A useful Seedance 2.0 prompt tells the model what must remain fixed, what changes during the clip, which source owns each reference role, how the camera moves, and when the shot ends. Write output settings in Oakgen's controls rather than burying them inside the prose. Start with a movement draft, repair the first visible failure, then select 4K only for the approved final when the delivery benefits from it.
ByteDance documents text, image, video, and audio input for Seedance 2.0. Oakgen currently offers two paths: text-to-video with optional image, video, and audio references, and image-to-video with a required opening image plus optional last image.
Write the Prompt Around the Current Controls
Choose Seedance text-to-video for a reference packet or image-to-video when the first frame must stay fixed.
Current Oakgen Seedance controls
| Field | Text-to-video | Image-to-video |
|---|---|---|
| Prompt | Required, up to 5,000 characters | Optional, up to 5,000 characters |
| Opening image | Optional reference images, up to 9 | One required image |
| Last image | No separate field | Optional |
| Video references | Up to 3, with 15 seconds total shown | No |
| Audio references | Up to 3, with 15 seconds total shown | No |
| Duration | 4-15 seconds | 4-15 seconds |
| Aspect ratio | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | Same choices |
| Resolution | 480p, 720p, 1080p, 4K | Same choices |
| Web search | Optional | Optional |
| Generated-audio switch | Not exposed | Not exposed |
No audio switch does not prove that every returned file is silent. Provider requests can use endpoint defaults. Inspect the returned file and plan a controlled sound pass for final delivery.
Choose text-to-video or image-to-video
Use text-to-video when composition can remain open or several references need separate jobs. Use image-to-video when the opening frame already locks the subject, product, set, and art direction. Add a last frame only when the ending composition matters and the path between endpoints is physically believable.
The first-frame, last-frame, and reference guide gives a fuller decision matrix.
The Seedance production prompt
GOAL: [deliverable, audience moment, and intended use]
SUBJECT LOCK: [identity, wardrobe, product geometry, materials, fixed details]
REFERENCE ROLES: [which image/video/audio owns which information]
ACTION: [one causal action with a clear start and end]
ENVIRONMENT: [place, time, motivated light, foreground/background]
CAMERA: [shot size, height, path, pace]
TIMING: [ordered beats and final hold]
CONSTRAINTS: [known failure risks only]
This structure separates stable facts from change. Do not repeat the character or product description with different wording in every beat.
Assign every reference a job
Reference uploads are not a mood board. Write role statements:
Image 1 controls the creator's face, hair, and wardrobe.
Image 2 controls the navy travel mug's shape, lid, handle, material, and color.
Image 3 controls the train interior, weather, and lighting.
Video 1 controls only waist-height camera path and slow forward pace.
Audio 1 controls only rhythm.
If two images own product geometry but show different caps, the conflict exists before generation. Remove one or decide which is authoritative.
Use the Seedance character and product consistency workflow when identity matters across several shots.
Camera language without fake precision
Camera terms can guide the composition, but a generated clip is not a calibrated camera rig. Start with the visible effect:
| Need | Prompt language |
|---|---|
| No reframing | “Locked medium shot at product height” |
| Approach subject | “Slow straight push forward at constant pace” |
| Reveal environment | “Camera moves backward from close to wide” |
| Follow movement | “Waist-height side tracking, subject moves left to right” |
| Change attention | “One slow focus pull from foreground product to creator” |
| Product depth | “Slow 20-degree orbit; do not reveal the back” |
Keep one camera path per beat. “Handheld orbit, crane up, whip pan, and zoom” names four conflicting systems.
Write action as cause and result
Weak:
A premium coffee machine in a beautiful kitchen, cinematic.
Production-ready:
The brushed-steel espresso machine sits on a stone counter. One finger presses
the circular button once. Espresso forms one clean stream and stops below the
cup rim. Locked medium close-up, warm dawn window light, no camera move, steam
burst, extra controls, text, logo, overflow, or second hand.
The second prompt defines contact, liquid, endpoint, framing, and failure risks. It does not claim the model will always perform them.
Use timed beats for 15-second clips
ByteDance publishes 15-second multi-shot output, and Oakgen exposes duration up to 15 seconds. Give each beat enough room:
0-5 seconds: Wide side view. The courier walks through the wet market once.
5-10 seconds: Medium view. The courier stops beneath the yellow awning and rests
both hands on the counter.
10-15 seconds: Close view. The vendor slides one paper parcel forward; hold the
parcel, red string, and warm counter light for the final two seconds.
The 15-second Seedance tutorial includes a continuity ledger and repair table.
Four copyable templates
These are starting briefs, not claimed tests.
Product detail shot
GOAL: Six-second product-page detail clip.
SUBJECT LOCK: Same matte-black headphones; oval cups, narrow silver hinge,
smooth headband, fixed proportions and materials.
ACTION: Right cup folds inward once and returns to open.
ENVIRONMENT: Charcoal studio sweep; soft strip light from camera left.
CAMERA: Medium close-up at product height; slow 20-degree orbit left to right.
TIMING: Orbit for four seconds; fold at second four; hold open for two seconds.
CONSTRAINTS: No logo, text, extra controls, duplicate product, full rotation,
shape change, or background change.
Creator reaction
GOAL: Six-second 9:16 reaction clip for a social edit.
