Seedance 2.5 Prompt Guide: Action, Multi-Shot and References

What Seedance 2.5 is built for

Seedance 2.5, from ByteDance, is the flagship model in Biggsfield's AI Video studio. It does three things that shape how you should prompt it. First, it generates native audio: lip-synced speech, sound effects and ambience come out of the same prompt as the picture. Second, it handles several shots in a single generation, so one prompt can hold a small sequence instead of one isolated clip. Third, it accepts multiple reference images, which lets you anchor specific people, products or places.

It also runs longer than most models. You can choose 5, 8, 10, 15, 20 or 30 seconds, at 480p, 720p or 1080p. That length is what makes multi-shot useful. A 30-second clip can carry an establishing shot, a dialogue beat and a reaction, all from one prompt.

Seedance 2.5 understands a particular structure. You get better results when you write the prompt the way the model reads it, instead of the way you would describe a scene to a friend.

Describing action: one clear beat per shot

The most common reason AI video looks mushy is vague action. "A woman in a café, cinematic" tells the model about mood but gives it nothing to animate. Give every shot a subject, a physical action and a camera, in that order. "A woman in a green raincoat pushes open the café door, shakes water off her umbrella and scans the room" is something the model can perform.

Use concrete verbs: pushes, turns, sets down, leans in, glances back. Avoid stacking abstractions like "elegant", "emotional" or "epic" in place of action. If you want emotion, show it through behavior. "She exhales and her shoulders drop" reads better than "she feels relieved".

Keep each shot to one main beat. A five-second shot rarely has room for someone to enter, sit, order, drink and leave. If the story needs more, split it into shots. That is what multi-shot is for.

Camera language belongs in every shot. Name the shot size (wide, medium, close-up), the movement (slow push-in, handheld follow, static locked-off, crane down) and, if it matters, the lens feel (wide-angle, telephoto compression). Biggsfield's AI Video studio can also add camera moves such as crash zooms, dolly zooms, FPV and bullet time in each model's own words, so you can pick a move and let the AI director phrase it.

Dialogue and sound in the same prompt

Because Seedance 2.5 generates lip-synced speech, how you write dialogue directly affects the result. The rule from the model guide is simple. Put spoken lines in double quotes, and say who speaks them and how. For example: The barista says, warmly and a little tired, "Last one of the night. You're lucky."

Naming the speaker matters most when there is more than one person in frame. Otherwise the model has to guess whose lips should move. The delivery note (whispering, laughing, flat, rushed) shapes both the voice and the facial performance.

Describe sound the same way you describe picture, specifically. "Rain drumming on the window, espresso machine hissing, low jazz from a speaker" gives the model a soundscape to build. Keep lines short. A single sentence of dialogue per shot is usually easier to sync cleanly than a paragraph.

Writing multi-shot sequences

Seedance 2.5 can render several shots in one generation if you write them in order and label them: Shot 1: ... Shot 2: ... Shot 3: ... Each shot gets its own camera. Do not describe one global camera move for the whole clip, because the point is that the framing changes between shots.

Treat each shot like a line on a shot list: subject, action, camera, and any dialogue or sound specific to that moment. Repeat the key identifying detail of your character in each shot (the green raincoat, the silver hair). That way the model does not drift between cuts.

Think about edit logic. Classic patterns work well: a wide establishing shot, then a medium, then a close-up; or action, then a reaction. Match your duration to the number of shots. Three or four shots fit naturally in 15 to 30 seconds, while a 5-second clip is best kept to one or two.

If you are planning a longer piece, Biggs Director lets you lay scenes out on a storyboard with per-shot camera body, lens, movement and genre presets, keep characters consistent and render the sequence. Seedance multi-shot is ideal for the moments inside those scenes.

Using reference images with @Image tags

Seedance 2.5 is a multi-reference model. When you attach reference images, refer to them in the prompt as @Image1, @Image2 and so on, in the order you uploaded them. For example: "@Image1 walks into the café and sits across from @Image2." The tag tells the model exactly which reference drives which part of the shot.

Give each reference a clear role. One image for the character, one for the product, one for the location is easier for the model to follow than three images of the same person from similar angles. Still describe what the person is doing and wearing in words. The reference handles identity, and the prompt handles action.

For recurring characters, Biggsfluencer generates a reference pack automatically: both sides, a profile, a laughing close-up and a full-body shot. Those images make strong references when you want the same face to show up across many clips. The AI Video studio also supports start and end frames, so you can animate a photo or lock both the first and last frame of a shot.

Four example prompts to adapt

Each example below follows the same pattern: shots labeled in order, concrete action, a camera per shot, quoted dialogue with speaker and delivery, and explicit sound. Swap in your own subjects and keep the structure.

Shot 1: Wide shot, static camera. A narrow Lisbon street at dawn, wet cobblestones, a tram bell in the distance. A cyclist in a yellow jacket rolls into frame and stops at a bakery. Shot 2: Medium close-up, slow push-in. The cyclist pulls off her helmet and smiles at the baker. She says, cheerful and out of breath, "Two of the usual, please." Ambient: street noise, birds, an oven door closing.

@Image1 stands behind a marble kitchen counter holding @Image2, a matte black coffee grinder. Shot 1: Close-up, locked-off camera. His hand twists the grinder lid; beans crunch loudly. Shot 2: Medium shot, gentle handheld drift. He pours the grounds into a filter and looks at the camera. He says, relaxed and confident, "Thirty seconds. That's all it takes." Sound: soft kettle hiss, morning birdsong outside.

Shot 1: Extreme wide aerial, slow descent. A lone hiker on a ridge above a sea of clouds at sunset, wind whipping her red scarf. Shot 2: Medium shot from behind, slow crane up. She plants her trekking poles and lifts her arms. Shot 3: Close-up on her face, static. She laughs and says, breathless, "We actually made it." Sound: strong wind, distant raven call, rustling fabric.

Shot 1: Medium two-shot, static, dim interrogation room lit by a single overhead lamp. A detective in a grey suit slides a photo across the table toward a young man in a hoodie. Shot 2: Close-up on the young man, slow push-in. His jaw tightens and he looks away. The detective says, quietly and coldly, "You were there." Shot 3: Close-up on the detective, static. He leans back and waits. Sound: buzzing fluorescent hum, a chair creaking.

Choosing resolution and duration

Price on Seedance 2.5 scales with both resolution and length, so pick settings for the job, not by default. For a 5-second clip with sound, 480p costs 184 credits (about $1.77 on Creator), 720p costs 395 credits (about $3.80) and 1080p costs 970 credits (about $9.33). An 8-second clip at 480p with sound costs 294 credits (about $2.83).

A practical workflow is to draft at 480p. Test your action, timing, dialogue sync and multi-shot structure there, since a broken prompt at 1080p is an expensive mistake. Once the shot works, move to 720p for social posts and previews, and 1080p for hero pieces where the extra detail is visible.

Match duration to content. Single-beat shots belong at 5 or 8 seconds. Save 15, 20 and 30 seconds for real multi-shot sequences where every shot earns its time. If you are unsure whether Seedance is the right model for a given idea, the AI Video studio can run one prompt on up to four models at once so you can compare takes and keep the best.

Questions

More guides