Image to video prompts: direct the motion, not the picture

The start frame is the shot

This is lesson 3 of Make a short film with AI. You have a shot list from lesson 1 and approved, consistent start frames from lesson 2. Now you animate them.

Image-to-video changes what a prompt is for. In text-to-video, the prompt has to describe everything: subject, setting, light, framing and motion. In image-to-video, the first four are already decided by the picture you uploaded. The model starts from that exact frame. Your prompt only has to say what happens next.

Start frame

→

twirl, lift, glance, push-in

Image promptNano Banana Pro · 38 creditsCopy

A golden retriever wearing a velvet dinner jacket and bow tie, seated at a Michelin-star restaurant, delicately twirling spaghetti with a silver fork, candlelight, white tablecloth, sommelier blurred in the background, editorial food photography, 85mm, shallow depth of field

Motion promptKling 3.0 · 5 s clip · 105 creditsCopy

The dog twirls the spaghetti on the fork, lifts it and glances at the camera with quiet satisfaction, candle flames flicker, subtle camera push-in, cinematic

Compare the two prompts above. The image prompt is long because it builds the scene: the velvet jacket, the candlelight, the blurred sommelier, the 85mm lens. The motion prompt does not mention any of that. It describes three beats of one continuous action (twirl, lift, glance), one environmental detail (the candles) and one camera move (a subtle push-in). The clip delivers exactly that: the dog twirls, raises the fork and ends looking into the lens, while the frame tightens slowly.

Anatomy of an image to video prompt

A good motion prompt has the same parts every time, in roughly this order: what the subject does, how it ends, what else moves, and what the camera does.

1. Subject and action first. Refer to the subject simply ("the dog", "the bottle"); the model can see it. Spend your words on the verb.
2. An end state. "Glances at the camera" gives the clip somewhere to arrive. Without one, motion tends to wander or loop.
3. Environment motion. One secondary movement (flames, haze, hair, water) makes a frame feel alive without competing with the action.
4. One camera move, with a speed. "Subtle", "slow" and "gentle" are instructions, not decoration.
5. A short style tag, optional. The image already carries the style; a single word like "cinematic" or "commercial" nudges the pacing.

Camera movement prompts: a vocabulary for AI video

The official prompting guides for Veo, Kling and Runway all use film-crew language. A named move like "dolly in" leaves the model far less to guess than "the camera goes closer". Each move also carries meaning. Pick it for what it tells the audience, not because it looks impressive.

Two practical notes. First, "zoom" is not the same as "dolly": a zoom magnifies without parallax and looks flatter, so ask for a dolly or push-in when you want depth. Second, the more complex the subject's action, the simpler the camera should be. Locked-off is underused in AI video and is often the best choice when a character has to do something precise.

Check that the camera did what you asked

Naming the move is necessary, not sufficient. We gave PixVerse V6 one perfume start frame (the same one behind the dolly-in clip in the next section) twice, once asking for a slow orbit and once for a locked-off camera, at 57 credits per clip:

Both clips push in on the bottle. The "locked-off" clip ignores the instruction completely: the bottle grows steadily in the frame for five seconds, and the light beam we asked to sweep across the glass barely moves. The "orbit" is mostly the same push-in with a slight turn, enough that by the last frame the label has rotated toward the edge of the bottle. In this test the low-cost model fell back on a slow push-in whatever we asked for. So watch every clip for the move you requested, and when the move matters, use a stronger model or pin the ending down with a last frame (next section).

Subject action, pacing and end states

A five-second clip has room for one camera move and one action. Ask for more and the model will either skip instructions or compress them into unnatural, sped-up motion. Pacing words are how you tell it to use the full five seconds:

• Speed: slowly, gently, gradually, steadily, in slow motion, suddenly.
• Sequence: "then", "before", "as". "Lifts the cup, then pauses" reads as two beats of one action.
• End state: "comes to rest", "settles", "looks up at the camera", "exits frame left". The last frame is your next cut point.

Start frame

→

slow dolly-in, light sweep

Image promptNano Banana Pro · 38 creditsCopy

Luxury perfume bottle of faceted amber glass with a brass cap, standing on wet black volcanic stone, water droplets, a single shaft of warm light cutting through haze, premium beauty campaign photography

Motion promptKling 2.5 Turbo · 5 s clip · 88 creditsCopy

Slow dolly-in on the perfume bottle as condensation droplets form and slide down the amber glass, a beam of light sweeps across the facets, haze drifting, luxury commercial

The perfume clip shows how far a little motion goes. The camera dollies in slowly, the beam of light shifts and brightens through the haze, and the bottle itself does nothing. Because the product stays still and the move is slow, the "Aurum & Amber" label is still legible in the last frame. Products and text survive motion when the camera, not the object, does the moving.

Start and end frame AI video

Sometimes you need a shot to finish on an exact composition: the character in the doorway, the product centered for the end card, a match cut into the next shot. For that, give the model both ends. In first and last frame mode, you upload a start frame and an end frame and the model generates the motion between them. Veo 3.1 is one model that supports it, and Kling 2.5 Turbo, used below, is another; check the model page for which durations each allows with an end frame.

