Image-to-video motion prompts
Turning a still AI image into a clip is mostly about describing motion clearly and specifically. Tools like Runway, Kling and Stable Video Diffusion each read motion language a little differently, but the core vocabulary — camera moves and subject motion — is shared. This builder pairs a sensible camera move with subject motion that suits your image type and formats it for your platform.
The two types of motion
Camera motion moves the viewpoint — the scene content stays put while the frame shifts around it. Camera moves are highly reliable across all i2v platforms because they affect the whole frame uniformly rather than requiring the model to reason about individual objects animating:
| Camera move | Effect |
|---|---|
| Dolly in (push in) | Slow approach — creates intimacy and focus |
| Dolly out (pull out) | Reveals surroundings — classic “zoom out” reveal |
| Pan left / right | Horizontal sweep across a wide scene |
| Tilt up / down | Vertical scan — great for tall subjects or landscapes |
| Orbit (arc shot) | Camera circles the subject — works best on single subjects |
| Handheld / float | Slight camera shake — adds realism and life |
Subject motion animates elements inside the frame. This is harder for the model because it must infer physics and anatomy from a single still image:
| Subject element | Suggested motion |
|---|---|
| Water, ocean | Gentle waves, ripple, flowing |
| Hair, cloth, leaves | Gently swaying, blowing in the breeze |
| Clouds, smoke, mist | Slowly drifting, dissipating |
| Fire, candle flame | Flickering, dancing |
| Portraits | Subtle blink, slight head turn — avoid large motions |
Platform differences
Runway Gen-3 / Gen-3 Alpha reads descriptive natural-language motion prompts well. Combine camera and subject motion in one sentence: for example, “slow dolly in, the subject’s hair gently blowing in the breeze.”
Kling has a motion brush feature that lets you paint regions and assign motion vectors directly, which gives finer control than text prompts alone. The text motion prompt still guides the overall behaviour.
Stable Video Diffusion is primarily controlled by a motion bucket / strength slider rather than text. The text you write here serves as guidance for tuning that slider — higher strength values animate more aggressively but risk frame distortion.
Practical tips for better clips
- One camera move, one subject motion. Stacking three or four motions (pan + zoom + spin + flowing water) produces warping artifacts because the model cannot simultaneously satisfy all constraints.
- Animate things that move in real life. Water, smoke, hair, cloth, and clouds produce convincing animation because the model has seen many examples of how they move. Rigid objects (furniture, architecture, vehicles) rarely animate convincingly without camera movement to carry them.
- Use subtle camera moves for faces and portraits. A slow dolly-in with a slight head-turn prompt works far better than asking for a full facial expression change — the latter often produces warped or uncanny results.
- For SVD, the slider rules. Use very low motion strength (3-5 out of 25) for still, photographic subjects. Reserve high strength for abstract or dynamic scenes where distortion is less visible.