FlyAIgh
Home/Blog/Guide

How to Write Seedance 2.5 Prompts: The Official Rules, Checked Against a Real 25-Second Clip

Published November 19, 20268 min read

One 25-second Seedance 2.5 prompt, seven timecoded beats, every beat within about a second or two of the script. Here are the rules from ByteDance’s own guide behind that, where our prompt broke them, and a template to copy.

Seedance 2.5 reads prompts differently from 2.0: it follows whole-second timecodes, takes up to 30 reference images, and generates sound with the picture. This guide condenses ByteDance’s official Seedance 2.5 prompt guide into steps you can follow, and checks the most important rule against a 25-second clip we generated.

The short answer

Write a Seedance 2.5 prompt in 4 parts, in this order:

  1. Asset map (only if you upload references): name every image, video and audio file by upload order and say what it is for. “Image 1 is the knight. Image 2 is his sword. Video 1 is the camera move to follow.”
  2. One-line summary: subject, place, event, style and camera, in one sentence.
  3. Timecoded beats: split the clip into whole-second ranges with no gaps (0–3s, 3–7s, 7–12s) and describe the action, camera move, dialogue and sound in each.
  4. Consistency notes: what must stay the same throughout, such as lighting, lens, location and sound.

Keep it to about 500 Chinese characters or 1,000 English words, describe what you want rather than what you don’t, and save negative phrasing for subtitles and audio (“no subtitles”, “no background music”). These rules come from ByteDance’s documentation; the sections below show what they look like in practice and where our own run followed or broke them.

What changes from Seedance 2.0

  • Timecodes work. ByteDance’s documentation says Seedance 2.0 does not respond to timestamps and only follows shot numbers, while 2.5 accepts whole-second timestamps. A “0–3s / 3–6s” prompt is a real timeline on 2.5.
  • One clip can run 30 seconds, long enough for a whole sequence, so a prompt now describes a progression instead of a single moment.
  • References scale up: up to 30 images, 10 videos and 10 audio clips per request on Seedance 2.5 (10 / 5 / 5 on 2.5 flash). With that many files, numbering them in the prompt becomes essential.
  • Multi-view character sheets are supported as subject references in 2.5; the 2.0 guide advised against them.

Expect a different look as well. ByteDance’s FAQ notes that 2.0 and 2.5 produce visibly different styles from the same prompt and image. Their suggested workaround for keeping 2.0’s style is to extend a 2.0 clip with 2.5 and choose MOV output, instead of regenerating it.

The structure, on a prompt we ran

In August we ran a 25-second single take on Seedance 2.5 at 720p and 16:9, with no reference images: a young flyer chasing a chained storm dragon off a floating market island. The Chinese prompt follows the structure above:

PartWhat the prompt said (summarized)
Style and consistency (placed first)Live-action fantasy look, ARRI camera, 35 mm anamorphic lens, shallow depth of field, golden-hour light with teal shadows, low backlight, volumetric cloud light
Beats0–3s the chain snaps · 3–6s the loose chain smashes fruit stalls · 6–10s rooftop chase with the camera tracking alongside · 10–12s dragon and girl leap off the cliff · 12–16s freefall past the island’s hanging roots · 16–19s she lands on the dragon and both punch into cloud · 19–25s they burst above the clouds and the camera pulls back to a wide shot with airships
Continuity lineOne continuous shot with no cuts or transitions; actions and camera moves flow from range to range
The complete 25-second Seedance 2.5 output, recompressed for the web and otherwise unedited.

All 7 beats appeared, each within about 1 to 2 seconds of its scripted range, including the physical event at 3–6s and every element listed for the final wide shot. The beat-by-beat check is in our Seedance 2.5 vs Wan 3.0 test.

The full prompt, verbatim (Chinese)

Published exactly as it was sent. The order differs slightly from the template below: the style and consistency block comes first, then the beats, then one continuity line.

