Seedance 2.5 prompt guide

The Seedance 2.5 prompt generator writes prompts in this order. For the model’s specs, see the Seedance 2.5 page.

BytePlus’s Seedance 2.5 tutorial orders a prompt as subject, action, scene, style, camera and cuts, then sound. The run sheet adds one part in front, how the clip runs, because on 2.5 that decides how the event is cut. Its prompt guide ends with what must stay the same.

1. 1Open with the frame, the length, and “one continuous take with no cuts” or the number of shots.
2. 2Name the subject with what must not change; with pictures, say which is which.
3. 3Write the event in the present tense, from start to end, in plain sentences.
4. 4Add the place, the style and the light, then each shot’s size and camera move.
5. 5End with the sound, and each line after its speaker with its language.

Seedance cuts between shots inside one generation: its Seedance 2.0 report calls multi-shot storytelling built in, and the Seedance 2.5 launch post says 2.5 links shots from setup to resolution. Mark shots with “Shot 1” labels or whole-second timestamps.

1. 1Give the total first, “15 seconds in three shots”, and the story in one sentence.
2. 2Label each shot with its span, “Shot 1 (0–4 s):”, and make the spans meet, with no gaps.
3. 3Give each shot a size, one camera move and its action; change the framing at every cut.
4. 4For a styled cut, say when and how: “At 8 s, a whip pan right.”
5. 5Close with what holds in every shot: subject, clothes, place and light.

• Allow about 3 seconds or more a shot: ByteDance warns that a crowded span brings extra cuts or drops a beat.
• Use timestamps for the story’s steps, not fast repeats: “nods three times a second” won’t land.
• Seedance 2.0 follows shot numbers but not timestamps, by ByteDance’s guide; 2.5 follows both.

How DreamBench’s film plan differs

One Seedance clip holds every shot in one generation: to fix Shot 3 you make the whole clip again, and it stops at 30 seconds. DreamBench writes a shot list instead, makes each shot as its own Seedance 2.5 clip of 3 to 12 seconds and cuts them together, so a film can run longer and one shot can be redone while the rest are kept.

On Seedance, a DreamBench shot may still hold a cut of its own, and each goes out with the references it needs, so the cast matches across the joins. See multi-shot video.

ByteDance’s guide says plain film terms work as written, from shot sizes to a long take or a dolly zoom. Follow a rarer term with what it looks like.

One move per shot; the next belongs after a cut. For lines and lettering, a locked-off camera keeps faces and words sharp.

Seedance 2.5 makes speech, effects and music with the picture. BytePlus’s tutorial lists 11 languages: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean. See native audio.

• Speaker, language, then the line: The chef says in Japanese, “まだ完璧じゃない。” ByteDance asks for the language before any line that isn’t Chinese.
• ByteDance’s examples put a line in curly braces, an effect in angle brackets and music in parentheses; quotation marks after a named speaker read more easily.
• Direct a line before it, “Her line (in disbelief):”, not after: ByteDance says words after a line can bring subtitles.
• “No BGM” can still let music in: name the words (music, score, melody) at the start and the end.
• A language off the list, such as German, isn’t promised: test a short line first.

On DreamBench, Seedance 2.5 has a sound switch, and a second costs the same with sound on or off.

The length is set before Seedance reads the prompt: too little and the model improvises, too much and it adds cuts or skips a beat.

On DreamBench, 480p costs the fewest credits a second: check that the shots land, then render the keeper at 1080p. Seedance 2.5 takes no seed, so that is a new clip, not a sharper copy. Frame it 1:1, 16:9, 9:16, 4:3 or 21:9, and name the frame in the prompt too.

ByteDance’s API takes up to 30 images, 10 videos and 10 audio clips; its guide finds 1 to 8 subjects in images work best. See reference to video.

• Number pictures in upload order and say what each is for: “@Image 1 is the florist; use @Image 3 for the light only.”
• Upload people in the order they appear, or characters can get mixed up, ByteDance warns.
• Bind names in the prompt, never by writing them on the picture.

On DreamBench a film takes up to 8 pictures, each tagged as a character, creature, location or prop; type @ and its label to call it, as in “@grandad hands @kite to the boy”. Every Seedance shot gets the sheets it needs with its prompt. Video takes text alone.

ByteDance offers a first frame, a first and last frame, or keyframes in order. A first frame sets the clip’s aspect ratio, and a last frame of another shape is stretched, so crop both alike. The prompt then describes only what moves. See first and last frame and image to video.

DreamBench exposes none of these for Seedance 2.5: Video sends text, and the film composer has no start or end slot. Your pictures go in as references, and the prompt decides how each shot opens and ends.

BytePlus’s tutorial says Seedance 2.5 doesn’t accept reference images or videos showing a real human face, so a photo of you or a customer is refused. Instead:

• Describe the person: age, hair, build, clothes.
• Let DreamBench design the character from your words; Seedance can still refuse one as a possible real person.
• Switch model: DreamBench names the refused reference and suggests another video model.

Seedance carries a subject across its own cuts; across clips, only what you repeat carries over. See character consistency.

