How to Describe What Moves and What Stays Still in an Image-to-Video Prompt

Load a still image into an image-to-video model and ask the subject to wave. The output shows the entire scene swaying: the tree behind the person moves, the building shimmers, the horizon slides sideways. You did not ask for that. You asked for a waving person. The problem is that you described subject motion but the model animated scene motion instead. Most image-to-video models treat the whole frame as a single animated surface by default.

Learn the distinction that determines whether your backdrop stays put or slides: the difference between subject motion and scene motion in prompting, and the exact controls each model offers to restrict animation to the subject. You will learn why default prompting almost always moves the background, which tools let you isolate the subject, and the trade-off you accept when you push for strong motion.

Image-to-video models do not understand a scene as a 3D world with independent layers. They encode your source image into a latent space, inject noise, and denoise across frames using temporal attention layers that predict pixel displacement from a single reference frame. They have no persistent object model. When you write “a woman waves”, the model must decide which pixels belong to the woman and which belong to the backdrop, and it often guesses wrong. Ambiguous depth, similar textures, or a subject occupying less than half the frame all worsen the guess.

The result is background sliding: the backdrop moves as if painted on a separate layer drifting out of sync with the foreground. The model's temporal attention loses the conditioning signal from the source image over the clip duration, and it starts interpolating motion patterns from its training distribution rather than anchoring to the static background pixels. Longer clips make this worse. Motion coherence degrades visibly after 4 to 6 seconds on most models, even when the advertised maximum is 10 seconds.

The structural cause is the single-reference-frame limitation. The model has exactly one still image to work from. It cannot reason about parallax, occlusion, or the fact that the building behind the subject should not move when the person does. What the model does infer is that something in the image has moved, and since it lacks a segmentation model to restrict that motion to a region, it distributes displacement across the entire frame.

Subject motion means the object in the foreground changes position, rotates, or articulates while the backdrop remains static relative to the frame.

Scene motion means the virtual camera moves through the scene, creating parallax changes in the backdrop and shifting the perspective on everything in the frame.

Your prompt must signal which of these you want. A prompt that says only “a cyclist pedals down a road” leaves both options open. The model defaults to scene motion because whole-frame displacement produces fewer visible artefacts during training. Shifting every pixel a little is easier than keeping a region still while another region deforms.

The failure mode you encounter when you ask for subject motion without constraining the background is subject drift. The cyclist's body morphs because the model tries to simulate both the pedaling motion and a background slide simultaneously. The two motions interfere, and the temporal attention mechanism resolves the conflict by blending the subject with the moving background pixels. Source-image fidelity drops. By the fourth second of a 5-second clip, the cyclist's face may have changed completely.

You have three strategies for constraining motion to the subject. They depend entirely on which model you are using. Not every model supports every strategy.

Region-Mask Support (Pika 2.2)

Pika 2.2 is the only commercial model with dedicated region-mask support for content isolation. In the Pika interface, paint a mask over the area you want to animate, the subject, and write a prompt that only affects those pixels. The rest of the image remains completely static. This is the most reliable method for selective animation because it removes the model's ambiguity about which pixels belong to the subject.

Motion Brush (Runway Gen-4, Kling 2.6, Seedance 1.0, Hailuo AI)

Runway Gen-4, Kling 2.6, Seedance 1.0, and Hailuo AI each offer a Motion Brush tool. Paint the direction and intensity of motion onto specific regions of the image. Unlike a region mask, which simply freezes the unmasked area, the Motion Brush tells the model where to apply motion and where to leave the pixels unchanged. The difference matters: a region mask is binary (move or do not move), while a Motion Brush can specify that the subject moves left while the backdrop stays still, reducing the chance of background sliding even when the mask boundary is not perfectly defined.

Negative Prompting and Motion Strength Adjustment (All Models)

On models without region masks or Motion Brushes, LTX-Video 0.9.5, Stable Video Diffusion 1.1, Wan 2.2, Luma Dream Machine, use negative prompting to discourage background motion. A prompt like “a dancer spins, the background is completely still, no camera movement, no background motion, no parallax” reduces but does not eliminate background sliding. These models interpret negative prompts as weak constraints, and the backdrop still drifts on any clip longer than 3 seconds. Lower the motion strength or motion bucket ID (on SVD, 1 to 255) to reduce the overall motion magnitude, which in turn reduces the amplitude of the background drift. The trade-off: the subject's motion also weakens. You cannot have strong subject motion with a completely static backdrop on models that lack region-level controls.

