Seedance Prompt Guide for Seedance 2.0 and 2.5

Seedance does what the words tell it, and fills in everything else on its own. Most disappointing renders come from prompts that leave the important parts to chance: no camera direction, no sound, three actions crammed into five seconds. This guide shows how to write prompts for Seedance 2.0, 2.0 Fast and 2.5, the models you can use today, with examples you can paste into the generator.

A note on Seedance 3: it hasn't been released yet (see our release date tracker). Nothing below depends on it, and the habits here are the ones every Seedance release so far has rewarded.

Start with one shot, described in five parts

Before you think about cuts or length, describe a single shot. Five parts, in this order, cover almost everything the model needs:

1. Subject: who or what, with one or two visual details. "An old fisherman in a yellow raincoat."
2. Action: one main thing that happens. "He hauls a net over the side of the boat."
3. Camera: where it is and how it moves. "Low angle from the deck, slow push in."
4. Light and look: time of day, weather, colour. "Grey dawn, mist on the water, muted colours."
5. Sound: what we hear. "Waves slapping the hull, gulls, the creak of rope."

Put together:

An old fisherman in a yellow raincoat hauls a net over the side of a small boat. Low angle from the deck, slow push in. Grey dawn, mist on the water, muted colours. Waves slapping the hull, gulls, the creak of rope.

That's 40-odd words, and every one of them is something the model can show or play. Leave out feelings ("he feels hopeful"), backstory ("after a long night at sea") and instructions the camera can't see.

Camera words that work

Seedance follows ordinary film language. These phrases give predictable results:

Use one move per shot. "Push in, then orbit, then pull back" in five seconds usually turns into a wobble.

Ask for the sound you want

Seedance generates audio in the same pass as the picture. If you don't describe it, the model picks something, often music. Three tips:

• Name two or three sounds that belong in the scene: "rain on a tin roof, a distant dog barking".
• Put dialogue in quotes and say who speaks: The girl whispers, "Did you hear that?"
• Say what to leave out, briefly, at the end: "no music". One short exclusion works; a long list of things to avoid tends to confuse the model.

Several shots in one clip

Seedance can cut between shots inside one clip. Write the shots in order, label them, and repeat whatever has to stay the same:

Shot 1: close-up of a young woman in a red scarf checking a paper map on a windy cliff path. Shot 2: wide shot from behind as the same woman in the red scarf walks toward a white lighthouse, grass bending in the wind. Wind roaring, a bell ringing in the distance.

Two shots fit comfortably in 5 to 8 seconds on Seedance 2.0. Three or more need 10 seconds or longer.

Timed beats for 30-second clips on Seedance 2.5

Seedance 2.5 makes up to 30 seconds per generation and is built to tell a short story with a beginning, a turn and an ending. With that much time, timestamps keep the pacing under control:

0–8 s: A delivery cyclist in a green jacket weaves through morning traffic, camera tracking alongside, horns and bike bells. 8–16 s: The same cyclist in the green jacket skids to a stop at a bakery, grabs a paper bag from the counter, the baker shouts "Go, go!". 16–24 s: Handheld shot following the cyclist up a narrow staircase, breathing hard, the paper bag rustling. 24–30 s: A door opens, an old man smiles and takes the bag, warm light from inside, quiet piano.

What makes this work:

• One main action per beat. The model has time to finish each one.
• The subject is restated ("the same cyclist in the green jacket"), which helps keep the face and clothes consistent across cuts.
• Sound in every beat, so the audio changes with the scene.

Longer clips cost more, so test the beats at 480P or at 15 seconds first, then render the final version.

Image-to-video prompts

In image to video, your picture is already the first frame. The prompt only has to describe what changes:

• Weak: "A woman with long dark hair in a blue coat stands on a bridge at sunset."
• Better: "She turns her head toward the river, wind lifts her hair, the camera slowly pulls back to show the bridge, traffic hum and water below."

If you add a last frame, describe the journey between the two images, not either image: "the empty room fills with furniture piece by piece as the camera stays still".