1. 01

   Make the end frame from the start frame

   Use an edit or reference model to create the end frame from the start frame, changing only what should change: pose, camera position or time of day. Same character, same light, same lens.

2. 02

   Keep the gap plausible

   The model has to invent a physically believable path between the two images. A head turn or a camera arc works; a character teleporting across the room produces morphing.

3. 03

   Describe the transition

   The prompt names the move that connects the frames: "slow orbit to the right as she turns to face the camera". Do not re-describe either image.

Here is one from our bear film. The start and end frames are the ride and the arrival from lesson 2, and the prompt describes the move between them:

Start frame

End frame

lands on the end frame

Motion promptKling 2.5 Turbo · 5.1 s clip · 88 creditsCopy

The bear slows the motorbike to a stop beside the lighthouse, climbs off and looks out to sea, slow tracking shot that settles into a locked-off wide

Kling 2.5 Turbo starts on the side-on ride, swings round behind the bear for a rear view along the coast with the sun ahead, and finishes on the end frame's composition with the lighthouse above the road. The two stills put the sea on opposite sides of the frame, and the camera move is how the model reconciled them. Notice what it did not do: the prompt says the bear "climbs off and looks out to sea", but the end frame shows him still seated, so the clip ends with him on the bike. When the prompt and the end frame disagree, expect the frame to win. Fix the frame, not the words.

First and last frames are also the cleanest way to build longer sequences: the end frame of shot 4 becomes the start frame of shot 5. In PromptLab, the Playground's Start End Frame tool lists only models that accept both frames, so you do not have to check each model's inputs yourself.

Common failure modes and how to fix them

Almost every bad clip is one of a handful of problems, and most have the same root cause: too much happening in too little time. Two of our own text-to-video renders show it.

The tea party prompt asks for tea that "flows upward". The model poured it down into the cup, the way tea behaves in nearly every video it has seen, and its slow push-in ends with the hare's head out of the top of the frame because no end framing was given. The octopus prompt asks for seven arms doing three jobs at once. The clip shows two pitchers pouring into a row of cups before the camera drifts down to the latte art; the grinding and wiping never appear. At 305 credits, it was also one of the most expensive clips in this course.

Start frame

→

slow rotation, 57 credits

Image promptFlux 2 Pro · 8 creditsCopy

Tortoiseshell aviator sunglasses floating above rippled desert sand at dusk, the lenses reflecting a pink and orange sky, long shadow, luxury eyewear campaign, hyper-detailed

Motion promptPixVerse V6 · 5 s clip · 57 creditsCopy

The sunglasses rotate slowly in mid-air, fine sand drifting below them, the sky colours shifting from pink to orange, product commercial

Rotation is the classic stress test for product geometry. This full turn ran on PixVerse V6 for 57 credits and holds together well at playback speed, with sand puffing up below. Step through it frame by frame and the temples and hinges are less precise mid-turn than in the front view. That is the trade-off of a cheap model, and the reason to test motion cheaply first and render the final on a stronger one.

Image to video prompt examples you can copy

Starting points for common shots. Each assumes the start frame already shows the subject, setting and light, so the prompt only covers motion. Adapt the bracketed parts.

• Product hero: "Slow dolly-in on the [product], a soft highlight glides across its surface, [haze / steam / droplets] drifting, ends centered on the label, luxury commercial"
• Character reaction: "[Character] slowly looks up from [object] toward the camera, a slight smile forms, hair moves in a light breeze, subtle push-in, cinematic"
• Establishing shot: "Slow crane up revealing [landscape] beyond, clouds drift across the sky, birds cross the frame, locked horizon, epic"
• Movement: "[Character] rides [vehicle] steadily along [road], camera tracking alongside at the same speed, [hair / fur / scarf] rippling in the wind"
• Food: "[Sauce / honey / syrup] pours slowly onto [dish] and spreads, steam rises, locked-off macro shot, soft morning light"
• Fashion: "The model turns slowly from profile to face the camera, the fabric of the [garment] sways, slow orbit to the left, editorial"
• Ending shot: "[Characters] stand still watching [horizon], the sun sinks, very slow pull-back to a wide silhouette, then holds"

A repeatable workflow, and what it costs

The image-to-video clips in this lesson ran on 3 different models at very different prices:

That spread suggests a working method. Test a motion prompt on an inexpensive model first: if the action reads correctly at 57 credits, the prompt is probably right, and you can spend the premium credits on fidelity rather than on finding out the prompt was wrong. One caveat from the perfume test: if a cheap model ignores a camera instruction, the model may be the limit rather than the prompt, so try the same prompt on one stronger model before rewriting it. Then run the final take on two or three stronger models side by side and keep the best. The PromptLab Benchmark does this in one step with the same start frame and prompt on up to four models, and the catalog lists which models support image-to-video, first and last frames, and at what price.

Which model to use for which shot is the subject of the next lesson: how to choose the best AI video model for faces, products, landscapes and action, with side-by-side tests.

Questions people ask