AI Video No Motion: Why Does My AI Video Barely Move?

TL;DR: An AI video with no motion is usually a prompt that described a scene instead of an action. But not always. Of the eight common causes, five are wording problems a rewrite fixes, two are parameter problems where rewriting is wasted effort, and some are limits of the model you are on.

Why does my AI video barely move?

Because you probably wrote a photograph. Most static AI video comes from a prompt full of nouns and adjectives with no verb attached to the subject, and the model rendered exactly that: a beautiful frame, held.

That is the common case and it is fixable in one edit. It is not the only case, which is why "add more action words" fails so often as universal advice. Sort the problem before you rewrite. Our post on image prompt troubleshooting made the same argument for still images, and video splits the same three ways.

Prompt problems. The model did what you wrote, and what you wrote had no movement in it, or had movement cancelled out by something else in the same sentence. Rewording fixes these.

Parameter problems. Your wording is fine. A field in the request is set wrong, or the duration is too short for the motion to complete. Rewriting the sentence changes nothing here, and it is the most common wasted effort in this category.

Capability limits. The model produces low motion on this subject, in this style, at this length, and it will keep doing that. Vendors are quieter about this than about image limits, but it is real and no phrasing reaches it.

Which of these eight causes is yours?

Find the sentence that sounds like what you said when the clip finished rendering.

Does your model even have a motion strength setting?

Usually not, and this is where most advice goes wrong. Motion strength dials exist under different names, on a minority of surfaces, and they do not port between vendors. Checking takes a minute and saves an afternoon.

The details behind that table:

Vidu is the one vendor here with a real motion-amount field, and it documents the field as inert on its current models. movement_amplitude is an optional string, defaulting to auto, accepting auto, small, medium and large, described as the movement amplitude of objects in the frame. Its text-to-video page then adds: "This parameter does not take effect when using the q2 & q3 model" (Vidu text-to-video, read 29 August 2026). On Q3, the dial is there and turning it does nothing.

Seedance publishes a camera lock rather than a motion amount. camera_fixed is a boolean defaulting to false, and BytePlus describes true as "Fix the camera. ModelArk will append the fixed camera instruction to the user's prompt, but the actual result is not guaranteed." Its supported-models list names Dreamina Seedance 1.5 pro, 1.0 pro and 1.0 pro fast, and not the 2.x models (BytePlus ModelArk video generation API, read 29 August 2026). If you inherited a request body with camera_fixed: true on a supported model, that alone explains a locked-off frame.

Runway exposes no motion field on any of its image-to-video branches. Its published schema carries thirteen model options; the Gen-4.5 branch accepts model, promptImage, promptText, ratio, duration, seed, outputFormat, proresProfile and contentModeration. No motion, no amplitude, no strength. Two branches describe the text field as "An optional text prompt describing motion or changes in the output video" (Runway API schema, read 29 August 2026). On Runway, motion is a writing job.

Grok Imagine Video 1.5 takes model, prompt, duration, aspect ratio and resolution. Duration is documented as "The allowed range is 1–15 seconds" and there is no motion parameter of any kind (xAI video generation, read 29 August 2026).

Luma Ray 3.2 is the same story with better prompt advice. Its video options are resolution, duration, hdr, anchor frames and keyframes, with no motion field. Its prompt parameter is documented as "A text description of the video, 1–6,000 characters. Be specific about subject, motion, camera movement, lighting, and pacing." (Luma video generation, read 29 August 2026).

Kling 3.0 Omni's text-to-video request body is prompt plus settings for multi-shot, audio, resolution, aspect ratio and duration (Kling 3.0 Omni text-to-video, read 29 August 2026). No motion parameter. What Kling does ship is region-level control in the app: the Motion Brush, which the quickstart describes as letting you "Add motion to up to 6 elements at once, or use the Static Brush to keep your background perfectly still" (Kling Motion Brush guide, read 29 August 2026).