Reference prompts: give every file one job

In reference to video, uploads are named in order: Image 1, Image 2, Video 1, Audio 1. Say what each one is for:

The man from Image 1 sits at the kitchen table from Image 2, reading a letter. Copy the slow camera move from Video 1. Use Audio 1 as the background music. He looks up and says, "It's from her."

Rules of thumb:

• One or two clear photos of a face hold a character better than many mixed ones.
• A reference clip should show only the move you want copied.
• Name every file in the prompt. An unnamed upload is a file the model has to guess about.

10 prompts you can copy

Each one is written for a 5 to 10 second clip on Seedance 2.0 unless noted.

Product, 16:9

A matte black wireless speaker on a wet slate surface, water droplets jump on its top in time with a deep bass note, macro close-up, slow orbit, dark studio with one blue rim light.

Food, 9:16

Overhead shot of hands folding dumplings on a floured wooden board, one dumpling is placed into a bamboo steamer, steam rises, soft kitchen light, the quiet tap of fingers and a simmering pot.

Travel, 21:9

Drone shot rising over terraced rice fields at sunrise, low clouds drifting between the hills, a farmer walking along a ridge, birdsong and distant wind.

Dialogue, 16:9

Medium shot of two old friends on a park bench, autumn leaves falling. The woman laughs and says, "You still cheat at chess." The man shrugs and says, "Only when I'm losing." Soft afternoon light, children playing in the distance.

Sport, 9:16

A skateboarder pops an ollie over a concrete gap, the camera tracks low beside the board in slow motion, late afternoon sun flare, wheels clacking on the landing.

Nature, 16:9

A fox steps out of a snowy forest at dusk, pauses and looks straight at the camera, static camera, blue hour light, snow crunching, an owl calling.

Fantasy, 16:9

A paper lantern drifts up from a night market and turns into a glowing jellyfish as it rises over the rooftops, camera tilts up to follow it, soft synth pad, crowd murmur fading.

Social ad, 9:16

Shot 1: close-up of a tired office worker rubbing her eyes at a desk at night. Shot 2: she opens a can of iced coffee, a bright fizz, the room light warms. Shot 3: she smiles at the camera and says, "Okay, one more slide." Upbeat ukulele.

Image mode (upload a portrait as the first frame)

He looks up from the camera, a slow smile, the camera pushes in slightly, a light breeze moves his hair, café chatter in the background.

30-second story on Seedance 2.5, 16:9

0–10 s: A girl in a yellow raincoat launches a toy sailboat into a park pond, camera low at water level, rain pattering. 10–20 s: The wind pushes the boat out of reach, the same girl in the yellow raincoat runs along the bank, worried, splashing through puddles. 20–30 s: A dog swims out and brings the boat back in its mouth, the girl laughs and hugs the wet dog, the rain stops and sunlight breaks through.

When the result isn't right

A cheaper way to get the final clip

1. Write the prompt and render it on Seedance 2.0 Fast at 480P.
2. Fix the prompt using the table above until the motion and timing are right.
3. Render the keeper on Seedance 2.0 at 1080P, or on 2.5 if it needs more than 15 seconds.

Each render's price shows on the button before you generate, and failed renders are refunded. Current per-second rates are on the pricing page.

FAQ

How long should a Seedance prompt be?

Long enough to cover subject, action, camera, light and sound: often 30 to 80 words for one shot, more for timed 30-second clips. This site accepts up to 4,000 characters.

Do the same prompts work on Seedance 2.0 and 2.5?

Yes. They take the same kind of prompt. On 2.5, add timed beats when you use the extra length.

Should I write prompts in English?

The examples here are in English, the way we render them. The models accept other languages too; whichever you use, keep the descriptions concrete.

Will these prompts work on Seedance 3?

Seedance 3 isn't released, so nobody can say for sure. Our view: the five-part shot and timed beats are the most likely habits to carry over.

Seedance is a trademark of ByteDance. This site is an independent service and is not affiliated with or endorsed by ByteDance.