SUBJECT LOCK: Use image 1 for the same woman, short black bob, green knit sweater,
small gold stud earrings, natural skin texture.
ACTION: She reads one phone message, looks up, and gives a restrained smile.
ENVIRONMENT: Quiet cafe window seat in soft overcast daylight.
CAMERA: Locked medium close-up, eye level, natural phone-camera framing.
TIMING: Read for three seconds, look up over one second, hold for two seconds.
CONSTRAINTS: No speech, hand wave, new jewelry, wardrobe change, or camera move.
Reference-led product-in-use clip
GOAL: Ten-second commuter product demonstration.
REFERENCE ROLES: Image 1 owns creator identity and gray coat. Image 2 owns navy
travel-mug geometry. Video 1 owns camera height and slow forward pace only.
ACTION: Creator closes the mug lid, releases it, then lifts the mug once.
ENVIRONMENT: Quiet morning train, cool window light, warm carriage reflection.
CAMERA: Medium close-up with one slow straight push forward.
TIMING: Close lid by second four, release by second six, lift by second eight,
hold product through second ten.
CONSTRAINTS: Preserve face, coat, mug handle, lid, rim, taper, material, color,
scale, and set. No generated label or extra hand action.
First-to-last frame packshot
Use Seedance image-to-video for this one.
FIRST FRAME: Open product on a desk, front three-quarter view.
LAST FRAME: Same product closed in the same position and light.
ACTION: One hand closes the product and leaves frame.
CAMERA: Locked; no zoom, pan, tilt, orbit, or focus change.
TIMING: Complete closure by second four; hold final frame for two seconds.
CONSTRAINTS: Same geometry, surface, hinges, controls, color, shadow, scale,
and background. No text or logo redraw.
Negative constraints in Seedance
Oakgen does not expose a separate Seedance negative-prompt field. Put a short constraint line at the end of the main prompt. Tie it to likely failures:
No duplicate mug, moving handle, new label, camera orbit, wardrobe change,
extra hand action, or lighting transition.
Do not write a second screenplay made entirely of exclusions.
Draft first, render 4K after motion passes
Oakgen currently exposes a 4K option for Seedance 2.0. That is a host control, not a claim that ByteDance documents native 4K for the exact served endpoint.
Use this order:
- Draft at 720p when testing action, timing, camera, and identity.
- Repair the first visible failure with one change.
- Confirm the source, prompt, duration, ratio, and reference roles.
- Choose 1080p if the final destination does not need 4K.
- Choose 4K for an approved hero shot, larger crop, or high-resolution master.
- Inspect the returned file properties and every critical detail.
Higher resolution cannot repair product drift, broken contact, or a rushed timeline. The Seedance 2.0 4K guide contains the full QA sheet.
Repair prompts by symptom
| Failure | Prompt repair |
|---|---|
| Static shot | Replace appearance adjectives with one visible action |
| Rushed action | Remove verbs or extend duration |
| Camera ignores path | Keep one path and pace |
| Character changes | Strengthen one identity block; reduce unseen rotation |
| Product deforms | Add missing angles or reduce orbit/contact |
| Hands break | Start with contact established; widen the shot |
| Ending drifts | Define a final state and hold, or use a last frame |
| Text fails | Reserve a clean tracked area and add type in post |
Return to the symptom map above and change one variable when the first repair does not hold.
Web search and current facts
Seedance forms currently expose an optional web-search switch. Use it only when the scene depends on current information. It does not replace fact-checking, approved brand copy, or final editorial review. Keep it off for fictional scenes and reference-led product work.
FAQ
What is the best Seedance 2 prompt format?
Use goal, subject lock, reference roles, one main action, environment, camera, timing, and short constraints. Put duration, ratio, and resolution in Oakgen's controls.
How long should a Seedance prompt be?
Long enough to assign every job, but short enough that each clause changes the shot. A structured prompt can be much shorter than a paragraph of style references.
Can Seedance 2.0 use video and audio references?
Yes. ByteDance documents both upstream, and Oakgen's text-to-video form currently accepts as many as three videos and three audio files alongside image references.
Does Oakgen Seedance generate audio?
Oakgen exposes no audio-generation toggle. The provider request can use endpoint defaults, so inspect the returned file and complete a separate sound pass when the mix must be approved.
Does Seedance support a last frame?
Oakgen's Seedance image-to-video form currently exposes an optional last image. Text-to-video does not show a separate last-frame field.
Can I generate Seedance 2.0 video in 4K?
Oakgen currently displays a 4K output option for Seedance 2.0. Describe it as an Oakgen option, not a native-4K ByteDance claim, and inspect the delivered file.
Should I render drafts in 4K?
No. Resolve motion, timing, identity, geometry, and framing at a lower review resolution. Use 4K after the shot passes.
Can I use exact logos and text in the prompt?
Keep approved logos, claims, prices, disclaimers, and calls to action for post-production. Generate a stable surface and track the approved artwork onto it.
Continue the Seedance workflow
- Seedance 2.0 complete guide
- Seedance 2.0 4K guide
- Seedance 2 character and product consistency
- Seedance vs Kling vs Veo for creators
Turn the Prompt Into a Controlled Render
Paste one template into Oakgen, assign every reference role, and approve movement before spending on the final resolution.


