直接回答:想让 Seedance 2.5 生成带对白的视频,每句台词都放进引号里,并在台词周围写清四件事:谁在说、用什么语言和口音、什么语气、在片子的第几秒说。一句能用的写法是:The woman says quietly in French, "On rentre ?"。台词总量要和时长匹配,大约每秒两到三个英文单词的量;两句台词之间给说话人安排一个看得见的动作。Seedance 2.5 在同一次生成里同时出画面和声音,Yovid 上的「Sound」默认开启。字节跳动没有公布它能说哪些语言,所以标题里的「任意语言」指的是:你想要哪种语言就写明哪种,然后先试。官方演示里出现过英语、法语和中文对白。下面 15 条提示词覆盖独白、两人对话、广告口播、短剧和五种语言,另附两条 30 秒分段脚本。

图片提示词:电影剧照风格,两个三十多岁的成年人在安静的深夜餐馆卡座里隔桌交谈。左边的女人身体前倾,话说到一半,表情坦诚认真;右边的男人握着咖啡杯在听。室内是暖色钨丝灯光,窗上有雨痕,窗外是蓝色街灯,浅景深,皮肤质感自然,35mm 胶片感。
最后核实:2026-10-28
目录
Seedance 2.5 怎么生成对白
字节跳动把 Seedance 2.5 定位为音视频联合生成模型。落到使用上就是:人声、环境声和音效跟画面在同一次生成里出来。你不用另录配音再去对口型,提示词本身就是剧本。
在 Yovid 上,和对白相关的设置有这些:
| 设置 | 对对白的影响 |
|---|---|
| Sound(声音) | 默认开启。关掉后生成的是无声视频,也就没有对白。 |
| 时长 | 每次 4 到 30 秒。时长越长,越能容纳一来一回的对话,而不只是一句台词。 |
| 分辨率 | 480p、720p 或 1080p。Lite 套餐下 Seedance 2.5 最高 720p。 |
| 画面比例 | 16:9、9:16、1:1、4:3、3:4 或 21:9。竖屏广告和短剧用 9:16。 |
| 提示词长度 | 最多 10,000 个字符,写一段 30 秒的剧本绰绰有余。 |
| 使用条件 | 任意付费套餐或任意积分包。注册赠送的免费积分不能用于 Seedance 2.5。 |
按积分包基础价,一条 10 秒 720p 视频消耗 61 积分,30 秒消耗 183 积分。订阅套餐每秒更便宜,点生成前按钮上会显示你账户的准确消耗。套餐和积分包价格见价格页。如果同一场戏打算多试几版,可以先用 Seedance 2.5 草稿模式在 480p 下检查节奏。
有两点限制要心里有数。第一,字节跳动的官方演示里有英语台词、一段发生在药店的法语对话和一句家庭厨房里的中文台词,但官方没有列出模型能说的全部语言。第二,口型是生成出来的,不保证每次都对得上:说话人正对镜头、台词短的时候,嘴型通常更贴合,所以发布前要逐个检查近景镜头。
对白提示词公式
每句台词写成一句话,按固定顺序包含五个部分:
说话人 + 语气 + 语言和口音 + "台词" + 接下来的动作
例如:The older man says slowly, in English with a soft Irish accent, "You kept the key." He sets it on the table.
- 说话人:用看得见的特征指代人物,比如「穿黄色雨衣的女人」「个子高的那位厨师」。模型得知道该动哪张嘴。
- 语气:一两个词写清语气和音量,比如低语、平淡、笑着说、快哭了。
- 语言和口音:英语也要写明,比如 in French、in Spanish with a Mexican accent、in Mandarin Chinese。
- 引号里的台词:用目标语言和它本来的文字书写。不要在括号里附翻译,模型可能会把翻译也念出来。
- 接下来的动作:一个眼神、一次停顿、退后一步。对方回话时,画面才有事可做。
然后让台词量和时长匹配。正常对话的语速大约每秒两到三个英文单词,场景里还要留出反应的时间:
| 视频时长 | 合适的台词量 | 适合拍什么 |
|---|---|---|
| 5 秒 | 一句,不超过 10 个英文单词 | 一个反应、一句包袱、一句口号 |
| 10 秒 | 两三句短台词 | 一问一答 |
| 15 秒 | 三四句 | 一小段对话加一个动作 |
| 30 秒 | 六到八句,拆成分段时间 | 一场完整的小戏 |
单人
独白是最稳的对白镜头。说话人放在中景或更近的景别,正对镜头或稍微侧一点,再写清背景声,让人声落在一个真实的空间里。
两人
开头把两个角色各描述一次,用服装或位置区分开。之后每一句都要标明是谁说的,哪怕很短。写一句「She answers」就够了;没有说话人的光引号,最容易导致声音串到别人身上。
三人及以上
一条视频里出现三个人说话是可以的,但每多一个说话人,台词落到错误面孔上的概率就更高。给每个人固定在画面里的位置(左、中、右),台词写短,一次只让一个人开口。较长的群戏,就把对白拆进分段时间里。
15 条可复制的对白提示词
提示词都用英文写,台词保留各自的语言。每张卡片上的按钮会把提示词带进 Yovid 生成器,并预先选好 Seedance 2.5;确认 Sound 已开启,再设置时长和画面比例。也可以自己粘贴到文生视频生成器。
单人独白
1. 夜班护士,10 秒
A 10-second medium close-up of a tired nurse in her forties in blue scrubs, sitting on a bench in an empty hospital corridor at night. She rubs her eyes, looks up at the camera and says quietly in English, "Twelve hours, and the coffee machine is still broken." She lets out a short laugh. Fluorescent hum, a distant monitor beeping, soft cool light.
