Seedance 2.5 Prompt Guide: Plan and Review a 30-Second Video

Start here: a useful prompt in five minutes

If you want to try the method before reading the full production workflow, use seven parts: format + one visible job + reference roles + timed action + audio + exclusions + final frame. The order matters. It moves the prompt from the viewer's takeaway to the things the model may borrow, then ends with a frame a reviewer can recognize.

Create an 8-second, 16:9 product shot.
Show the same compact matte-blue lamp opening into a stable desk light.
Use @Image1 only for the lamp's shape, color, circular base, and silver dial.
00-03s: closed lamp on a dark desk; slow push-in.
03-06s: one hand turns the dial once; the lamp opens in one continuous motion.
06-08s: the warm light turns on; the hand leaves; hold on the open lamp.
Use quiet room tone and one soft click. Keep one lamp and one hand.
Exclude logos, labels, extra controls, captions, music, cuts, and camera changes.

This short prompt is deliberately plain. It asks one question: can the product keep its identity through the hardest interaction? If that works, the next version can add the map, camera arc, second sound, and longer ending. If it fails, eight seconds gives you a cleaner clue than a polished 30-second sequence with several things going wrong at once.

What Seedance 2.5 changes—and what it does not prove

ByteDance officially introduced Seedance 2.5 on July 31, 2026. Its launch material describes a single generation of up to 30 seconds, continued extensions, and up to 30 images, 10 video clips, and 10 audio clips as references. It also describes targeted editing at particular times, along with green-screen, camera-perspective, and white-model controls. BytePlus announced API availability the following week.

Those are product capabilities, not a promise that every 30-second story will be coherent or every supplied detail will survive. ByteDance's own launch article says complex-motion physics and interactions between multiple subjects still need improvement. That qualification matters. A model can accept fifty references while still confusing which jacket belongs to which person. It can expose a local edit control while changing an approved shadow or prop outside the intended region.

The practical shift is from describing a pretty shot to preparing a small production contract. In practice, that can fit on one page. What must the viewer understand? Which facts stay fixed? What is each reference for? What would make the final file good enough to use?

Begin with the shortest useful scene

Thirty seconds is the ceiling, not the assignment. If the useful action takes eight seconds, make an eight-second shot. If the hardest part is a hand rotating a dial without changing the product, test that interaction before surrounding it with an opening scene, a location change, music, and a hero ending.

Begin with one visible job. “Make a cinematic product video” leaves the important decision unstated: what should the viewer understand? “Show that the same compact lamp changes from a closed travel form into a stable desk light” is reviewable. The transformation can pass or fail. The product can keep or lose its shape. The final state can be visible or missing.

The source brief in the downloadable kit separates approved facts from attractive inventions. It asks for the destination, release owner, protected details, unsupported claims, final-frame state, and stop conditions. That mirrors the safer planning method in our broader AI product video workflow, while this guide concentrates on Seedance-specific reference and edit control.

Turn every reference into a contract

A mood board says, “These are all roughly what I like.” A reference contract is more precise: “Use this file for the product body throughout the clip; leave its background, text, hands, and camera angle behind.” That difference becomes more important as the reference count grows.

Give each source one primary job whenever possible. An image might establish a character's face, a product's geometry, or a room's materials. A video might establish only a camera path or the physics of an action. An audio clip might provide voice identity, room ambience, or beat timing. If one file has several jobs, enumerate them. Then state what must not transfer.

Start with the smallest set that can express the idea. Adding twenty similar product images can introduce disagreements in color, scale, lens distortion, labels, and minor hardware details. The model then has to reconcile conflicts that the production team did not resolve. A smaller, owner-approved set with explicit roles is often more controllable than a full folder of loosely related material.

A worked example that can fail honestly

The Northline Pocket Lamp is a fictional fixture. It has a matte-blue cylindrical body, one silver rotary dial, a circular flat base, and a warm-white light. Its brief leaves out price, battery duration, brightness rating, waterproof claims, certification, logos, and customer outcomes because none has an approved source. Those omissions define the honest edge of the example; they are not gaps for the model to fill.

