Seedance 2.5 answers a prompt differently from the generation before it. Two changes matter before you write a word: it reads timestamps in whole seconds, and it can take a blockout video as the only source of camera, motion and lighting. Both are documented on the vendor prompt guide.
What prompt control changed from Seedance 2.0
The guide lists four differences from the previous generation, and all four are about what a prompt can address rather than about wording. Timestamps: 2.0 answers shot numbers only, 2.5 answers whole seconds. Multi-view: 2.0 discouraged it as a subject reference, 2.5 supports it. Aspect ratio: 2.0 offers six fixed values, 2.5 accepts anything from 0.4 to 2.5 through the input material. And MOV output, which holds colour and brightness steady in edit and extension work.
Three of the four are input-side controls, and that is the pattern: 2.5 guesses less about what you meant. The parameter comparison covers the limits these controls operate inside.
Timestamps: whole seconds, contiguous coverage
Four forms of the same idea appear in the official examples: an interval pair such as 0s-3s and 3s-8s, a bracketed form such as [0-10秒], a shot-anchored form such as 0-3s（首帧参考 <2pic>）, and a shot list where each 镜头 carries its own range. The skeleton looks like this.
写实自然纪录片风格，电影级真实光影，温暖的午后, 在森林草坡上, 一只圆滚滚的熊猫幼崽从坡上滚下来.
0s-3s：<第一段发生什么，连同它自己的机位与光线>
3s-8s：<第二段发生什么>
<全局约束：机位高度、环境音、不要出现什么>Two habits come out of those examples. The intervals should cover the clip without gaps, since each beat carries its own visuals, action, camera and sound. And a segment describes what happens across a stretch of time, never how often something happens per second. The how-to guide lands these beats in a request.
White-model references: coarse geometry, torso only
The vendor splits blockouts into two granularities. A coarse blockout carries motion, camera, path and lighting information and is meant for locking trajectories, entrances and exits, camera paths, cut positions and lighting changes; the guide asks for 仅用简单几何体拼接表示人物/物体/动物等. A fine blockout is a complete model structure whose materials, colours or style get re-rendered.
One boundary is easy to miss and costs generations: 白模视频主体建议仅保留躯体，避免包含【四肢/翅膀】等涉及精细运动的元素；如已包含，需补全其动作序列，以防止渲染结果出现四肢僵化。 Blockouts can also be paired with reference images for subject, scene and prop. The full published prompt appears in example two below.
Storyboards and keyframes are different contracts
The two are not interchangeable. For multi-panel boards the guide says 生成画面不是严格对齐分镜图, treats the grid as 大致的剧情参考, and recommends simple line art with the prompt filling in what the board leaves out. For keyframes it says 生成画面对齐关键帧 and requires an ordering sentence first: 以图片 x 至图片 x 的顺序作为关键帧, as in the published 以图片 1 至图片 7 的顺序作为关键帧. If strict alignment matters, the grid is the wrong tool.
How much material to feed
The guide carries a recommendation table, and its numbers beat any general advice about reference quality.
| Input | Recommendation |
|---|---|
| Images | up to 30, each within 4K |
| Video | up to 10 clips, 30 seconds total |
| Audio | up to 10 clips, 30 seconds total |
| Subject images | 1 to 8 subjects works well, 9 to 12 is less stable |
| Subject video and audio | 1 to 5 subjects, 5 to 10 seconds each |
| Multi-view subject images | beyond 5 subjects, split views into separate files |
| Storyboard grid | under 15 panels, line art rather than text-heavy |
| Blockout | coarse, built from simple geometry |
| Video to edit | under 20 seconds |
| Reference images for a video edit | 1 to 5 works well, 6 to 8 is less stable |
| Extension input and output | MOV for the best audio and picture match |
Past the comfortable range the vendor does not say the input fails, only 稳定性下降、可能需要抽卡. Feeding more material does not move the bill either, because billing runs on output tokens, which the price breakdown sets out in full.
Worked examples: three prompts quoted in full
Example one: an eight-second panda clip, text only
Five parts make up this prompt: the register, subject and scene, two timestamped beats, and one paragraph of global constraints. The parameters are duration: 8 and ratio: 16:9.
写实自然纪录片风格，电影级真实光影，温暖的午后, 在森林草坡上, 一只圆滚滚的熊猫幼崽从坡上滚下来.
熊猫黑白毛发蓬松真实，体型小而胖，动作笨拙可爱。场景为绿色森林斜坡，地面覆盖青草、苔藓、三叶草、泥土、小石块、枯枝和少量小黄花，虚化背景中有高大树干与树林。镜头为低机位中远景，轻微手持感，整体基本固定，始终保持熊猫在画面中。
0s-3s：一只熊猫幼崽趴在绿色草坡上，身体圆滚滚, 它刚开始顺着斜坡慢慢侧滚，动作笨拙，草叶被身体轻轻压弯。微风拂过, 阳光从左上方穿过树林，形成斑驳光影。
3s-8s：熊猫滚到画面右下方，动作逐渐停止，从侧躺变成趴卧。它的圆脸朝向镜头，前爪压在草地上，熊猫趴在前景草丛中，身体调整到舒服的姿势，头部小幅抬起又放低, 发出轻微哼唧声。
低机位，轻微手持感，轻微跟随熊猫向右下方移动。景深自然，前景草叶轻微虚化，熊猫主体清晰，背景树林柔和虚化。自然环境音，风声、熊猫滚动时轻微柔软的扑通声, 整体温暖、真实、自然。{
"model": "doubao-seedance-2-5-260628",
"content": [{ "type": "text", "text": "<the prompt above>" }],
"ratio": "16:9",
"duration": 8
}The lesson is the last paragraph. Everything that applies to the whole clip sits in one place at the end rather than repeating inside every beat.