Start with a photograph of a woman standing in a field, arms at her sides. You want her to raise her right hand in a wave while the grass and sky stay frozen. Here is how the prompt structure changes between a model with region-mask support and one without.

Prompt on Pika 2.2 (region-mask enabled): Paint a mask over the woman's arm and torso. Write the prompt “the woman raises her right hand and waves, slow motion, arm rotates at the shoulder, hand opens”. Leave the background unpainted. The model animates only the masked region. The grass does not sway. The sky does not shimmer. The clip runs 4 seconds and the backdrop is pixel-identical to the source image across every frame.

Prompt on Luma Dream Machine (no region mask): Write “a woman standing in a field raises her right hand and waves, the background is completely still, no camera move, no wind, no grass movement, no sky motion, only the woman moves”. The first two seconds of the 5-second clip show the woman's arm moving correctly, but by the third second the background grass has started to drift left. By the fifth second the horizon line has shifted 15 pixels. You get subject motion with background sliding. It is usable for b-roll where the audience does not scrutinise the edges, but not for a hero shot.

Prompt on Stable Video Diffusion 1.1 (negative prompt, motion bucket 80): Set the motion bucket ID to 80 (moderate motion). Write “a woman raises her right hand and waves, static background, no wind, no movement in the field, only the arm moves”. The model outputs 25 frames at 6 fps. The arm motion is present but weak. The backdrop holds still for approximately 2 seconds, equivalent to 12 frames, before the grass begins to pulse. By frame 20 the woman's face has started to distort. This is the motion magnitude versus source fidelity trade-off in action: the lower motion bucket reduced the background drift but also reduced the subject animation to a visible but unsatisfying level.

Stronger subject motion always increases the risk of subject drift and background sliding. This is a structural consequence of predicting pixel displacement from a single reference frame. The model cannot know that the woman's arm should move 40 pixels while the building behind her should move 0 pixels. It has only one image as its anchor, and every frame after the first is generated from the previous frame's output, accumulating error.

The practical range you can expect:

• Low motion (arm moves 10 to 20 pixels): Background holds still on most models through 4 seconds. Subject articulation is minimal but visible. Source-image fidelity at the clip end is good.
• Moderate motion (arm moves 30 to 60 pixels): Background begins to drift after 2 to 3 seconds on models without region masks. Subject face may change subtly. Fidelity is fair to good.
• High motion (arm moves 80+ pixels): Background slides noticeably by the second second. Subject drift is likely, face may change, limbs may warp. Fidelity is poor. Only models with region masks (Pika 2.2) or Motion Brush (Runway Gen-4, Kling 2.6) can maintain fidelity at this level.

If you need strong subject motion on a model without region-level controls, accept that you will need to post-process the clip: extract frames with FFmpeg, composite the static backdrop over the moving subject using a video editor, or accept the background drift as a stylistic choice.

The words you choose matter less than the structure you impose. Image-to-video models are text-conditioned on the motion axis, but their text conditioning is weak. Many models effectively ignore the text prompt for motion when the image conditioning is strong. Write prompts that leave no ambiguity about what moves.

Follow these patterns:

• State what moves first, what does not move second. “The dog runs across the grass. The grass is completely still. The camera does not move. The fence behind the dog does not move.”
• Use negation explicitly. “No background movement, no camera movement, no wind, no swaying, no parallax.” The model does not understand the concept of a static backdrop unless you tell it what not to animate.
• Name the axis of motion. “The dog runs from left to right. The dog's legs cycle. The dog's body stays centered in the frame.” This tells the model that the subject translates laterally while its body articulates, and that the camera does not track the movement.
• Keep motion magnitude phrases vague. “Slow wave” or “gentle turn” are better than “dramatic spin” because lower-magnitude phrasing reduces the model's tendency to over-animate the backdrop as a side effect.
• Avoid words that imply camera movement. “Approach”, “pull back”, “pan”, “zoom”, “track”, “circle”, “dolly” all trigger camera-motion generation in models like Luma Dream Machine, which is trained to produce smooth camera moves as a primary feature.

No current model except Pika 2.2 has a dedicated region mask for content isolation. The table below summarises the regional controls available on each model that supports selective animation in some form. The absence of a control means the model cannot reliably animate only the subject.