整体风格:好莱坞奇幻大片实拍质感,写实风格CG,电影感构图,ARRI摄影机拍摄,35mm变形宽银幕镜头,浅景深;整体色调:黄昏暖金色为主调,阴影偏青蓝色,金橙与青蓝互补色对比;光线:低角度夕阳侧逆光,轮廓光明显,云层体积光;质感:皮革、鳞片、石材纹理细节丰富,材质有真实反光;胶片颗粒感轻微,柔和光晕,高动态范围,无过度锐化,参考《阿凡达》《驯龙高手》真人版的影像质感。史诗奇幻电影,一镜到底无剪辑,25秒连续镜头。架空世界:漂浮在万米云海之上的浮空岛屿,岛上是石砌市集小镇,岛底垂挂巨大岩石根系与瀑布。主角:16岁东方少女,深色短发,额头推着黄铜圆框飞行风镜,穿棕色皮质飞行短装;巨龙:二十米长青蓝色鳞片风暴龙,颈部有深色鬃鳞和半截断裂铁链,半透明翼膜泛金光。两者形象全程严格保持一致。第0-3秒: 市集中景,巨龙被粗铁链锁在石柱上,周围商贩叫卖、行人穿行;巨龙猛然暴起长啸,铁链崩断,铁环飞崩迸出火星。
第3-6秒: 断裂的锁链横扫市集,砸翻一排水果货摊,果子滚落、帆布棚飞起,人群惊叫四散;镜头快速向右横摇,跟住巨龙冲进石板街道。
第6-10秒: 巨龙拖着半截断链沿狭窄街道狂奔,撞开木桶和吊灯笼;少女逆着逃散人流翻身跃上屋顶,在连绵赤陶瓦面上全速奔跑追赶;镜头与她平行横移跟拍,背景云海与远处浮岛快速掠过。
第10-12秒: 巨龙冲到浮岛断崖边缘,张开双翼纵身跃下悬崖俯冲入云;少女紧随其后从岛缘飞身跃下。
第12-16秒: 镜头紧跟少女垂直向下俯冲,自由落体,画面中浮岛底部倒悬的岩石根系、藤蔓和细瀑布飞速上掠,流云扑面,强烈速度感;少女收紧身体,逐渐逼近下方展翼滑翔的巨龙。
第16-19秒: 少女稳稳落在巨龙背上,双手抓住颈部鬃鳞,巨龙猛然振翅拉起;镜头围绕龙身环绕半圈,两者一同扎入厚厚云层,云雾扫过镜头,能见度骤降。
第19-25秒: 冲出云海的瞬间画面骤亮,金色夕阳铺满天空;镜头缓缓后拉并升高成大全景:少女骑在振翅巨龙背上,前方无尽云原悬浮着大小浮岛,鲸形飞艇在云间缓缓游弋,飞鸟群掠过,逆光剪影,史诗感定格。
全程一个连续镜头,无剪辑无转场,各时间段动作与运镜自然衔接,锁链断裂与货摊翻飞有真实物理细节,云层有体积光,电影级动态光影。

What it does not prove: this is a single run of a single prompt. It shows that Seedance 2.5 can follow a 7-range timeline closely, not how often it does.

Timecodes the model follows

  • Use whole seconds and cover the whole clip: “0–3s … 3–7s … 7–12s”, not “0–3s … 5–6s”. A gap invites the model to improvise.
  • Balance each range. Too little and the model fills it in its own way; too much and you get extra cuts or dropped actions. Each range in the dragon prompt holds a main action and a camera instruction.
  • Point events can use “at the 5-second mark”, and relative timing such as “3 seconds later” also works.
  • Don’t timecode rapid repetition (“shakes her head 3 times a second”). Describe the action generally instead.
  • For a transition, give the moment and the method: “At 5s, a fast whip pan left into a dissolve.”

Numbering reference files

When you upload references, the prompt has to say which file is which. ByteDance’s guidance:

  • Number by upload order (Image 1, Image 2, Video 1, Audio 1) and bind each one in the text. Don’t rely on a name written inside the picture: writing “John” on an image and then “John walks in” in the prompt is a known cause of swapped or duplicated characters.
  • Say what to take from each file: “the spell-casting motion in Video 1 and the orbiting camera in Video 2”, “the lighting and color of Image 1”.
  • Don’t re-describe what a reference already shows. If Video 1 contains the exact camera move, write “follow the camera movement in Video 1” instead of narrating it.
  • With several characters, upload them in the order they first appear and number them in that order. ByteDance lists mismatched characters as a common failure and this as the fix.
  • Stay within the stable range: 1 to 8 subjects in images. Up to 12 can work with more retries; past 5 subjects, separate single-view images are steadier than one multi-view sheet.

On FlyAIgh’s Generate page, reference images are sent in slot order, so Image 1 is the first slot. In Canvas, each card in a video’s asset panel has a name you @-mention in the prompt, and only the cards you mention are sent, in panel order.

Negative phrasing, sound and subtitles

Seedance 2.5 generates audio with the picture, so write sound into the beats: ambience, effects and dialogue, in the language you want spoken. ByteDance lists native dialogue in more than 10 languages.

  • Keep negative phrasing for subtitles and audio: “No subtitles.” “No background music; only ambient and action sounds.” “No audio.” Elsewhere, describe what should appear.
  • Subtitles can appear even when you forbid them. ByteDance says this is reduced in 2.5 but not eliminated. Their mitigations: don’t repeat words after a line or attach separate tone and action notes to individual words, write lines as Character’s line (emotion): “content”, and use reference videos without burned-in subtitles.