The viewer's job is simple: understand that the lamp changes from a closed travel form into a desk light. The most important risk is not whether the scene looks expensive. It is whether the transformation invents hinges, changes the dial, stretches the body, or produces a second lamp. The final four seconds must settle on the same product at the left side of a dark desk, illuminating a paper map without captions.

This fixture is intentionally modest. Long dialogue, multiple people, rapid environment changes, and complicated hand-object contact would make a poor first diagnostic. Once the core transformation passes, one additional difficulty can be added and versioned.

To reproduce the plan, supply four rights-cleared inputs that match the roles in the table: a product identity image, a final-composition image, a camera-motion clip, and an ambience-and-clicks audio clip. The public kit leaves those inputs as placeholders. Bundling invented substitutes would make the package look more tested than it is and would hide the source decisions a real project needs to make.

Write the prompt from the outside in

A reliable Seedance 2.5 prompt reads more like a compact directing note than a pile of style words. Set the outcome and format first. Bind the references. Lock the facts. Then describe the story in timed beats, followed by camera, look, audio, and exclusions.

GOAL AND FORMAT
Create a 30-second, 16:9 product story.
The viewer should understand that the same compact lamp changes from
a closed travel form into a stable desk light.
End on the open lamp at the left side of a dark desk, casting a warm
pool of light across a paper map. Hold the final state for four seconds.

REFERENCE ROLES
- Use @Image1 only for the lamp's cylindrical shape, matte-blue material,
  circular flat base, and single silver dial. Preserve these for the full clip.
  Do not transfer its background, camera angle, hands, or any text.
- Use @Image2 only for the desk layout, paper map, and warm light composition
  during 22-30s. Do not use it to redesign the lamp.
- Use @Video1 only for the slow push-in and gentle 20-degree arc during 0-8s.
  Do not copy its subjects, objects, environment, or audio.
- Use @Audio1 only for quiet room tone and two soft mechanical clicks.
  Do not add speech, melody, or a voice.

LOCKED FACTS
The lamp remains one object with one silver dial. Its body stays matte blue.
The base remains circular and flat. The light is warm white.
There is no logo, screen, label, price, rating, or written claim.

TIMECODED ACTION
00-05s — A closed lamp rests alone on a dark wooden desk. Medium-wide shot.
The camera slowly pushes in. Nothing else moves.
05-13s — One unbranded hand enters, steadies the base, and rotates the single
silver dial once. The lamp opens through one physically continuous movement.
13-22s — The warm light turns on. The camera completes a gentle arc while the
same lamp illuminates the map. Keep product scale and screen direction stable.
22-30s — The hand leaves. Settle into the final composition from @Image2.
No cut, no text, no second product. Hold the last frame for four seconds.

CAMERA AND LOOK
Natural commercial realism, restrained contrast, dark blue shadows, warm light.
One coherent push-in and arc; no orbit, whip-pan, drone move, or sudden zoom.

AUDIO
Quiet room ambience. One soft click when the dial turns and one subtle click
when the lamp locks open. No music, narration, dialogue, or subtitles.

CONSTRAINTS
Preserve product geometry, part count, dial position, matte-blue color, and scale.
Do not add labels, logos, controls, cables, people, extra hands, cuts, or claims.
If the hand interaction cannot remain plausible, simplify its motion rather
than inventing another mechanism.

The prompt is specific where failure would matter and quiet elsewhere. It does not name five lenses, three lighting styles, and six emotional adjectives. It gives the camera one path. It gives the product one change. It tells the story where to land.

When the script itself contains claims, prepare it separately from the visual direction. Our source-linked video script workflow shows how to bind narration and shots to approved facts before either is handed to a generator.

