Seedance 2.0 Prompts: Reference Mode and Audio Cues

ByteDance launched Seedance 2.0 with multimodal reference input and up to 15-second multi-shot audio-video output. That makes reference roles useful: an image can anchor a character, a clip can suggest movement, and audio can guide sound. Do not ask a single reference to control appearance, camera, setting and performance at once. For a text-only test, omit the reference lines and keep the scene plan intact.

Character and setting: “Use image 1 only for the appearance of the original courier: short dark hair, green jacket, brown satchel. Use image 2 for the station's architecture and lighting. Shot 1: courier enters the station. Shot 2: courier stops at a timetable. Keep the jacket and satchel unchanged.”

Camera reference: “Use video 1 for the slow sideways camera move, not its actors or location. Show an original red paper boat drifting across a shallow fountain. Keep the boat centered as the camera moves.”

Sound reference: “Use audio 1 for the rhythm of soft footsteps. A botanist crosses a glass greenhouse in one continuous shot; do not copy speech from the reference.” Only attach assets you have rights to use. These are templates, not tested generations.

Give each reference one job

For a reference-enabled interface, prepare an authorized character image of an adult courier in a yellow raincoat and a location image of a red-brick station entrance. Attach them with the interface’s reference controls and note which is the character reference and which is the location reference. A typed filename alone does not attach an image.

Complete example brief: “Use the attached courier image for the character’s appearance and yellow raincoat. Use the attached station image for the location and brickwork. The courier walks toward the entrance, pauses, and looks up at the clock. Medium-wide framing with a slow forward camera move. Keep the raincoat and station architecture consistent. Audio: footsteps and quiet street ambience, no dialogue.”

Check the character, setting, action and sound against their assigned references. If the output confuses roles, test one reference and one action first, then add the second reference back. If your interface lacks multiple reference inputs, use its supported workflow instead. This proposed task has not been generated or measured.

Write sound as part of the scene: voice identifies who speaks and what they say; foley names an action-linked effect; ambience locates the environment. Example: “The botanist says, ‘The first leaves are here.’ Audio: footsteps on wet stone, quiet greenhouse ventilation, no music.” Avoid conflicting commands such as “silent scene” plus detailed dialogue. If the model places a sound at the wrong moment, make the related action simpler and try again.

Seedance 2.5 extends the family with longer storytelling and stronger reference/editing controls, according to ByteDance. Keep your 2.0 subject, shot and sound structure, then adapt asset roles and length to the newer model's current interface. Do not assume a 2.0 prompt or result transfers unchanged. The speed-oriented 2.0 Fast variant exists, but output quality depends on settings and the scene; compare your own drafts before choosing it.

ByteDance model-level comparison:

The counts describe ByteDance's announced model interfaces; they do not establish identical limits in CrePal. For a 2.0 prompt, preserve its subject and action, then decide whether a longer 2.5 generation actually helps the story. Do not simply paste more shots because a newer model permits longer output.

Before writing a mixed-reference prompt, make a small asset list: Image A holds the character's clothing, Image B holds the room layout, Video C suggests a camera move, and Audio D provides a timing cue. Then say which property each asset may influence. For a cafe scene: “Keep the red jacket from Image A; use the table arrangement from Image B; borrow only the slow lateral camera movement from Video C.” This avoids asking one reference to control identity, set design and motion simultaneously.

Check the chosen model interface before uploading: ByteDance's announced model limits do not establish what a particular CrePal project exposes. If a reference type is unavailable, convert the job into a concise text instruction or select a supported workflow; do not claim the same fidelity. After generation, compare each asset's assigned property with the output. Record the model version and reference roles next to the prompt so a later 2.5 adaptation has a clear starting point.

For a text-first multi-shot brief, start with the Seedance shot templates before adding reference roles. To compare a single cinematic shot with another model family, keep the action fixed and use the Veo prompt structure; check current model availability in the generation service.

CrePal's video creator page describes scene-level model selection and review before export. Use that workflow to check whether a Seedance option and the reference controls described here are actually present in the current interface before preparing a paid run. ByteDance's linked model announcements document model limits; CrePal's page does not guarantee that every control is exposed in every account or route.

Related official reading: CrePal's Seedance 2.0 lip-sync guide. CrePal's Seedance 2.0 lip-sync guide recommends beginning with one visible speaker, one short line and limited head or camera movement, then checking whether speech and mouth motion align. Use that as an acceptance test for the audio examples here; split longer dialogue into separate shots if sync drifts. The guide itself labels its troubleshooting examples illustrative; no synced clip from this batch is documented.

Evidence note added 5 November 2026. The prompts in this guide are editorial examples. The linked CrePal pages document product workflows or illustrative use cases; they do not verify an output from these exact prompts. No prompt-output pair is published here.

To document a result for this topic: Record the model label actually available to the account, settings, prompt and clip; inspect motion and lip synchronization. Keep the exact prompt, model/version, settings, output URL, revisions and human review decision together.

Start with one example from this guide, then compare the result with your intended subject, style and details.

Create with CrePal

CrePal's Terms identify SIGNOO PTE. LTD. as the service provider. The prompt examples here are instructions to try, not recorded generation tests.

• Privacy
• Terms
• Send feedback
• Email support