2. 日出时的登山向导,10 秒
A 10-second shot of a weathered mountain guide in his fifties wearing a red jacket on a rocky ridge at sunrise. He turns to the camera and says calmly in English with a Scottish accent, "Slow steps from here. The wind does the rest." Wind gusting over the mic, loose stones underfoot, golden light on his face, slow push in.
3. 开幕前的舞台演员,15 秒
A 15-second close-up of a young actress in a plain black dress standing alone on a dark stage under a single spotlight. She takes a breath, her voice shaking at first, and says in English, "I said I would leave. I never said I would forget." She pauses, steadies her chin, then says firmly, "Tonight I do both." Silent theater, the faint creak of the floorboards, slow push in to an extreme close-up.
两人对话
4. 午夜餐馆卡座,15 秒
A 15-second scene in a late-night diner booth, rain on the window. A woman in a dark knit sweater leans toward a bearded man holding a coffee mug. She says in English, quietly but sharply, "You could have called." He looks down at the mug and answers flatly, "I know." She waits, then says more softly, "So why didn't you?" He looks up at her without speaking. Low jukebox music, rain tapping the glass, slow alternating close-ups.
5. 搬空的公寓,10 秒
A 10-second scene in an empty apartment full of cardboard boxes, afternoon sun through bare windows. A young man taping a box looks up at his sister in the doorway and says in English, in a tired but happy voice, "That's the last one." She smiles and answers, "Then let's go before I start crying." Tape screech, echo in the empty room, handheld medium shot.
6. 开餐前的两位厨师,15 秒
A 15-second scene in a busy restaurant kitchen before dinner service. A tall head chef in a white jacket tastes a sauce from a spoon and says to the shorter sous chef on his left, in English with a London accent, "More salt. And taste it before you send it." The sous chef nods fast and says, "Yes, chef." Pans sizzling, extractor fans roaring, steam rising, tracking shot along the pass.
广告口播和代言台词

图片提示词:电影感照片,一位四十多岁、神情亲切的面包师站在社区小面包店的柜台后,端着一木托盘新鲜酸面包,正对镜头说话。柔和的早晨窗光,空气里飘着面粉,暖金色调,浅景深,表情自然。
7. 面包店老板对镜头说话,15 秒,竖屏
A 15-second vertical 9:16 ad in a small neighborhood bakery in morning light. The baker, a woman in her forties with a flour-dusted apron, holds a tray of sourdough loaves and speaks to the camera in a warm, friendly tone in English: "We start at four every morning, so you don't have to. Fresh bread until noon, every day." She sets the tray down and smiles. Oven door closing, light chatter, slow push in.
8. 不露脸的产品旁白,10 秒
A 10-second product shot of a matte green water bottle on a wet stone surface beside a mountain stream, water drops sliding down its side, slow orbit around the bottle. A calm male voiceover says in English, "Cold for a day. Light enough to forget it's there." No one on screen, no music, only the stream and the voice.
短剧场景

图片提示词:电影剧照风格,两个年轻人在雨夜的公交站。提着小行李箱的男人回头,看向候车亭下刚叫住他的女人。湿漉漉的柏油路面映着红色尾灯和一盏路灯,雨丝在灯光里落下,情绪紧张,变形宽银幕质感,浅景深。
9. 雨夜公交站,15 秒,竖屏
A 15-second vertical 9:16 short drama scene at a rainy bus stop at night. A young man with a small suitcase walks away from the shelter. A woman in a long coat calls after him in English, her voice breaking, "If you get on that bus, don't call me." He stops in the rain, turns his head and says quietly, "I was going to ask you to come." Heavy rain, a bus engine approaching, red tail lights on wet asphalt, slow push in on his face.
