How to Write Seedance 2.5 Prompts: Differences from 2.0, Timestamped Shots and Reference Mapping

I. How Seedance 2.5 differs from 2.0

The official summary of Seedance 2.5 is "long storytelling × strong references × precise editing × multilingual." Every spec below comes from Volcengine Ark's official Seedance 2.5 and Seedance 2.0 series tutorials:

Which to use: 2.0 when you need 4K output; 2.5 when you need coherent clips longer than 15 seconds, more than 15 references, or per-second pacing.

II. The official four-part structure

The official framing: treat Seedance 2.5 as a visual content producer and write a structured prompt the way a director would. A complete prompt has four parts, in order:

Describe the picture positively; the negative phrasing the docs support is for subtitles and audio — e.g. "no subtitles", "no BGM".

III. Writing timestamps

On 2.0 the official approach is "Shot 1 / Shot 2", with the model pacing each shot from the story; 2.5 responds to timestamps in one-second units, so a 30-second story can be paced second by second. Three forms are supported:

• Ranges: "0–3s … 3–7s … 7–15s" or "[1s-4s] … [4s-8s] … [8s-12s]"
• Points: "at 5s, a fast whip-pan left as the transition"
• Relative: "he stands frozen; 3 seconds later the people around him shake their heads", "after the shutter clicks, the frame holds for 1 second."

Official cautions:

1. Keep the timeline continuous — avoid gaps like "0–3s … 5–6s"
2. Match content to duration — too little and the model improvises, too much and it over-cuts or drops beats
3. Do not use timestamps to control frequency, e.g. "shake head 3 times per second"
4. Either timestamps or "Shot N" are fine for splitting the beats.

IV. Reference mapping: the more references, the clearer "who is who" must be

The official mapping rules:

1. Number in upload order (Image 1 / Video 1 / Audio 1) and bind every reference once in text; use a list when there are many people
2. Do not rely on names written on the image — labeling a portrait "Zhang San" and then writing "Zhang San at school…" invites mix-ups or duplicated characters
3. Be specific about what is borrowed, and which part for partial references: "the spell-casting motion from Video 1, the orbiting camera from Video 2", "the lighting and filter of Image 1"
4. When a reference is already precise, just point to it instead of re-describing: "strictly follow the motion and camera of Video 1, in the same order"
5. With multiple characters, upload references in order of first appearance and renumber the prompt to match — this avoids character swaps.

V. Camera language, action and expression

VI. Dialogue, subtitles and sound

The official 2.5 tutorial keeps the same sound symbols as 2.0: music in ( ), sound effects in < >, dialogue in { }, subtitles in 【】. For non-Chinese dialogue, state the language before the line. Because 【】 means subtitles and ( ) means music, avoid those brackets for section labels or tone notes inside the prompt.

Negative control: for subtitles write "no subtitles" or "do not add dialogue captions"; audio can be controlled separately for SFX, BGM and dialogue — e.g. "no BGM, only ambience and action sounds", "no sound at all."

Still getting subtitles after "no subtitles"? The docs say this is much improved over 2.0 but cannot be fully avoided; to reduce it:

1. Do not repeat words after a line and tack on tone, expression or action descriptions
2. When a line needs emotion, split it as "Character line (emotion): content"
3. Make sure your reference video has no burned-in subtitles.

When on-screen text must be exact: split English words into separate letters and ask for them to appear in sequence; for Chinese, render the correct copy as an image, supply it as a reference, and cite that image number after the text description.

VII. Writing edit and extend prompts

2.5 infers the task type from the prompt, so edits and extensions need trigger words:

• Edit: include one of "edit video, add, delete/remove, modify/replace/change to" — e.g. "add some small animals to @video1", "change the person in @video1 to @image1", "remove the background music from @video1"
• Extend: include one of "extend forward/backward, continue" — e.g. "extend @video1 backward; the character from @image1 drops from the sky…".

For edits, state the scope and the change, optionally limited by timestamp: "in Video 1, change the man drinking coffee at 4–6s to mopping the floor; keep everything else unchanged." Official tips: source videos under 20 seconds edit best (they must be 4–30s), and using MOV for both input and output keeps color and audio continuity better when editing or extending.

VIII. A 30-second template you can adapt

Below is a text-to-video example in the four-part structure (pure text, so no reference mapping): a one-line summary, timestamped beats, then global notes. Dialogue, SFX and music are marked with { }, < > and ( ). Swap the subject, setting and lines to reuse it:

For reference-to-video, put reference mapping first — e.g. "Image 1 is the woman, Image 2 is the clerk, Image 3 is the store interior; Image 1 uses the voice from Audio 1" — and trim the character descriptions in the body accordingly.

IX. Troubleshooting at a glance

X. FAQ

Q: Can I reuse Seedance 2.0 prompts on 2.5? A: Structurally, yes — "Shot N" splitting and the ( ) < > { } 【】 sound symbols carry over. But the two versions have different aesthetic preferences, so the same prompt will not look the same; to use 2.5's longer length, swap shot numbers for timestamps and re-pace for 30 seconds.

Q: Should I write prompts in Chinese or English? A: Either. The docs list 11 supported languages for prompts and audio on 2.5; for non-Chinese dialogue, state the language before the line.

Q: What about videos longer than 30 seconds? A: 2.5 supports multiple rounds of extension: use keywords like "extend backward" or "continue" with the generated video as input. Or split the story into segments of 30 seconds or less, generate each, and cut them together.

XI. Generate shot prompts from a reference video with VideoLens

Writing a 30-second prompt by hand means laying out every second. VideoLens runs it in reverse: give it a reference video (upload, or paste a Douyin/Xiaohongshu link) and it breaks the clip down shot by shot, outputting shot prompts with anchored characters/locations/props, per-shot camera and action, dialogue and SFX. Both Seedance 2.0 and 2.5 are supported — pick one before uploading: 2.5 splits into 30-second blocks with up to 6 keyframes each, 2.0 into 15-second blocks.

A Seedance 2.5 prompt is essentially a timed shot script: state what each reference is for, set the tone in one line, write each time segment's picture, camera, action and sound, then add the constraints that hold across the clip. Your 2.0 habits still work on 2.5, so pick the version that suits the job — and if you would rather not start from scratch, give VideoLens a reference video and it returns a shot prompt split to the chosen version's length limit, ready to tweak.