Find the cheapest useful failure first

A full 30-second request is a poor place to discover that the hand grows an extra finger or the dial becomes a screen. Use a staged ladder:

1. Reference check: inspect the final source files, rights, hashes, and role assignments. Resolve disagreements before upload.
2. Hard-interaction draft: test only the dial turn and opening action at a shorter duration or lower-cost setting when the product permits it.
3. Story draft: add the opening and final frame, but keep audio simple and remove unnecessary scene changes.
4. Audio draft: introduce the two authorized effects and check whether their causes remain synchronized.
5. Final candidate: request the intended duration and resolution only after the earlier failure modes are understood.

Keep each rejected output. A failure tells you whether the problem is identity, action density, camera conflict, reference leakage, physics, sound, or the final-state instruction. Deleting failures creates a misleading success rate and makes the next prompt revision guesswork.

Review with timestamps, not adjectives

“Looks good” is hard to act on. Download the encoded video and write exact observations: “At 00:08.4 the silver dial changes into a black rectangle,” or “At 00:23.1 a second lamp appears behind the map.” A timestamp makes the problem reproducible and tells a targeted edit where to begin.

Technical review matters too. Confirm the actual duration, dimensions, frame rate, video and audio streams, black frames, unintended freezes, clipping, and basic playback. A successful provider status proves that a task finished; it does not prove that the file is intact or suitable for a campaign.

Use targeted editing as a controlled experiment

When one detail fails, resist the urge to rewrite the entire prompt. Ask for one observable change inside one interval and preserve everything else as an explicit review requirement.

Edit the source video only between 08.0 and 10.5 seconds.
Keep one silver rotary dial visible while the hand turns it once clockwise.
Remove the temporary black rectangle that replaces the dial.
Keep the lamp shape, action timing, hand, camera path, light, sound,
and every frame outside this interval unchanged.
Do not add text, controls, cuts, people, or music.
Return to the same protected lamp state by 10.5 seconds.

“Keep everything else unchanged” is not a guarantee. Compare the edited candidate with the accepted source before and after the interval. Check audio continuity across the edit boundaries. If the repair changes several approved elements, reject it and narrow the operation or return to the last good source.

Know which Seedance surface you are documenting

Seedance 2.5 appears through several ByteDance surfaces, and their controls, regions, prices, and limits are not interchangeable. The official launch page pointed consumers to Jimeng AI and Doubao Pro. Dreamina presents Seedance 2.5 globally, while the BytePlus availability announcement describes enterprise and API access in supported markets. BytePlus states that its service is not available in the United States.

As checked on September 15, 2026, the BytePlus LAS enhanced-generation documentation lists the model ID dreamina-seedance-2-5-260628, a 4-to-30-second duration range, synchronized audio, and asynchronous task states such as queued, running, succeeded, failed, and expired. The returned task identifier is retained for a limited period, and the result link is temporary. Archive the permitted output promptly and record its hash, so the only evidence is not left inside a provider history page.

The same API page limits Seedance 2.5 output to 480P and 720P in that workflow; 1080P is not supported for Seedance 2.5 there, and 4K is listed only for Seedance 2.0. A Dreamina subscription screen or another product surface may expose different finishing options. Write the exact platform beside every resolution claim instead of publishing “Seedance 2.5 is 4K” as a universal fact.

Estimate cost before the final request

Pricing changes, and consumer credits are not the same as API billing. The BytePlus LAS documentation checked for this guide describes an enhanced-generation unit price of USD 0.303 per second multiplied by a resolution and input-mode coefficient. For Seedance 2.5 without a video input, the documented coefficients were 0.6785 for 480P and 1.525 for 720P.

These are calculations from one dated BytePlus LAS operator table, not a quote for Dreamina, Jimeng, Doubao, another reseller, or a future billing period. Input video can change the billable duration and coefficient. The provider says only successful generations are charged in this operator path, but a successful file that fails your review still consumes production budget. Measure cost per accepted result, not only cost per request.