Example two: a blockout plus ten keyframes, thirty seconds
A blockout video paired with ten keyframe images for a 30-second short. The first sentence hands the blockout one job, the timeline then walks segment by segment, and the close names the output specification.
以白模参考视频 <video1> 作为整支视频唯一的运镜、镜头节奏、景别变化、主体运动轨迹和镜头调度参考，严格保持白模视频的镜头顺序、机位变化、运动方式和节奏，不改变镜头结构，不新增镜头，不改变主体运动逻辑。结合各阶段关键帧参考图，生成一部 30 秒电影级 3D 动画短片，整体风格梦幻、童话、温暖，具有儿童幻想色彩，角色外形与各阶段的关键帧保持一致，不要改变角色形象，人物表情情绪随场景变化而改变。
- 0-3s（首帧参考 <2pic>）镜头由全景俯视缓缓推向地上的小女孩，小女孩坐在房间的地毯上玩飞机，小女孩起身左转，右手使劲一挥将手里的飞机放飞。
- 3-5s（参考 <3pic>）飞机从左至右穿过房间悬挂的星星挂件，小女孩乘坐飞机进入幻想的天空。
- 5-8s（参考 <4pic>）镜头继续侧跟，以小女孩为中心的环绕。
- 8-10s（参考 <5pic>）镜头环绕到小女孩驾驶的飞机背面、飞机开始缓缓俯冲海面。
- 10-19s（参考 <6pic>、参考 <7pic>）小女孩继续向海底游去，一条鳐鱼游过来入画。
- 19-23s（参考 <8pic>）小女孩从时空裂缝冲出来到幻想宇宙。
- 23-24s（参考 <9pic>）前景小女孩和星球开始前翻转并，逐渐幻化消失。
- 24-28s（参考 <9pic>）俯视镜头持续推进，小女孩躺在地的毯上熟睡。
- 28-30s（参考 <10pic>）镜头继续推进绘本，随后爸爸进入画面。
整体要求：所有画面均参考对应关键帧，白模视频仅作为运镜、镜头运动和角色动画参考，不参考画面内容。最终输出 30 秒 16:9 横构图宽屏视频。Two things make it worth copying. The split is explicit, so the blockout cannot leak its grey geometry into the render. And the closing line names output length and aspect ratio in the prompt, which the request also exposes as fields.
Example three: a nine-panel storyboard with two subjects
This prompt binds four images by role before the story starts: image1 for shot structure, image2 for environment and colour, image3 and image4 for the two characters. Then come 主体设定, 整体风格, a negative list and nine timestamped shots with dialogue. The negative list is what most prompts skip.
黑白、单色、灰度、去色；手绘、素描、线稿、插画、漫画、动画；分镜稿 / 故事板、草图；移轴微缩、玩偶感、塑料 CG、油腻过曝 CG。It names visual families rather than individual words. Saying what the render must not look like is how a style instruction survives a nine-panel grid.
Editing, audio editing and extension
Each published editing prompt fixes what stays before it says what changes. The clearest is the ageing prompt.
保留 @视频 1 的构图、机位、光线与表演节奏，只改写画面里女主的样貌与神情：让她从二十多岁自然地老去到六十岁，眼神里的隐忍慢慢化开，泪光滑过眼角，嘴角一点点扬起，最后破涕为笑。全程一镜到底，不跳切、不闪烁，五官随年龄渐变而不漂移。The constraint list is the prompt. Translation follows the same shape, changing only the spoken dialogue, its language and the lip movement. Extension is the shortest of the four, and the page notes separately that the output format should be MOV.
Two request-side rules belong with these prompts. An edit task needs ratio set to adaptive and duration set to minus one, and its reference video must run between four and thirty seconds. The API guide covers where those fields sit.
What the vendor has not published
Four gaps matter here because third-party guides fill them with invented detail. No frame rate appears anywhere. No stability percentage exists, only 稳定性下降、可能需要抽卡. No 2K output tier exists, since the published resolutions are 480p, 720p and 1080p. And no seed parameter is documented, so advice about holding a seed constant is not coming from ByteDance.
Frequently asked questions
Does Seedance 2.5 respond to timestamps? Yes, to whole-second timestamps. Seedance 2.0 answers shot numbers instead.
What is a white-model reference? A blockout video carrying motion, camera, path and lighting information. Two granularities exist, and the vendor recommends the coarse one.
How do I align output strictly with a storyboard? Use a keyframe set rather than a multi-panel grid, and open the prompt with 以图片 x 至图片 x 的顺序作为关键帧.
How many reference items fit in one request? Fifty: thirty images, ten videos, ten audio clips, with video and audio capped at thirty seconds each in total.
A prompt for Seedance 2.5 is a shot list with roles attached. Name what each input controls, give every beat its own seconds, and collect everything that applies to the whole clip into one closing paragraph.