot determine which pixels should move. Every pixel displacement prediction carries a risk of introducing an artefact, so the model's loss function penalises movement that deviates from the source image. If the source image provides no spatial structure that favours one displacement over another, the safest output is zero displacement, a static clip. Depth cues are what break that deadlock.

The three-layer rule is the minimum. Foreground layer: objects that overlap the frame edge or are close enough to the lens that they would move fastest in a real camera move. Midground layer: the primary subject of the shot, at a distance where its displacement rate relative to the foreground creates parallax. Background layer: the horizon or far wall, which moves slowest or appears static. A landscape with a foreground rock, a midground tree, and a mountain in the background gives the model three velocity bands to calculate against. A portrait with the subject 2 metres from a flat wall gives it one.

Stable Video Diffusion 1.1, which has a motion strength slider (0-100) and camera control for pan, tilt, zoom, still fails on single-layer images because the model architecture, latent diffusion with temporal attention, has no built-in 3D scene understanding. The spatial data has to come from the image itself. Composing the still with these three layers is the only reliable workaround.

Subject motion, a person walking, an arm waving, a bird taking off, is harder for image-to-video models than camera motion because it requires the model to move a discrete object across the frame while keeping its shape, texture, and identity intact. The failure rate is higher, and the compositional requirements are stricter.

Headroom for the Direction of Movement

The subject must have empty space in the direction it will move. If the subject will walk right, there must be at least 30 percent of the frame width to the right of the subject. If it will jump upward, there must be headroom above the subject equal to at least half the subject's height. Without that headroom, the subject will drift into the frame edge and the model will either crop it or mutate it to fit, limb mutation or partial identity shift. Kling 2.6's start-and-end-frame feature can mitigate this because you can specify the final position, but even with two frames, the model needs the headroom in the source image to generate the intermediate frames.

Limb Separation as a Compositional Requirement

A limb touching the body, another limb, or the frame edge creates a continuity problem for the model. It cannot determine where the limb ends and the body begins, so the temporal attention mechanism merges them over the clip duration. The result is limb mutation: an arm that fuses into the torso, fingers that merge into a blob, a leg that becomes part of the background. The fix is to compose the still with clear separation. Arms held away from the body by at least 10 pixels at 1080p resolution. Fingers spread, not touching. Hair that does not obscure the neck or shoulders. These are not aesthetic choices; they are engineering constraints.

The Motion Strength Trade-Off

Every model, Runway Gen-3 Alpha with its motion strength 0-10 slider, Pika 2.2 with its 0-10 slider, Kling 2.6 with its 0-100 slider, makes explicit the trade-off between motion magnitude and source-image fidelity. Higher motion strength increases the risk of subject drift, limb mutation, and identity shift. The model's structural limitation, predicting pixel displacement from a single reference frame, means that stronger motion always degrades fidelity. The composition rule is: for high-motion-strength settings, increase headroom and limb separation proportionally. A motion strength of 8 on Runway Gen-3 Alpha requires at least 50 percent frame-width headroom in the movement direction and limbs separated by at least 20 pixels. At motion strength 3, the same subject might work with 20 percent headroom and 10-pixel separation.

The practical failure that undoes all composition work is this: you spend 20 minutes composing a still with three depth layers, perfect headroom, and clear limb separation, the first 2 seconds of the clip look convincing, but by second 5 the subject has drifted, the background has warped, and the source image is barely recognisable. This is not a composition failure; it is the model's structural limitation. The temporal attention mechanism loses the conditioning signal from the source image as the clip length increases. The reliable usable clip length is 40 to 60 percent of the advertised maximum, 4 to 5 seconds on a model that claims 10 seconds. The remedy is to generate shorter clips and concatenate them with FFmpeg, accepting a seam artefact at each join, or to use Kling 2.6's start-and-end-frame feature, which gives the model a fixed endpoint and dramatically reduces drift over the full clip duration. Do not trust the advertised length. Expect to work with clips under 6 seconds for any composition that requires the subject to remain recognisable.