Seedance prompts: control motion and preserve the subject

Published on November 14, 2026 by Steven Jones

Good Seedance prompts separate three instructions: what the subject does, what the camera does and which details must survive from the source image. Otherwise, a clip can look convincing while doing the wrong thing. A camera orbit may become a rotating object, or a gentle product movement may quietly redraw the label.

This guide focuses on Seedance 2.5 image-to-video prompts, not generic AI video prompting. To follow the examples, open the DIY AI Studio video generator, choose Image to Video and select Seedance 2.5. Use one landscape reference image and change only the motion instruction between paired attempts.

1. Hold the image, model, duration, resolution and aspect ratio constant.
2. Test subject movement with a locked camera.
3. Test camera movement with a stationary subject.
4. Inspect the complete clip, retain failed takes and record credits charged per accepted result.

Evidence status: The prompt pairs below are a reproducible test specification, not completed Seedance 2.5 experiments. No Seedance-specific acceptance rate, retry count or generation charge is claimed here. Add results only after the original clips and job records have been retained.

Use the Seedance version your interface actually exposes

As checked on 14 November 2026, DIY AI Studio lists Seedance 2.5 under the model key seedance_25. Its video interface supports text-to-video or image-to-video, one starting image, 16:9 or 9:16 output, five or ten seconds, and 720p or 1080p. The present Studio integration does not expose a repeatable random seed, multiple reference videos, a last-frame image or the full editing controls described for the underlying model.

BytePlus documents the standard 2.5 model as dreamina-seedance-2-5-260628. Its wider capabilities include timestamped shot instructions and multi-asset references, but those are not interchangeable with controls available in every third-party interface. Seedance 2.0 and Seedance 2.5 Flash are separate versions. The exact backend build behind Studio’s displayed model name is not independently exposed in this check.

For model-specific syntax and advanced reference handling, use the official BytePlus Seedance 2.5 prompt guide. The short prompts here deliberately stay within Studio’s single-image route.

---

---

Use a source image with details that can actually fail.

For the proposed comparison, reuse DIY AI’s fictional TIDAL SOAP bottle photograph from our original image-to-video source and comparison. It has amber glass, a dark-teal cap, a cream label, an empty outlined-circle logo and three identifiable text lines. The existing comparison documents a 1536 x 864 JPG reference and the exact input hash. It also shows why correct label text alone does not make a clip acceptable if the product becomes cropped.

Keep the same JPG pixels for every Seedance attempt. Do not crop one version, sharpen another or upload a screenshot of a screenshot. The label, cap, glass colour and complete bottle silhouette give you separate checks for identity preservation. A plain, featureless object would make a flattering result much easier to mistake for faithful reproduction.

Test 1: move the bottle while keeping the camera fixed

Start both prompts with this identical reference instruction:

Use the uploaded photograph as the first frame of a single continuous five-second product shot. Preserve the amber-glass bottle, dark-teal cap, cream label, outlined-circle logo and the text TIDAL SOAP, CEDAR + SEA SALT and 250 ml. Keep the original lighting and background. The entire bottle remains visible throughout. No cuts, subtitles or background music.

Prompt S0, stationary control:

The bottle remains completely still for all five seconds. The camera is locked at its starting position with unchanged framing.

Prompt S1, subject-motion change:

The camera stays locked at its starting position. Between seconds 1 and 4, the bottle turns approximately 15 degrees clockwise around its vertical axis, as though on a concealed turntable. Its base stays centred on the same point. It settles and remains still for the final second.

The intended difference is the bottle’s orientation, not the size of the whole scene. Check whether the label turns naturally with the glass, whether the cap and base remain connected, and whether the background moves despite the locked-camera instruction. Some perspective change to the label is expected during a genuine rotation; invented letters or a replacement logo are not.

Test 2: move the camera while keeping the bottle still

Reuse the same source, reference instruction and generation settings. Submit a fresh control rather than assuming a previous static render is deterministic.

Prompt C0, stationary control:

The bottle remains stationary and upright for all five seconds. The camera is locked at its starting position. The cap and base stay fully inside the frame.

Prompt C1, camera-motion change:

The bottle remains stationary and upright. From seconds 1 to 4, the camera makes a very shallow forward dolly, ending approximately 5% closer to the bottle. Keep the full cap and base visible. The camera stops for the final second. The bottle does not rotate.

A true push-in should increase the bottle’s apparent scale without changing its identity or orientation. A bottle that bends towards the lens, rotates to show another side or acquires new packaging has failed the requested camera-only action. Compare the cap-to-base framing at the opening, middle and ending, not just the last attractive still.