• Describe each character in identical words in every clip: “a florist in his sixties with a white mustache and a green apron”.
• Keep the cast small: ByteDance says scenes where several subjects interact aren’t fully stable yet.
• Keep one model: by ByteDance’s guide, 2.0 and 2.5 give one prompt a different look.
• The Seedance 2.5 page says ByteDance’s tools can extend a clip twice; on DreamBench the next shot gets the same references.

Paste these into DreamBench’s Video page or any tool that runs Seedance 2.5.

One take: the wok stall

Widescreen 16:9, one continuous 20-second take with no cuts. At a night-market stall, a cook in a black cap tosses noodles in a flaming wok, plates them and slides the bowl to a waiting girl, who grins. The camera starts on the flames and pulls back to the counter. Warm bulbs, blue night. Roaring burner, crowd chatter; no music.

Three shots: the beekeeper

Widescreen 16:9, 15 seconds in three shots. Shot 1 (0–4 s): wide, a beekeeper in a white suit walks between hives in a lavender field. Shot 2 (4–10 s): medium, slow push in as she lifts a frame of bees. Shot 3 (10–15 s): close-up, honey runs onto her glove. Low gold sun from the left. Humming bees; no music.

Two speakers: the mountain hut

Widescreen 16:9, 12 seconds in three shots, a mountain hut at night. Shot 1 (0–4 s): wide two-shot, an old guide in a red wool hat pours tea for a young climber. Shot 2 (4–8 s): the guide says in English, “We leave at four.” Shot 3 (8–12 s): the climber nods: “Then I’ll sleep now.” Firelight, wind outside; no music.

Two shots, a shout in Portuguese

Vertical 9:16, 10 seconds in two shots. Shot 1 (0–6 s): low tracking shot as a girl in a yellow shirt dribbles past two boys on a dusty field. Shot 2 (6–10 s): close-up as she scores and shouts in Portuguese, “Golaço!” Hard afternoon sun. Children shouting, the ball in the net.

A line in Indonesian: the satay stall

Vertical 9:16, one continuous 8-second take. A woman in a green headscarf fans the coals at a satay stall and holds out a plate, saying warmly in Indonesian, “Silakan, masih panas.” Locked-off medium shot, smoke in orange dusk light. Sizzling meat, scooters; no music.

Product with lettering: the coffee tin

Square 1:1, one continuous 10-second take. A hand lifts the lid of a matte black coffee tin on slate and steam rises. Slow push in as the white letters “N”, “O”, “R”, “D” appear one by one on the label. Soft light from the left. The pop of the lid; no music, no subtitles.

Timestamps: the surfer

Widescreen 16:9, one continuous 10-second take. 0–3 s: a surfer in a black wetsuit paddles into a green wave and pops up. 3–7 s: she carves down the face, spray fanning. 7–10 s: she kicks out and raises a fist. Drone shot tracking alongside. Morning sun. Surf, wind; no music.

Music: the cellist

Vertical 9:16, one continuous 15-second take. A cellist in a long gray coat plays alone on an empty train platform at dawn. The camera orbits her once, ending on her face. Pale blue light. (Music: a slow solo cello melody that builds.) A distant train horn.

Social hook to camera

Vertical 9:16, one continuous 5-second take. On a sunny balcony, a young man holds a tiny houseplant up to his phone and says, “This plant survived three house moves. Here’s how.” Handheld selfie framing. Street noise; no music, no subtitles.

Stylized: the clay fox

Widescreen 16:9, one continuous 10-second take, stop-motion clay style. A clay fox in a red scarf builds a snowman, sets a pebble nose on it and steps back proudly; the nose falls off. Locked-off at the fox’s height. Soft studio light. Crunching snow, a surprised yelp.

With references: the florist

Widescreen 16:9, 15 seconds in three shots. @Image 1 is the florist, a man with a white mustache and a green apron; @Image 2 is his shop. Shot 1 (0–5 s): wide, he locks the shop in light rain. Shot 2 (5–10 s): tracking, he cycles off with a bouquet. Shot 3 (10–15 s): close-up, an old woman opens her door and smiles. Rain, a bicycle bell.

On DreamBench, the last one belongs in a film, with your labels, such as @florist, in place of the numbers.

• One prompt run twice gives two clips; Seedance 2.5 takes no seed.
• ByteDance says complex motion can look physically off, and multi-subject scenes aren’t fully stable.
• Subtitles can appear despite “no subtitles”: ByteDance’s fixes make them rarer, not impossible.
• One generation stops at 30 seconds; Video refuses a prompt past 6,500 characters.

• One clip: on Video, pick Seedance 2.5, paste the prompt, choose 4 to 30 seconds, the frame, the resolution and sound. Your shot labels reach the model as written.
• A prompt: the Seedance 2.5 prompt generator writes one take or two or three timed shots. It writes the prompt only.
• A film: describe it, add up to 8 pictures and pick Seedance 2.5; DreamBench plans, casts, makes and cuts the shots, and you can redo any one.
• Not offered here: first and last frames, video or audio references, extending a clip, or editing one by timestamp.

The composer shows the price in credits before you generate; the first time, create a free account and choose a plan. For Kling, see the Kling 3.0 prompt guide; for prompts from words alone, text to video.