Best Image-to-Video Prompts for Consistent Motion

This week my For You feed kept handing me the same clip — a flat product photo that suddenly moves. The bottle turns, light sweeps the label, quick push-in on the cap. Six accounts, same supplement-and-skincare corner, all clearly built from one still. Nobody shot anything. They wrote an image to video prompt, dropped in a product photo, and let the model handle the motion.

If you're cutting TikTok Shop or affiliate assets this week — you've got the product images, but every attempt to make them move falls apart — this is for you. The ones that work aren't running better tools. They're running prompts with structure. Below is the formula I use for consistent motion, copy-and-adapt examples by use case, and the mistakes that quietly wreck your clips.

Quick version: lock the subject, describe one motion, keep the camera honest, and match the action to a 5–8 second window. Most "AI looks broken" results come from prompting for too much at once.

Why image-to-video prompts need structure

Text-to-video and image-to-video are not the same job. With a still as your input, the image already carries the composition, color, and style. Your words only have one real task: say what moves. Runway makes this explicit — its image-to-video prompting guide tells you to focus the prompt almost entirely on motion and to refer to subjects in general terms so the model knows what to animate.

That's why dumping a paragraph of scene description into image to video prompts backfires. You're re-describing things the image already shows, and the model burns its attention rebuilding the frame instead of moving it. Structure isn't about sounding cinematic. It's about handing the model fewer, clearer jobs so the output doesn't drift.

The prompt formula for consistent motion

Here's the skeleton I keep coming back to for ai video prompts built off a still:

[Subject] + [one clear motion] + [camera behavior] + [what stays fixed] + [pacing].

That's it. Google's own Veo 3.1 prompting guide treats prompting as directing rather than describing — you're telling the model what to do, not painting a picture it already has. Five parts, each doing one thing.

Subject lock

Name the subject in plain language and tell the model to keep it stable. "The sneaker stays centered and keeps its shape and laces" does more than a pile of beauty-shot adjectives. The point is to stop the model from re-inventing your product mid-clip. If your subject has fine detail — text on a label, a logo — flag it as something to preserve, not something to redraw.

Motion direction

One motion. Not three. "The cap slowly unscrews" beats "the cap unscrews while the bottle rotates and liquid pours and the camera spins." Pick the single action that carries the shot. You can stack more later, in a second variation — but the first version should test one move so you actually know what the model can hold.

Camera movement

Decide if the camera moves at all. A slow push-in or a slight orbit reads as intentional. Most of the rest reads as drift. Use real camera words — push in, pan left, locked off, slow dolly — because the models were trained on them. If you want zero camera motion, say the camera is locked and still; "no movement" is weaker than a positive instruction.

Background stability

This is the part people skip, and it's where clips fall apart. Don't pin the background, and it warps. Add a line like "background stays still and in focus" or "plain backdrop holds steady." For product and faceless content especially, a quiet background is what makes the motion look deliberate instead of glitchy.

Duration and pacing

Match the action to the clip length before you generate. Most image-to-video models output short windows — Runway's Gen-4 runs 5 or 10 seconds, Veo 3 lands around 8. OpenAI's Sora 2 prompting guide makes the same point about rhythm: a 4-second shot fits one or two beats, an 8-second one a few more. Ask for a full unbox-rotate-zoom in five seconds and every step gets rushed and the motion smears. One beat per few seconds — that's the budget.

Prompt examples by use case

These are starting points, not finished lines. Adapt the subject, then make variations. A few AI image-to-video prompt ideas I actually reach for:

Product reveal

"The perfume bottle stays centered and keeps its shape. Light slowly sweeps across the glass, then the camera pushes in gently on the cap. Background stays dark and still."

Built for the first 1.5 seconds of a TikTok Shop hook. Google DeepMind's Veo model generates native audio on its 8-second clips, so if you're on it, you can let a subtle whoosh land on the push-in instead of adding it in edit.

Character moment

"The person turns their head toward the camera and gives a small nod. Hair and clothing move naturally. Camera holds still."

Keep human motion small. Big gestures are where faces melt. A nod or a half-smile is enough to break the "it's just a photo" feeling.

Faceless loop

"Steam rises slowly and continuously from the cup. The camera is locked off and the shot starts and ends on the same frame for a seamless loop. Background stays fixed."

Loops are the easiest faceless win — pin everything, animate one element, match the start and end frame.

Before-and-after

"The messy desk stays in frame, then quickly tidies itself, items sliding into place. Camera locked, background stable."

Transformation clips need one continuous motion, not a hard cut. Let the change happen inside the shot.

Common prompt mistakes

The biggest one: over-describing the image. Runway's Gen-4 video prompting guide warns that restating details already in your input can actually reduce motion and produce unexpected results — the model gets busy and stops moving. Describe the motion, trust the image for the rest.

Second: negative prompting. "No distortion, no warping, no extra hands" tends to do less than telling the model what you do want. Positive instructions land harder.

Third: expecting the prompt to fix a bad input. A blurry, busy, or low-res starting image gets worse once it's animated — the flaws move too. Fix the still first.

And the honest one: prompts reduce drift, they don't kill it. Some clips will still warp on the second or third second no matter how clean your wording is. When that happens, regenerate with a simpler motion or stabilize in your editor. Don't burn an hour rewording a prompt that's fighting a hard generation. I'll usually run a batch of variations and keep the two that held — faster than chasing one perfect line. When I need a stack of versions off one image, I run them through AI Inspo's template remix and sort after.

Conclusion

A clean image to video prompt isn't about clever wording. It's structure: lock the subject, name one motion, keep the camera and background honest, and fit it all into a short window. That alone moves most clips from "obviously AI" to "good enough to post."

Don't aim for the perfect line. Write the formula once, make five variations, and ship the ones that hold. The prompt is a starting point — the testing is where the wins live. Go make a batch this week and see which motion your audience actually stops for.

Previous posts:

• GPT Image 2 Capabilities
• Best GPT Image 2 Prompts for Social Creatives
• GPT Image 2 Image Editor for Social Creatives