Why matching settings still cannot prove which words worked

Studio currently exposes no seed-setting control for this route. Two outputs with identical prompts may therefore differ before you change a single word. The S0/S1 and C0/C1 comparisons can reveal useful behaviour, but one clip per condition cannot establish that the changed phrase caused the outcome.

For an initial editorial check, create the four named clips and keep every result. If the budget allows, repeat each condition with unchanged inputs before drawing a stronger conclusion. Never quietly regenerate only the weaker prompt until it produces a winner. That selects for luck rather than prompt quality. If the model or provider changes mid-test, start a new cohort instead of combining the results.

Score subject preservation separately from motion accuracy

Check	Pass condition	Reject or flag when
Subject identity	Original bottle, cap, colour and logo remain recognisable	Packaging changes, cap morphs or logo gains new marks
Lettering	All three label lines remain legible at the declared viewing size	Letters change, vanish or become invented text
Subject motion	S1 turns the bottle; S0 does not	Background moves, object slides or requested turn never occurs
Camera motion	C1 pushes in gently; C0 remains locked	Object rotates instead or the push-in becomes excessive
Editing-ready ending	Complete bottle, including cap and base, remains in frame	Either end is cropped or the final second is unstable

Review at normal speed first. Then inspect the first frame, two intermediate points and the last frame at native resolution. Record the first visible failure rather than treating a technically completed generation as a usable clip. A slight natural shift in shadows may be tolerable for a social post; changed product branding would fail an advert even if the camera movement were perfect.

When preservation wording fails, change the workload.

Two recurring practical problems deserve more attention than another list of cinematic adjectives. Creators report camera requests that move the subject instead, and image-to-video attempts that soften detailed artwork despite lengthy instructions to retain every pixel. The latter is particularly noticeable with fine ink lines and watercolour texture. The model regenerates video frames; a reference image does not guarantee pixel-level preservation.

If the bottle rotates but the label warps, try a smaller 5-degree turn before adding more preservation clauses. If a forward dolly clips the product, reduce the requested approach or start from a wider source. If an illustration’s background changes, consider animating a limited foreground element and compositing it back onto the original static artwork. This is more reliable for exact brand assets than assuming the model will protect every texture through motion.

Also avoid overloading a five-second test with dialogue, two camera moves and multiple action beats. BytePlus supports timestamped sequences in Seedance 2.5, but timestamps express intent rather than guaranteeing a precise motion path. One physical event and one camera instruction are enough to reveal which part of the brief is being missed.

Your search, your sources

Make DIY AI a preferred source

See more of our reporting in Google Top Stories, AI Overviews and AI Mode.

Add DIY AI on Google

Record retries and calculate the cost of an accepted clip

Save each output with its prompt variant and the job identifier. For the first cohort, use the following ledger; leave results unfilled until the files and actual charges are available.

Clip	Single changed instruction	Attempts	Accepted	Charged credits	Failure reason
S0	Stationary subject and camera	Not run	Not assessed	Not measured	Pending
S1	15-degree subject rotation	Not run	Not assessed	Not measured	Pending
C0	Stationary subject and camera	Not run	Not assessed	Not measured	Pending
C1	Shallow camera push-in	Not run	Not assessed	Not measured	Pending

Cost per accepted clip = total credits actually charged for every attempt in a cohort divided by its accepted clips. Include paid creative failures, not just the chosen render. Keep failed technical jobs and refunds as separate entries. If nothing passes, report no accepted clip rather than dividing by zero or assigning a misleading zero cost.

Before spending on retries, decide whether the failure is fixable in editing. A clip with a useful middle but a cropped ending may suit a short cutaway, yet it fails this full-duration product brief. Changing the acceptance rule after viewing the result makes cost comparisons unreliable. Keep the same rule across all four prompts and record any trimmed versions separately.

Which Seedance prompt should you keep?

Use the subject-motion pattern when a product or character must act inside an otherwise stable scene. Use the camera-motion pattern when the starting image already shows the correct product pose, and you only need movement around it. For exact lettering and packaging, start with the least ambitious action that communicates the shot; extra motion creates more opportunities for unwanted reconstruction.

The useful result from this experiment will not be the prettiest isolated frame. It will be the prompt that reliably delivers the requested movement, preserves the source and gives an editor a complete ending at a recorded cost. Until the Seedance 2.5 clips are generated and checked, those outcomes remain open questions rather than editorial verdicts.