The five causes a rewrite really does fix

These are the ones where the sentence is the problem. Every pair below is before and after on one intent.

Cause 1: you described a scene, not an action

The single most common cause. Google's video generation prompt guide gives action its own section and says "Action brings the subject to life, describes movements, interactions, and subtle expressions" (Gemini Enterprise video generation prompt guide, read 29 August 2026). A prompt with no such sentence is a still-image prompt pointed at a video model.

The fix is mechanical: give the subject a verb, then say how the verb progresses across the clip.

Before: A cozy coffee shop on a rainy afternoon, warm light, wooden tables,
        steam rising from a cup, moody atmosphere, cinematic.
After:  A barista slides a steaming cup across the counter and turns back to
        the machine, wiping her hands on her apron as rain streaks the window
        behind her.

Before: A futuristic city at night, neon signs, wet streets, flying cars,
        blade runner aesthetic, highly detailed.
After:  A flying car banks left between two neon towers and drops toward the
        wet street below, its lights sweeping across puddles as it descends.

Before: A golden retriever in a field of wildflowers at golden hour, shallow
        depth of field, beautiful bokeh.
After:  A golden retriever bounds through waist-high wildflowers toward the
        camera, ears lifting on each stride, then skids to a stop and shakes
        pollen from its coat.

Cause 2: you told the subject to hold still

This hides in plain sight because it looks like an action. "Standing", "sitting", "posing", "gazing" and "resting" are verbs of stillness. Google's guide lists "standing still, sitting" among its examples of basic movements, which is the point: those are instructions and the model follows them.

If your intended motion is elsewhere in the frame, say so rather than assuming the model will infer it.

Before: A chef standing at a wooden counter with vegetables and a knife.
After:  A chef dices a red pepper in fast, even strokes, sweeping the pieces
        into a bowl with the flat of the blade.

Before: A woman sitting by a window, looking out at the storm.
After:  A woman turns from the window as lightning flares behind her, her
        reflection sliding across the glass as she moves.

Before: An astronaut posing in front of a landing craft on Mars.
After:  An astronaut plants a sample tube in the red soil, straightens up, and
        walks back toward the landing craft as dust drifts past her boots.

Cause 5: the camera moves and nothing in it does

Camera language is easy to write and models follow it well, which is the trap. A slow dolly-in over a subject you never gave a verb produces a moving frame around a frozen picture. That reads as a technical failure and is a faithful render.

Kling documents a sharper version of this. When a motion trajectory is drawn on something that cannot move in the physical world, the guide says: "Since it's difficult for a building to move horizontally in the real world, the motion trajectory for the element will be interpreted as a camera movement path." Your requested subject motion is converted into camera motion. Same source and date as above.

The rule: every camera instruction needs a subject instruction beside it. If you want the camera vocabulary itself, our camera movement reference covers thirty terms and what each one actually does.

Before: Slow dolly in on a man at a desk in a dark office.
After:  Slow dolly in on a man at a desk as he stops typing, leans back, and
        rubs his eyes.

Before: Drone shot orbiting a lighthouse on a cliff at sunset.
After:  Drone shot orbiting a lighthouse as its beam sweeps out over the water
        and waves break against the rocks below.

Before: Handheld push through a crowded market.
After:  Handheld push through a crowded market as vendors reach across their
        stalls and shoppers step aside to let the camera pass.

Cause 6: your prompt contradicts itself

"Serene", "calm", "still", "tranquil" and "quiet" are mood words that also describe motion, and models weigh them. Pair "a serene, still lake" with "a dynamic action shot" and you have written two instructions that cancel.

Most current video models publish no negative prompt field, so "no static shots" is not read as an exclusion. It puts "static" and "shot" into the positive description. Say what should move instead.

Before: A calm, serene, peaceful mountain lake at dawn, dynamic and full of
        energy, no static shots.
After:  Mist rolls off a mountain lake at dawn as a heron lifts from the
        shallows and beats slowly across the water.