Failure patterns worth testing before launch

The table below is a set of review hypotheses, built from the official constraints and recurring risks in generative-video work. It is not a report of failures observed in the unrun Northline example. Use the symptoms as a viewing checklist, then record only what appears in your own output.

Rights and privacy are part of the prompt workflow

A reference-rich model makes provenance more important, not less. Record where every image, clip, voice, effect, and music file came from; what right permits the intended use; and which attributes the model is allowed to transfer. Remove customer information, internal screens, unreleased products, and location clues before upload.

BytePlus currently says its Seedance 2.x API path does not accept ordinary reference uploads containing real human faces. It describes an allowlisted material-library route for verified, authorized likenesses and offers virtual-human assets. Treat that as a current platform rule, not a universal legal safe harbor. Consent, performer agreements, publicity rights, trademarks, music rights, model disclosure, and destination policies still need human review.

Keep public figures, film characters, living artists' names, customer logos, and competitor products out of the workflow unless the project has a documented reason and the necessary rights. A visually convincing result can be the highest-risk version of the mistake.

Download the reference-and-review kit

Download the complete Seedance 2.5 workflow kit. It includes the source brief, reference-role plan, timecoded prompt template, run log, review scorecard, evidence gate, synthetic self-test results, and integrity inventory.

You can also download the prompt template, reference plan, run log, review scorecard, and evidence gate separately.

python3 seedance-2-5-evidence-gate.py --self-test
python3 seedance-2-5-evidence-gate.py .

The first command should report eight passing synthetic tests. The untouched public kit should return HOLD because the references are placeholders and every provider run is NOT_RUN. That is the honest starting state. The gate checks the record structure; it cannot watch the video, authorize a likeness, verify a claim, or approve release.

If you need deterministic assembly after accepting generated shots, the Codex and HyperFrames short-video workflow covers audio-first timing and encoded-file checks. Keep the generative and deterministic evidence separate.

Questions people ask about Seedance 2.5 prompts

Has Seedance 2.5 actually been released?

Yes. ByteDance Seed's official launch post is dated July 31, 2026, and BytePlus announced availability on August 6. Access still varies by product, account, and region.

Is Seedance 2.5 free?

Do not treat “free” as a permanent model property. Dreamina may present trials or plan credits, while BytePlus documents paid usage. Check the exact account and region before promising free access, and distinguish trial credits from unlimited use.

Does Seedance 2.5 generate 4K or 1080P?

Not through the BytePlus LAS Seedance 2.5 generation path documented for this guide. That page currently lists 480P and 720P for Seedance 2.5, excludes it from 1080P, and reserves 4K for Seedance 2.0. Other product surfaces may apply separate enhancement steps, so name the platform and date whenever you report resolution.

Should I use all fifty references?

No. Fifty is a capacity limit, not a quality target. Use the smallest rights-cleared set that covers identity, environment, motion, camera, and audio without contradiction. Add a reference only when you can name its job and exclusions.

Can this workflow tell you whether Seedance beats another video model?

Only after a matched evaluation. Use the same source assets, prompt intent, duration, resolution, audio requirement, repeated runs, cost records, and human rubric. This source-reviewed guide contains no such model run, so it cannot name a winner—and neither can a launch reel or one viral clip.

Why does a detailed 30-second prompt still fail?

Detail does not remove conflicting instructions, overloaded action, bad references, or difficult physics. Reduce the scene to the first failed interaction, test one change, and rebuild from the last accepted state. The shortest useful prompt is often better than the longest impressive prompt.

The release decision belongs after the video

Seedance 2.5's most interesting feature is not simply that it can make a longer clip. It is that generation, references, extensions, and local edits can be treated as one production loop. That loop becomes valuable only when the team preserves evidence across it.

Freeze the job. Bind each reference. Direct observable beats. Retain every attempt. Review the exact encoded file. When the product, person, story, physics, audio, or rights miss the brief, make the next test narrower. Release starts only when the evidence and the video tell the same story.