Models With Region-Level Controls

Pika 2.2: Region-mask support for content isolation. Paint the area to animate. The rest of the image is perfectly static. No other model offers this.

Runway Gen-4: Motion Brush for regional motion direction and amplitude. Paint where motion should occur and how strong it should be. The brush does not freeze unmasked areas, it reduces their motion magnitude. The backdrop may still drift slightly at high motion strengths.

Kling 2.6: Motion Brush for regional motion. Same principle as Runway. Paint the motion direction and intensity on one region and leave others untouched. Background holds still better than Runway at moderate motion levels.

Seedance 1.0: Motion Brush with regional direction and intensity. Effective for single-subject isolation when the subject occupies a clear portion of the frame. Background drift is minimal at low-to-moderate motion but reappears at high motion.

Hailuo AI: Motion Brush with regional direction and intensity. Similar performance to Seedance.

Models Without Region-Level Controls

Luma Dream Machine: No Motion Brush, no region masks. Relies entirely on the text prompt to constrain motion. Background sliding is common on any clip above 3 seconds. Use the Camera Motion panel to set “no camera movement” as an explicit control, which helps slightly.

Wan 2.2: No Motion Brush, no region masks. Start and end frame support (keyframe interpolation) can help if you set the end frame to be identical to the start frame for the background area, but the model does not accept partial end frames; the entire frame must be specified.

Stable Video Diffusion 1.1 and LTX-Video 0.9.5: No regional controls. Motion bucket ID (SVD) or frame count (LTX-Video) are the only ways to reduce overall motion magnitude, which reduces background drift as a side effect.

Veo 3, Sora, and Google Veo 2: No regional controls. Motion is driven entirely by the text prompt and the model's internal interpretation. Background sliding is the expected default behaviour. Veo 3 can produce 4-, 6-, or 8-second native clips, but motion coherence drops off after 4 seconds on any subject that occupies less than 60% of the frame.

Compare these two prompt sets for the same source image: a man sitting at a desk, hands on a keyboard. You want him to type. You do not want the desk, the bookshelf behind him, or the window to move.

Failed prompt (animates the background): “A man types on a keyboard. His fingers move quickly across the keys. The scene is an office with bookshelves and a window.”

The model generates a 5-second clip. The man's fingers blur and mutate; the model has no skeletal model and guesses finger positions. The bookshelf slides left, the window light changes, and the desk appears to wobble. By the third second, the man's face has shifted subtly. This is the default behaviour of every model listed above when given an unconstrained prompt.

Working prompt (constrains motion to the subject): “A man sits at a desk. Only his fingers move, they press keys on a keyboard. The man's body, head, and face do not move. The desk does not move. The bookshelf behind him stays completely still. The window does not change. No camera movement. No wind. No animation outside the fingers. The fingers move slowly and precisely.”

On Pika 2.2 with a region mask painted over the keyboard area, this prompt produces a 4-second clip where only the fingers move. On Runway Gen-4 with Motion Brush set to low intensity on the keyboard region, the fingers move while the background drifts less than 5 pixels. On Luma Dream Machine without regional controls, the background drifts after 2.5 seconds, but the subject's face does not change because the prompt's explicit negation of face movement reduced the model's tendency to generate facial micro-expressions.

The single most important thing to understand about image-to-video subject motion versus scene motion prompting is that the model has exactly one image to work from. It cannot see the back of the subject. It cannot reason about depth. It cannot know that the subject's arm is in front of the backdrop and therefore should occlude it when it moves. Every frame after the first is a prediction generated from the previous frame's output, and each prediction step increases the drift from the source image.

This is why region masks and Motion Brushes work. They provide the model with a spatial constraint that the single reference image cannot supply. Without them, the model treats the entire frame as a single deformable surface, and your subject motion becomes scene motion because the model has no way to distinguish foreground from background.

The practical consequence: if you need reliable subject-only animation, use Pika 2.2 with region masks. If you cannot use Pika, accept that you will get some background drift and plan your shot composition accordingly. Keep the subject large in the frame, occupying at least 50% of the pixels, and keep the backdrop simple: solid colours, shallow depth of field, minimal textures. A busy background with trees, water, or patterned wallpaper guarantees visible background sliding on any model without regional controls.