Before: A quiet, minimal, still-life composition of a workshop, with the tools
        moving energetically.
After:  A hand reaches into the workshop frame, lifts a chisel from the bench,
        and taps it twice against a block of wood.

Before: A tranquil, motionless forest scene with a fast-paced chase through it.
After:  A deer breaks from the treeline at a sprint, weaving between trunks as
        leaves scatter behind it.

Cause 7: too many actions, so none of them complete

A clip is a budget. Ask for six actions in five seconds and the model spends under a second on each, which reads as twitching rather than moving. This is the cause most often misdiagnosed as "the model is weak", and the fix is subtraction. One clear action, or two that chain naturally, beats a list.

Before: A man walks into the kitchen, opens the fridge, pours a glass of milk,
        checks his phone, sighs, and sits down at the table.
After:  A man opens the fridge, pours a glass of milk, and sets the carton back
        on the shelf.

Before: A dancer spins, jumps, lands, turns to the camera, smiles, and walks
        off stage as the lights change.
After:  A dancer spins twice and lands facing the camera, arms dropping to her
        sides as she catches her breath.

Before: Waves crash, seagulls fly, a boat passes, a child runs, a kite lifts,
        and the sun sets.
After:  A child runs along the waterline trailing a kite that lifts and steadies
        above the waves.

Which causes are parameter problems, not prompt problems?

Two of the eight. On these, rewriting the sentence is the wrong move and will feel like the model is ignoring you.

Cause 4: a motion parameter left where you found it

Covered in the table above; the practical version is short. Open the request body or the advanced panel and read it. If there is a motion amount field, note its value and whether the vendor says it applies to your model. If there is a camera lock, note whether it is on.

Two traps worth knowing. On Vidu Q3, movement_amplitude accepts values and does nothing, so neither a small setting nor a large one is your problem. On Seedance, camera_fixed applies only to the 1.5 pro and 1.0 pro models, so a value copied from an older recipe is silently inert.

Do not carry one vendor's field name to another. There is no movement_amplitude on Runway and no camera_fixed on Kling. A parameter that does not exist on the surface you are using is not a control, it is a typo.

Cause 8: the clip is too short for the motion to happen

A walk across a room takes about four seconds. Ask for it in a five-second clip that also has to establish the room and you get a step and a half, which looks like almost nothing happened.

Duration is a request field, not a wording problem. Published ranges differ per model: Grok Imagine Video 1.5 allows 1 to 15 seconds, Kling 3.0 documents 3 to 15, Vidu Q3 Pro documents 1 to 16. Our duration reference by model has the full matrix.

Two ways out, both of which work:

Before (5s): A woman walks the length of a hotel corridor, stops at a door,
             takes out a key card, and lets herself in.
After  (5s): A woman stops at a hotel door and taps a key card against the
             reader, pushing the handle down as the light turns green.

Before (5s): A time-lapse of a flower opening from bud to full bloom.
After (12s): A time-lapse of a flower opening from bud to full bloom, petals
             unfurling one layer at a time.

What if the still image gives the model nothing to move?

This is cause 3, and it sits between the buckets honestly enough to deserve its own answer.

An image-to-video model reads one frame. That frame carries no velocity, no intent and no next position, so the prompt is the only place motion can come from. Runway's schema says so on two of its branches: the text field is "An optional text prompt describing motion or changes in the output video". Same source and date as above.

The trap is writing a caption. If your text restates what the picture already shows, you have given the model a redundant description and no instruction, and a near-frozen clip is the correct output.

This is broader than it looks. On grok-imagine-video-1.5, xAI documents that text-to-video "uses text-to-image then image-to-video under the hood: the model generates a first frame from your prompt, then animates it" (same source and date). Even when you never uploaded an image, you may be animating one.

Kling's own Motion Brush guidance points the same direction: "be sure to add a text prompt that matches the movement of the area/element", in an element plus motion format. Its worked example takes the intended result of a cat jumping over a plate, succeeds with a prompt about a cat jumping over the bowl in front of it, and fails with a prompt about a cat walking forward.

