Seedance Prompts: Syntax, Shots and Examples for Seedance 2.0 and 2.5

What Seedance is and which versions are current

Seedance is ByteDance's family of AI video models. As checked in November 2026, two versions are in use: Seedance 2.0, which comes in standard, fast and mini variants, and the newer Seedance 2.5. You reach them through BytePlus ModelArk, Volcengine Ark, Dola and BytePlus Playground, and the Runway API also offers Seedance, among other platforms.

• Seedance 2.0: 4–15 second clips; 480p and 720p (1080p and 4K on the standard model only); 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive aspect ratios.
• 2.0 modes: text to video, first frame, first and last frame, omni reference (up to 9 images, 3 videos and 3 audio files), editing, and extension forwards, backwards or by joining up to 3 clips.
• Seedance 2.5: 4–30 second clips, up to 1080p in 10-bit, up to 50 references and dialogue in 11 languages.

Limits, available modes and pricing depend on the platform you use and change often. Check them there before planning a long or high-resolution video.

How a Seedance prompt is structured

The official guide recommends a specific order. Sticking to it keeps the model from skipping things and tells you what to adjust when a result goes wrong:

• Precise subject and action: who is in the shot and what they do.
• Scene: place, era, weather, key objects.
• Light and colour: time of day, quality of light, palette.
• Camera: shot type and movement.
• Style: realistic, documentary, animation and so on.
• Quality: level of detail and sharpness.
• Constraints: what must not appear, written as a sentence (there is no negative field).

Give each subject two or three stable visual traits (a yellow raincoat, a short grey beard, a brass lantern) and always call it the same thing, for example 'the keeper'. Change the label or the traits between shots and the model may treat it as a different character.

Seedance 2.0 versus 2.5: shots or a timeline

This is the practical difference between the versions. In Seedance 2.0, a multi-shot video is written as 'Shot 1: … Shot 2: … Shot 3: …', with a single camera move per shot. Exact timestamps such as '0-3s' are unreliable in 2.0, so leave them out.

Seedance 2.5 does understand timing: whole-second segments with no gaps ('0-3s… 3-7s…') or a single moment ('at the 5-second mark'). The recommended shape is a one-sentence summary (subject, place, event, genre and camera), then the timed plot, then coherence notes. It also accepts images of the subject from several angles and storyboards of up to 15 panels.

Shot 1: [subject with 2-3 traits] [action] in [place]; [one camera move]. Shot 2: [same subject label] [next action]; [one camera move]. [Light and colour], [style], [quality]. <[sound effects]> （[music]） [Label] says {[dialogue]} [Constraints: avoid generating any text or subtitles.]

[One-sentence summary: subject + place + event + genre + camera]. 0-3s: [what happens]. 3-7s: [what happens]. 7-10s: [what happens]. <[sound effects]> [Character]'s line ([emotion]): {[dialogue]} Coherence: [traits that stay the same]. [Constraints: No BGM.]

Seedance's own syntax: dialogue, audio, text, references and constraints

Seedance uses literal brackets to separate audio and text from the visual description. Copy them exactly, full-width parentheses included:

• Music in （ ）: （soft piano melody）.
• Sound effects in < >: <rain on a tin roof>.
• Dialogue in { }: The baker says {It's ready}.
• Subtitles in 【 】: 【Open since 1962】.
• Language: one language per line of dialogue, and don't mix Chinese and English. For a less common language, say so: says in Japanese {…}.
• In 2.5, dialogue can carry an emotion, as in Baker's line (cheerful): {…}, and voice-over is marked (V.O.).

References are named in upload order: 'Image 1', 'Video 1'. State each one's role ('Use Image 1 as the character's appearance') and be strict when editing: 'Strictly edit Video 1'. The guide suggests a mix of 4 or 5 assets and no more than 4 people. Seedance 2.0 does not directly accept reference images or videos with real human faces, so use fictional characters and material that is yours or licensed.

There is no negative prompt field. Ask for what you don't want with constraint phrases at the end: 'avoid generating any text or subtitles', 'do not generate a watermark/logo', 'no twin effect'. In 2.5, 'No BGM' (no background music) and 'generate only environmental and action sounds' also work.

