Seedance 2.5 Prompt Guide: Control Camera, Motion, Lighting, and Timing
Write better Seedance 2.5 prompts with a shot-based framework for subject, action beats, camera movement, lighting, audio, timing, references, and continuity.
The fastest way to ruin an AI video prompt is to write it like a poster description.
“Cinematic cyberpunk woman in a rainy city, beautiful lighting, dynamic camera” may produce an attractive frame, but it does not direct a shot. Where does the character begin? What happens first? What moves? When does the camera move? What should remain unchanged at the end?
Seedance 2.5 is designed for longer, reference-driven generation. Dreamina’s official product page says it can combine scripts, images, video, music, and other inputs, with up to 30 seconds in standard mode. That makes prompt structure more important, not less.
This guide gives you a repeatable way to write Seedance 2.5 prompts without turning them into unreadable novels.
Think in shots, not adjectives
A usable video prompt answers seven questions:
- Who or what is the subject?
- Where does the shot happen?
- What visible actions occur, and in what order?
- What does the camera do?
- How does light change?
- What sound belongs to the scene?
- What must remain consistent?
Use this formula:
Subject and identity + environment + timed action beats + camera and framing + lighting and color + audio + continuity constraints + ending state
The ending state matters because a longer generation needs a destination. “She walks through the market” can wander forever. “She reaches the blue stall, places the envelope on the counter, and looks toward the bell as the camera stops” creates a final composition.
Build a compact identity block
When a character recurs, define three to five visible anchors:
Clearly adult woman in her thirties; square face, short black bob with one silver streak above the left eyebrow; rust-red courier jacket; narrow navy trousers; brass message tube across her back.
Repeat this wording across scenes. Do not call the jacket “rust-red” in one prompt, “burnt orange” in the next, and “copper leather” later unless you want variation.
Avoid invisible biography in the visual prompt. “She regrets leaving her hometown” is not directly visible. Translate it into behavior:
She hesitates at the old station sign, touches its scratched edge, then continues without looking back.
Write action as numbered beats
Seedance 2.5’s 30-second standard mode can hold several actions, but a loose list may happen out of order. Divide the scene into beats.
Example:
0–5s: courier enters frame left and walks beside the stopped tram.
5–12s: she hears a bell, slows, and turns toward a narrow market lane.
12–22s: she crosses to a blue flower stall while pedestrians pass behind her.
22–30s: she removes the brass tube, places it on the counter, and holds still as the camera settles into a medium close-up.
Time codes communicate sequence and relative duration. They do not guarantee frame-perfect editing. Keep the number of events realistic. A 30-second prompt should not contain a chase, costume change, argument, explosion, and sunrise.
Direct the camera with one main move
Conflicting camera terms are a common failure. “Drone shot, handheld close-up, locked-off dolly, fast orbit” asks for several incompatible visual grammars.
Choose:
- starting shot size;
- camera height and angle;
- one main movement;
- ending shot size;
- pace.
Example:
Begin in a wide eye-level shot six meters ahead of the courier. The camera dollies backward at her walking speed, then arcs gently right as she reaches the stall, ending in a stable medium close-up. No cuts, no handheld shake, no orbit.
Useful movement terms include:
- dolly in/out: camera physically approaches or retreats;
- truck left/right: camera moves sideways;
- pan/tilt: camera rotates from a fixed position;
- crane: camera rises or descends through space;
- arc: partial movement around the subject;
- locked-off: no camera movement;
- handheld: deliberate small instability;
- rack focus: focus moves between depth planes.
Use cinematic vocabulary only when you understand the visual result.
Describe movement through mechanics
“Move naturally” is vague. Describe balance, direction, contact, and pace.
Weak:
She dramatically grabs the package.
Stronger:
She plants her left hand on the counter, lifts the brass tube from her right shoulder with both hands, rotates it horizontally, and sets it down without sliding.
For R2V guidance, a clean performance reference can communicate more than prose. Dreamina promotes green-screen or white-model references for motion and spatial control. Record full limbs, visible contact, a stable camera, and the target aspect ratio. Obtain permission from the performer.
Treat lighting as an event
Lighting should come from a source and remain spatially coherent.
Example:
Overcast blue daylight from the open market roof; warm tungsten bulbs inside the flower stall. As she approaches, the warm light gradually reaches the right side of her face while the left side remains cool. No flashing neon and no unexplained color changes.
Useful lighting information:
- time of day;
- source position;
- hardness or softness;
- color contrast;
- atmospheric effect;
- intended change over time.
Avoid stacking “cinematic, volumetric, neon, golden hour, moonlight, studio light” in one scene.