Our still image to video handoff guide goes deeper on this. The short version, as pairs:

Before: A woman in a red coat standing on a bridge at night.
After:  She turns her head to follow a passing train, coat lifting in the
        draught as it goes by.

Before: A bowl of ramen on a wooden table, steam rising.
After:  Chopsticks lift a tangle of noodles clear of the broth, steam bending
        toward the camera as they rise.

Before: A vintage car parked outside a diner.
After:  The car pulls out of the parking space and swings left onto the road,
        neon reflections sliding across its hood.

Two rules make image-to-video prompts work: write only what is not already visible, and start the sentence with the thing that moves.

What if none of it works?

Then you have found a capability limit, and the honest answer is to stop rewriting.

A model that produces low motion on a given subject may simply do that. This is the conclusion our image prompt troubleshooting post reached for stills, and it holds here: after three genuinely different attempts, rewriting is not diagnosis, it is hope.

Some of this is documented: Kling tells you physically impossible motion becomes camera motion, and BytePlus tells you camera_fixed results are "not guaranteed". Most of it is not, because low motion is a quality characteristic rather than a hard limit, and vendors do not publish those.

What to do instead:

1. Change one variable, not the whole prompt. Same prompt, different model. If the second model moves, it was the first model.
2. Change the shot, not the wording. Wide static subjects resist; close subjects with a body part in motion do not.
3. Anchor the motion instead of describing it. Keyframes, a start and end frame, or Kling's Motion Brush give the model positions to interpolate rather than a sentence to interpret.
4. Check you are not fighting a different problem. If the clip has motion but the motion is wrong, that is a different diagnosis, and our post on morphing and warping covers it.

We will not tell you a rewrite hits some percentage of the time. Nobody has that number, and the ones you see quoted are invented.

Six more before-and-after pairs

A bank to work from, covering the patterns that come up most in product and social work.

Before: A product shot of a watch on a marble surface, luxury, cinematic
        lighting.
After:  A hand sets the watch down on marble and withdraws, the second hand
        sweeping on as the strap settles flat.

Before: An office team collaborating in a bright modern workspace.
After:  A woman writes on the glass wall while two colleagues turn their chairs
        toward her, one leaning forward to point at a line she has drawn.

Before: A cat by a sunny window, peaceful and still.
After:  A cat stretches out along the windowsill, tail flicking once, then
        flattens its ears as a bird lands outside.

Before: Rain on a city street at night, moody, atmospheric, cinematic.
After:  Rain hammers a city street as a cyclist cuts through a puddle, spray
        arcing out behind the back wheel.

Before: A close-up of a coffee cup, steam, warm morning light.
After:  Steam curls off the coffee and bends sideways as a hand lifts the cup
        out of frame.

Before: A mountain landscape with clouds, epic and dramatic, wide shot.
After:  Cloud shadows race across the valley floor as the wind pushes a bank of
        mist over the ridge.

The after-prompts share one shape: subject first, verb second, one continuing motion, and one detail that has to change across the clip. That is close to the whole technique. Our seven-part video prompt structure puts it in a fuller framework.

A three-minute diagnostic before you rewrite anything

Run this in order. It is deliberately boring.

1. Read the request fields. Duration, and any motion or camera field the vendor publishes. Fix anything wrong before touching the prompt.
2. Underline the subject's verbs. Zero is cause 1. One verb of stillness is cause 2. Four or more is cause 7.
3. Check for cancelling words. Calm, serene, still, quiet, tranquil, and any attempted negative phrasing.
4. Check the camera-to-subject ratio. If camera instructions outnumber subject instructions, that is cause 5.
5. For image-to-video, cover the picture and read the prompt alone. If it is a caption rather than an instruction, that is cause 3.
6. Rewrite once, generate, compare. One change, so you learn something either way.
7. After three genuinely different attempts, change the model. That is a capability limit and rewriting will not reach it.