Camera and movement

Seedance understands plain-English film vocabulary: medium shot, close-up, camera fixed, slow push-in. The rule that matters most is one move per shot. Ask the camera to push in and orbit the subject in the same shot and it will usually do one of them badly, or neither.

• Static: camera fixed.
• In or out: the camera slowly pushes in / pulls back.
• Tracking: the camera follows behind the subject.
• Need two moves? Split them into two shots (2.0) or two time segments (2.5).

Examples to copy

Every example uses an invented subject. Swap in your own details and keep the structure.

Shot 1: The keeper, an old man with a yellow raincoat, a short grey beard and a brass lantern, climbs the spiral stairs of a stone lighthouse at night; the camera follows behind him. Shot 2: The keeper (yellow raincoat, grey beard) pushes open the top hatch onto a stormy sky; the camera slowly pulls back. Shot 3: Close-up of the keeper's wet face lit by the lantern; camera fixed. Cold blue night light with warm lantern highlights, realistic cinematic style, high detail. <wind howling, metal hatch creaking> （low cello drone） The keeper says {Here comes the storm!} Avoid generating any text or subtitles. Do not generate a watermark or logo.

A retired clockmaker resets the tower clock of a mountain village at dusk, a quiet documentary-style drama with a slow handheld camera. 0-3s: Close-up of the clockmaker's hands in green fingerless gloves turning a brass gear; golden light through a small window. 3-7s: Medium shot, the clockmaker (round glasses, grey wool vest, green gloves) checks his pocket watch and nods. 7-10s: The camera tilts up to the clock face as the hands click into place. <gears ticking, a distant church bell> Clockmaker's line (quiet, satisfied): {Right on time again.} Coherence: the clockmaker keeps the round glasses, grey vest and green gloves throughout. No BGM. Avoid generating any text or subtitles.

Use Image 1 as the first frame. The ceramic fox figurine on the shelf slowly turns its head toward the window as the morning light grows warmer; camera fixed. Soft natural daylight, handmade stop-motion look, fine clay texture. <a soft clink of ceramic> Avoid generating any text or subtitles. No twin effect.

A young cyclist with a red helmet and a canvas messenger bag delivers bread through a sleepy coastal town at sunrise, a light-hearted short with a tracking camera. 0-4s: Wide shot, she rides along the seafront; the camera tracks alongside. 4-8s: Medium shot, she hands a paper bag to a fisherman at a blue door. Narrator (V.O.): {Same route, every morning.} 【7 a.m. delivery】 <seagulls, bicycle bell> （gentle acoustic guitar） Coherence: red helmet and canvas bag in every shot.

Strictly edit Video 1. Change the season from summer to a snowy winter evening: snow on the rooftops and the street, warm light in the windows. Keep the camera movement, the framing and the pedestrians unchanged. <muffled footsteps on snow> Do not generate a watermark or logo.

Common mistakes

• Writing a negative list: Seedance has no such field, so use constraint phrases.
• Using exact timings in Seedance 2.0, or stacking several camera moves in one shot.
• Renaming the character or changing their traits between shots, which encourages face drift. One close-up reference plus one full-body reference helps.
• Dropping the brackets: dialogue without { } or music without （ ） bleeds into the visual description.
• Mixing languages within one line of dialogue.
• Chaining many extensions: seams can show at the joins and quality tends to drop.
• Expecting perfect text: in 2.5 text can come out misspelt (spell it out or supply an image of it), and subtitles sometimes appear even when you ask for none.

Using the PromptNow Seedance prompt generator

If you'd rather not assemble the prompt by hand, the PromptNow Seedance prompt generator does it for you. Pick the version (2.0 or 2.5), mode, number of shots, camera, lighting, length, aspect ratio and audio, and you get a prompt to paste into ChatGPT, Claude or Gemini. It returns the final prompt in that version's syntax, with the tool settings listed separately, two variants and what to check in the tool.

Don't use real people's faces, voices or names, or copyrighted characters or music, and label the video as AI-generated wherever the platform or the law requires it.