Prompt audio separately
Seedance 2.5 is marketed as an audio-video system, but sound still needs direction.
Separate:
- ambience;
- action sound;
- dialogue;
- music;
- silence.
Example:
Audio: light rain on the glass roof, distant tram motor, soft market footsteps, one clear brass bell at 6 seconds, tube contacting wooden counter at 25 seconds. No music, no narration, no random subtitles.
If dialogue is essential, keep it short and assign the speaker:
Courier, quietly at 26 seconds: “For the gardener.” No other speech.
Long dialogue increases lip-sync, voice, timing, and character-assignment risk. Consider recording final dialogue separately when exact words matter.
Use references as named departments
Do not upload a pile of files and hope the model knows which feature to copy. Create a reference manifest:
| File | Authority |
|---|---|
| character_front.png | Face, hair, jacket, body proportions |
| market_wide.png | Architecture and stall placement |
| tube_detail.png | Prop shape and material |
| walk_turn.mov | Body action and timing only |
| palette.png | Color relationships, not composition |
If two files disagree, decide which wins before generating.
Dreamina advertises as many as 50 inputs, but the best Seedance 2.5 prompt workflow starts small. Add a reference only to solve a diagnosed gap.
Add a continuity block
End with concise constraints:
Preserve the adult courier’s face, left-side silver hair streak, rust-red jacket, brass tube dimensions, and market layout throughout. Tube begins on her back and ends on the counter. No additional characters touch it. No text, logo, costume change, duplicated object, or cut.
Constraints should protect production facts, not list generic complaints such as “no bad anatomy” or “no low quality.”
A complete Seedance 2.5 prompt
30-second continuous cinematic shot, 16:9. Clearly adult courier in her thirties with square face, short black bob and one silver streak above the left eyebrow, rust-red courier jacket, navy trousers, brass message tube across her back. She enters frame left beside a stopped tram in a covered flower market. 0–5s: walks toward camera at measured pace. 5–12s: one brass bell sounds; she slows and turns toward a blue flower stall. 12–22s: crosses the lane while three background pedestrians pass without touching her. 22–30s: removes the tube with both hands, places it horizontally on the wooden counter, and looks toward the bell. Start wide at eye level; smooth dolly backward, gentle right arc near stall, finish in stable medium close-up, no cuts. Cool overcast roof light; warm stall bulbs gradually light right side of face. Audio: rain on glass, distant tram, footsteps, bell at 6s, tube on wood at 25s; no music or subtitles. Preserve face, silver streak side, jacket, tube, stall geometry, object journey, and adult appearance.
This prompt can be paired with five authorized references rather than fifty.
Diagnose failures one variable at a time
The action order is wrong
Reduce the number of beats and make the ending explicit. Separate actions with time ranges.
The character changes
Use fewer identity references, remove conflicting portraits, lock the anchor language, and shorten the scene for diagnosis.
The camera ignores direction
Choose one principal move. Remove style terms that imply a different camera grammar.
The result is visually busy
Reduce background actors, effects, props, and lighting colors. Video complexity compounds.
Dialogue or audio fails
Shorten dialogue, name the speaker, define the moment, and consider external final audio.
One object is wrong
Use Seedance 2.5’s advertised local-editing workflow if available. Check the repaired area across the full timeline for flicker.
Build assets before prompts
For anime production, establish the character and storyboard first. Elser AI can help creators develop consistent characters, comic panels, and visual sequences before they write the motion prompt. Export only approved frames with clear names.
A prompt cannot rescue an undefined face or an outfit that changes across source images.
FAQ
How long should a Seedance 2.5 prompt be?
Long enough to define the shot, often 150–300 focused words for a 30-second scene. Clear structure matters more than word count.
Should I use time codes?
Yes, for action order and pacing. Treat them as direction rather than guaranteed frame-accurate editing.
How many references should I upload?
Start with four to seven well-assigned references. Use more only when each has a distinct purpose.
Can I name a film director or living artist?
Describe broad visual properties instead of copying a living creator’s signature style. Use original or licensed reference material.
How do I stop random subtitles?
Explicitly request no text or subtitles, avoid reference images containing text, and inspect the output. Dreamina also claims cleaner output, but verification is still necessary.
Conclusion
A strong Seedance 2.5 prompt is a miniature shooting plan.
Define the subject, break action into timed beats, choose one camera move, give light a source, separate audio cues, assign every reference a job, and protect identity with a short continuity block. End the scene deliberately.
The model’s new reference capacity does not eliminate direction. It rewards it. The clearer your production decisions are before generation, the less you will depend on a lucky first result.
