The symptom in plain English: what “freezes the wrong parts” actually means

You feed an image into image‑to‑video, ask for “make it come alive,” and the model animates everything except the thing you care about.

Common versions of the same problem:

The background shimmers and the subject stays locked.
The camera “moves” (fake zoom/orbit) but the subject doesn’t.
The wrong object becomes the hero (e.g., clouds move, product doesn’t; hair moves, face doesn’t).
Text/logos warp while the main object remains still.

As of 2026‑06‑12, the most reliable fix is rarely a hidden setting—it’s a repeatable prompting workflow that reduces ambiguity, isolates motion, and iterates in small steps.

Root causes (ranked): ambiguity, competing motion cues, and missing subject locks

1) Ambiguity: the model can’t tell what must move

If your prompt says “animate the scene,” you’ve basically told the model everything is fair game. Tools generally respond better to natural, detailed language than vague instructions. (https://lumalabs.ai/learning-hub/best-practices)

2) Competing motion cues: multiple plausible movers

A busy image can contain many “tempting” motion candidates—water, hair, trees, reflections, screens, fabric folds. If you don’t explicitly crown one motion subject, the model may pick whatever looks most “animatable.”

3) Missing subject locks: you didn’t protect what must remain still

Most creators add the motion request, but forget to explicitly forbid motion elsewhere. In practice, you want both:

a motion contract (what moves)
a stillness shield (what does not move)

4) Camera motion masking subject motion

Camera movement can create the illusion of animation while your intended subject stays inert. Some tools expose camera motion options like pan/orbit/zoom. (https://lumalabs.ai/learning-hub/best-practices)

5) Limited “control knobs” beyond the prompt

In many image‑to‑video systems, granular control is still mostly prompt-driven rather than slider-driven. (https://www.tomsguide.com/ai/ai-image-video/i-put-luma-dream-machine-to-the-test-with-7-ai-videos-heres-how-it-stacks-up-to-sora)

Step 1 — Write a “motion contract” prompt (one mover, one action, one camera)

Start by defining the motion subject explicitly as a single noun phrase, and repeat it once to anchor attention.

Rule of three (debug mode):

One mover: “the red umbrella”
One action: “opens slowly”
One camera: “locked-off camera”

Keep it natural and specific about key elements like mood/lighting/style when relevant. (https://lumalabs.ai/learning-hub/best-practices)

Example motion contract (template):

Motion subject: [SINGLE NOUN PHRASE]. The [SINGLE NOUN PHRASE] performs [ONE VERB PHRASE]. Locked camera. Realistic motion.

Step 2 — Add a “stillness shield” for everything that must not move

Add a single line that forbids motion drift:

background
lighting
text/logos
secondary objects
overall composition

This “stillness shield” often prevents the model from spending its motion budget on shimmer.

Stillness shield (template):

Everything else remains still: background, lighting, shadows, reflections, text, logos, and all other objects stay unchanged.

If your shot includes text you must preserve, be explicit about the text too—many systems allow requesting text by describing it directly in the prompt. (https://lumalabs.ai/learning-hub/best-practices)

Step 3 — Use time-coding to prevent motion from spreading (0–2s / 2–5s / 5–8s)

Time-boxing localizes animation. Instead of “the product rotates,” give it a schedule:

0–2s: hold still (stabilizes the scene)
2–6s: do the action
6–8s: settle to stillness

Even if your generator outputs shorter clips (some systems commonly generate ~5 seconds per render), the same structure still helps because it conveys sequence and rest points. (https://www.tomsguide.com/ai/ai-image-video/i-put-luma-dream-machine-to-the-test-with-7-ai-videos-heres-how-it-stacks-up-to-sora)

Time-box template:

0–2s: hold perfectly still. 2–6s: [ACTION]. 6–8s: settle, stop moving, return to stillness.

Step 4 — Reduce camera motion to debug: lock the camera, then re-introduce movement

When motion goes wrong, remove camera motion first.

Pass A (debug): locked camera, only subject motion
Pass B (cinematic): add one camera move (slow push-in or slight orbit)

Many tools explicitly list camera motion modes (pan/orbit/zoom). (https://lumalabs.ai/learning-hub/best-practices)

The goal is to ensure the subject truly animates before you add anything that can hide failure.

Step 5 — Iterate with A/B micro-edits: the 6 fastest tests to run

Because you often don’t get granular controls beyond prompting, treat each render like a tiny experiment. (https://www.tomsguide.com/ai/ai-image-video/i-put-luma-dream-machine-to-the-test-with-7-ai-videos-heres-how-it-stacks-up-to-sora)

Quick A/B testing table (change one variable per render)

Test	Keep constant	Change only this	Why it works
A/B #1	Same image, same prompt	Motion verb (e.g., “tilts” vs “rotates”)	Verbs imply different body mechanics
A/B #2	Same motion subject/action	Camera (locked vs slow zoom)	Removes masking from camera movement
A/B #3	Same everything	Subject naming (more specific noun phrase)	Anchors attention on the right object
A/B #4	Same prompt	Add/reduce stillness shield detail	Stops background shimmer and logo warping
A/B #5	Same prompt	Time-boxing vs no time-boxing	Prevents motion spread
A/B #6	Same prompt	Reduce background detail via prompt (“simple backdrop”)	Lowers competing motion cues

Copy-paste prompt patterns: 8 fixes for the most common failure modes

Use these as starting points and swap bracketed fields.

