Camera Motion Prompts for AI Image-to-Video: A Practical Cheat Sheet

Most weak image-to-video clips fail for the same reason: the prompt describes the picture again instead of describing the movement. The model can already see your image. What it needs from you is how the camera moves, what moves inside the frame, and what must stay still.

This cheat sheet collects the camera moves that work reliably, the wording that tends to produce them, and the mistakes that cause warping and drift.

The three-part motion prompt

A good image-to-video prompt usually answers three questions, in this order:

1. Camera: How does the camera move? (push in, orbit, pan, static)
2. Subject: What moves in the scene? (hair in the wind, steam rising, a person turning their head)
3. Constraints: What must not change? (the logo, the face, the product shape, the background)

Example:

Slow push-in toward the cup. Steam rises gently from the coffee. The cup, label and table stay unchanged; no new objects appear.

Short and specific beats long and poetic. If you write a paragraph of adjectives, the model spends its effort on style instead of motion.

Camera moves that work well

Move	Prompt wording	Best for
Push-in	"slow push-in toward the subject"	Product hero shots, portraits, reveals
Pull-out	"slow pull-back revealing the room"	Showing context or scale
Orbit	"camera orbits 90 degrees around the subject, smooth"	Products, sculptures, characters
Pan	"slow pan left across the scene"	Landscapes, wide interiors
Tilt	"camera tilts up from the shoes to the face"	Fashion, full-body shots
Static	"locked-off static camera"	When only the subject should move
Handheld	"subtle handheld movement"	Casual, UGC-style social clips
Dolly zoom	"dolly zoom, background stretches"	Dramatic moments (use sparingly)

Two tips that apply to all of them:

• Say the speed. "Slow" and "smooth" reduce jitter. Fast moves amplify every artifact.
• Use one camera move per clip. "Push in, then orbit, then pan up" in a five-second clip almost always breaks. If you need several moves, generate several clips and cut them together.

Subject motion: keep it small

Small, physical motions look natural: blinking, breathing, fabric moving, water rippling, leaves swaying, light flickering. Large actions (running, dancing, turning around fully) are where faces and hands fall apart.

If a character needs to do something big, consider giving the model an end frame as well as a start frame, if your tool supports it. Anchoring both ends of the clip keeps identity much more stable.

Constraints: tell the model what to protect

This is the part people skip, and it matters most for commercial work. Useful constraint phrases:

• "the product shape and label stay unchanged"
• "the face stays consistent, no morphing"
• "background remains still"
• "no text, no new objects"

Not every model follows negative instructions equally well, so check the output frame by frame anyway, especially logos and hands.

Common mistakes

• Re-describing the image. "A red sneaker on a white background" adds nothing. Describe motion.
• Conflicting moves. "Static camera, slow orbit" confuses the model.
• Low-resolution input. Compression artifacts become flicker once the image moves. Start from the sharpest original you have.
• Busy source images. Lots of small details (text, patterns, crowds) tend to warp. Simpler compositions animate more cleanly.
• Judging one take. The same prompt can produce very different results. Generate three or four variations before you rewrite anything.

A quick workflow

1. Pick or generate a clean first frame (an image model helps if you need a new setting).
2. Write the three-part prompt: camera, subject, constraints.
3. Generate a few short variations with the same prompt.
4. Change one variable at a time: speed, move, or constraint wording.
5. Try the same prompt on a second video model; models differ a lot in how they handle motion.

Comparing models without juggling accounts

Step 5 is where most people give up, because every model lives behind a different account and billing page. Sora2 Hub is an AI image and video generator that puts several image and video models behind one credit balance, so you can run the same motion prompt on two or three models and compare the clips side by side.

Copy-paste starter prompts

• Product: "Slow 90-degree orbit around the bottle on a marble surface. Soft light glints on the glass. Bottle shape and label stay unchanged."
• Portrait: "Locked-off static camera. She blinks and smiles slightly; hair moves in a light breeze. Face stays consistent."
• Landscape: "Slow pan right across the valley. Clouds drift, grass sways. No people appear."
• Food: "Slow push-in toward the bowl. Steam rises, sauce glistens. Bowl and table stay still."
• Interior: "Slow pull-back from the sofa revealing the whole living room. Curtains move gently."

Start with these, change one thing at a time, and keep the camera move simple. That alone fixes most bad clips.

Comments

[baumgaerben]

Camera motion is exactly where most people under-specify. A static "cinematic shot" prompt gives you a Ken Burns pan at best — the model has no temporal intent to work with.

What actually moves the needle:

Direction + speed + easing in one clause

"Slow dolly-in from wide to medium, easing into a 3-second hold on subject's eyes" beats "camera moves closer" every time. The easing keyword (ease-in, ease-out, linear) tells the diffusion scheduler how to distribute motion across frames.

Compound moves > single-axis

"Orbit left 15° while pedastal up 0.5m, slight push-in" creates parallax depth that single-axis moves can't. Most models (Sora, Runway Gen-3, Kling, Luma) handle compound motion better when you name the axes explicitly: pan/tilt/dolly/truck/pedestal/roll.

Frame-count anchoring

"30-frame dolly-out" or "2-second truck left" forces the model to pace motion to duration. Without it, you get front-loaded movement then frozen tail frames.

Negative motion constraints

"No handheld shake, no zoom breathing, locked horizon" prevents the model from inventing micro-jitter that ruins clean plates.

One thing the cheat sheet could emphasize more: reference frame consistency. If you're doing image-to-video, the first frame is your composition lock. Any camera move that would crop the subject differently in frame 1 vs frame 30 needs explicit "maintain framing on subject" language — otherwise the model re-composes each frame independently — found it via LabAgent, site: labagent .tech