Some widely read guides recommend long lists of exclusions throughout a prompt (“no modern objects, no music, no dialogue”). ByteDance’s guide recommends describing what you want everywhere except subtitles and audio. Labeled sections are fine; if an exclusion isn’t about subtitles or sound, try rewriting it as a positive description of what belongs in that spot.

How long a prompt should be

ByteDance’s API reference recommends keeping a prompt to about 500 Chinese characters or 1,000 English words; past that, it says, the model tends to keep the main points and drop details. It is a recommendation, not a hard cap, and FlyAIgh sets no lower limit of its own for Seedance 2.5.

Our dragon prompt runs to about 930 Chinese characters, nearly twice the recommendation, and its 7 beats still landed. We did not run the same idea at a shorter length, so this run cannot show what the extra length cost. Start under the limit and add detail only where a run shows you need it.

Common failures and the official fixes

What goes wrongByteDance’s fix
Glowing eyes after strong emotional wordingPut “normal human eyes, no glowing eyes” first among constraints, and replace intense words such as “fanatical” or “extremely shocked” with neutral ones such as “amazed”.
A character takes another character’s face or voiceUpload references in order of first appearance and renumber the prompt to match.
On-screen text is misspelledSpell the word out letter by letter, or supply the exact text as an image reference.
Subtitles despite “no subtitles”See the section on negative phrasing, sound and subtitles.
2.5 output looks different from your 2.0 clipsExtend the 2.0 clip with 2.5 and MOV output instead of regenerating it.
An edited clip is about 0.3 s shorter than the sourceUse reference mode with a set duration, or trim the input to 8n+1 frames.

Edits and extensions also need a trigger word in the prompt (“replace”, “remove”, “add” for edits; “extend”, “continue” for extensions), and they lock the output aspect ratio to the source video.

A template to copy

Seedance 2.5 prompt template
[Asset map — leave out if you have no references]
Image 1 is <character>. Image 2 is <prop>. Video 1 is the camera movement to follow.

[Summary]
<Subject> <does what> in <place>, <style>, <camera approach>.

[Beats]
0–3s: <action>. Camera: <move>. Sound: <ambience / effect / line>.
3–7s: <action>. Camera: <move>. Sound: <…>.
7–12s: <action>. Camera: <move>. Sound: <…>.

[Consistency]
Same lighting, lens and location throughout; <character> keeps <key features>.
No subtitles. No background music; ambient and action sounds only.

Give each range a single main action. If a beat keeps getting dropped, give it more seconds before you give it more words.

Running Seedance 2.5 on FlyAIgh

Seedance 2.5 is available in Generate, Canvas and Director: clips from 4 to 30 seconds, 480p to 4K (1080p and above are upscaled from a native 720p render), aspect ratios from 9:16 to 21:9 plus adaptive, up to 30 reference images, and native audio. The price is shown before you generate; at 720p it is 15.3 credits per second.

Frequently Asked Questions

Do Seedance 2.5 prompts work on Seedance 2.0?

Partly. The structure carries over, but ByteDance’s documentation says 2.0 does not respond to timestamps and follows shot numbers instead. A timecoded 2.5 prompt loses its timing on 2.0; write “Shot 1, Shot 2” there.

Is there a character limit for Seedance 2.5 prompts?

ByteDance documents a recommended length rather than a hard cap: about 500 Chinese characters or 1,000 English words. Longer prompts tend to lose details. FlyAIgh does not set its own lower limit for Seedance 2.5.

How many reference files can one Seedance 2.5 prompt use?

Officially up to 30 images, 10 videos and 10 audio clips on Seedance 2.5, and 10, 5 and 5 on Seedance 2.5 flash. Seedance 2.5 on FlyAIgh accepts up to 30 reference images.

Is there an official Seedance 2.5 prompt generator?

ByteDance publishes a prompt-optimization skill for Seedance 2.5 (sd25-pe) that you install in an AI coding assistant and call before generating. We have not tested it, so treat it as a starting point and check the output against the rules here.

Can I make a Seedance 2.5 video longer than 30 seconds?

Not in one generation. Use an extension task, with a phrase such as “extend” or “continue” in the prompt, to continue an existing clip forward or backward.

Why do subtitles appear when I wrote “no subtitles”?

ByteDance says this is reduced in 2.5 but cannot be fully eliminated. Repeating words after a line of dialogue and reference videos that already contain subtitles are the usual triggers.

Write a Seedance 2.5 prompt

Prices and capabilities for each model are on their own pages, and the price of a run is shown before you start it.

Open Seedance 2.5