1) “Animate only the hair/cloth, keep body still”

Motion subject: the woman’s hair. The woman’s hair gently sways as if in a soft breeze. The body and face remain perfectly still. Locked camera. 0–2s still, 2–6s hair sways, 6–8s settle. Everything else remains still: background, lighting, shadows, reflections, text, logos, and all other objects unchanged.

2) “Product rotates, hands stay frozen”

Motion subject: the perfume bottle. The perfume bottle rotates slowly 20 degrees left then returns to center. Hands remain completely still. Locked camera. 0–2s hold, 2–6s rotate, 6–8s settle. Everything else remains still: background, lighting, reflections, labels, text, logos.

3) “Face expression only, no head movement”

Motion subject: the man’s facial expression. The man’s facial expression changes from neutral to a subtle smile; no head tilt, no eye drift, no body movement. Locked camera. 0–2s hold, 2–6s smile forms, 6–8s hold. Everything else remains still: background, lighting, hair, clothing, text.

4) “Eyes blink only (micro-motion)”

Motion subject: the cat’s eyes. The cat’s eyes blink once naturally. No other motion. Locked camera. 0–2s still, 2–6s single blink, 6–8s still. Everything else remains still: fur, whiskers, background, lighting.

5) “Only one object moves in a busy street photo”

Motion subject: the yellow taxi in the foreground. The yellow taxi in the foreground rolls forward slightly and stops. All pedestrians, buildings, signs, and other cars remain frozen. Locked camera. 0–2s hold, 2–6s taxi moves, 6–8s stop. Everything else remains still: background, lighting, text, logos.

6) “Keep background still, animate steam/smoke only”

Motion subject: the steam rising from the coffee. The steam rising from the coffee curls upward slowly. Cup, table, and background remain perfectly still. Locked camera. 0–2s still, 2–6s steam rises, 6–8s settle. Everything else remains still: lighting, shadows, text, logos.

7) “Water ripples only, keep subject locked”

Motion subject: the water surface. The water surface shows gentle ripples moving outward. The boat and shoreline remain fixed and unmoving. Locked camera. 0–2s hold, 2–6s ripples, 6–8s calm. Everything else remains still: background, lighting, reflections (except the ripples).

8) “Mechanical motion: one hinge/lever moves”

Motion subject: the laptop lid. The laptop lid opens smoothly from closed to slightly open, then stops. Keyboard, screen content, and desk remain still. Locked camera. 0–2s still, 2–6s lid opens, 6–8s hold. Everything else remains still: background, lighting, text, logos.

When it’s not your prompt: image selection issues (and how to compensate)

Sometimes the input image sets you up to fail. The model may “choose” motion based on what looks easiest to animate.

Common image pitfalls

Tiny subject: The intended mover occupies a small area, so background textures dominate.
Cluttered background: Too many objects compete for motion.
Motion blur: The subject already looks smeared; the model may avoid moving it.
Repeating patterns (grids, stripes, leaves): Can produce shimmer or crawling artifacts.
Occlusions: The motion subject is partially hidden, so the model “gives up” and animates something else.

Compensation strategies (prompt-level)

Declare subject prominence: “the [subject] is the clear focus in the foreground.”
Simplify the world: “background stays static and minimally detailed.”
For repeating textures: “no shimmering, no texture crawl; keep patterns stable.”
For occlusions: choose a different image if possible; otherwise specify the visible moving part (e.g., “the visible edge of the flag”).

A 5-minute preflight checklist before you render your “final”

I defined one motion subject as a single noun phrase and repeated it once.
I asked for one action (one verb phrase), not a list.
I used a stillness shield: background, lighting, text/logos, secondary objects remain unchanged.
I locked the camera for the debug pass.
I added time-boxing (hold → action → settle).
I planned one A/B change for the next render if this one fails.

FAQ

Why does image-to-video animate the background instead of my subject?

Usually the prompt doesn’t explicitly name a single motion subject, so the model picks an easy motion target (often textures like water, foliage, hair). More natural, detailed prompts help reduce that ambiguity. (https://lumalabs.ai/learning-hub/best-practices)

In many cases, control is still mostly prompt-driven rather than granular setting-driven—so micro-iteration and A/B testing are practical. (https://www.tomsguide.com/ai/ai-image-video/i-put-luma-dream-machine-to-the-test-with-7-ai-videos-heres-how-it-stacks-up-to-sora)

CTA: automate your motion tests with Veo3Gen

If you’re doing the A/B micro-iteration workflow above, it helps to run structured batches (one variable changed per render) and keep outputs organized.

Explore the Veo3Gen API to script prompt variants and iterate faster: /api
See plans and usage options before you commit to a big round of tests: /pricing

Try Veo3Gen (Affordable Veo 3.1 Access)

If you want to turn these tips into real clips today, try Veo3Gen:

Start generating via the API: /api
See plans and pricing: /pricing

Why Image‑to‑Video Freezes the “Wrong 90%” (and Animates the Wrong Thing): A Creator Troubleshooting Workflow for Veo3Gen (as of 2026‑06‑12)