10. 拆错的信封,10 秒
A 10-second short drama scene in a bright office meeting room. A woman in a gray blazer opens an envelope, reads it and freezes. Her colleague across the table asks nervously in English, "Is it the promotion?" She looks up slowly and says, "It's your resignation letter. Signed." Air conditioning hum, paper rustling, quick cut to his shocked face.
多语言对白
字节跳动的演示里有英语、法语和中文对白。西班牙语和俄语没有出现在官方演示中,所以第 14、15 条请当作测试:先在 480p 或 720p 下出一条短片,再决定要不要做长片。
11. 英语,爱尔兰口音,10 秒
A 10-second scene in a cozy village pub with a fire burning. An elderly barman in a wool vest polishes a glass and says to a young traveler at the bar, kindly, in English with a soft Irish accent, "You'll not find the road in this fog. Stay for one more." The traveler laughs and sits back down. Crackling fire, low murmur of voices, warm amber light.
12. 法语,10 秒
A 10-second scene on a Paris cafe terrace at dusk, string lights overhead. A woman in a beige trench coat stirs her coffee and says to the man opposite her, in French, in a teasing tone, "Tu es encore en retard." He raises his hands and answers in French, "Le métro, comme toujours." She shakes her head and smiles. Street noise, cups clinking, medium two-shot.
13. 普通话,10 秒
A 10-second scene in a small family kitchen at night, a pot steaming on the stove. A mother in her fifties in a cardigan turns from the stove and says softly in Mandarin Chinese to her grown son at the table, "饭快好了,先喝点汤。" He nods and answers in Mandarin Chinese, "好,妈。" Soup bubbling, a spoon against a bowl, warm overhead light, static medium shot.
14. 西班牙语,10 秒
A 10-second scene at a sunny outdoor fruit market. A cheerful fruit seller in her sixties with a straw hat holds up a ripe mango and says to a customer in Spanish with a Mexican accent, "Pruebe este, está dulcísimo." The customer takes a bite and answers in Spanish, "Tiene razón, me llevo tres." Market chatter, a radio playing faintly, handheld medium shot.
15. 俄语,10 秒
A 10-second scene on a snowy railway platform in winter, steam rising from an old train. A grandmother in a fur hat hugs her grandson in a student coat and says in Russian, in a warm and slightly worried tone, "Позвони, когда доедешь." He smiles and answers in Russian, "Обязательно, бабушка." Train whistle, crunching snow, soft gray daylight, slow push in.
如果画面里出现了你没要求的字幕,在提示词末尾加上「no subtitles, no on-screen text」。
两条 30 秒分段对白提示词
Seedance 2.5 一次能生成 30 秒,字节跳动自己的长镜头示例也是按时间段来分配动作。对白更需要这样写,因为每个时间段都在告诉模型台词该从什么时候开始。把时长设为 30 秒,再粘贴下面任意一条,卡片会把分段显示成时间轴。
16. 深夜电台主持人,单人
- 0–10sA cramped radio studio at 2 a.m., red ON AIR light glowing. A radio host in her thirties with headphones leans into the microphone and says in English, in a low warm voice, "If you're still awake, you're not alone. Lines are open."
- 10–20sShe reads a message on her screen, smiles and says, "Daniel from the night bus says hello to his driver." The camera slowly circles her desk, faders and coffee cups in the foreground
- 20–30sShe pushes a fader up as soft piano music begins and says, "Here's something quiet for the road home." Close-up on her face in the red light as the music rises
17. 兄妹卖掉家里的五金店
- 0–10sA dusty hardware shop at closing time, late sun through the front window. An older brother in a flannel shirt counts keys at the counter. His younger sister enters holding a folder and says in English, "The buyer signed this morning."
- 10–20sHe keeps counting and answers without looking up, "Dad built this counter himself." She puts the folder down beside him and says gently, "I know. That's why I wanted you to hear it from me."
- 20–30sHe stops, picks up one old brass key and hands it to her. He says quietly, "Keep this one." She closes her hand around it as the camera pulls back to a wide shot of the shop, the bell over the door ringing softly
分段、参考素材和编辑的写法,可以看 Seedance 2.5 提示词指南。
对白常见问题怎么改
| 问题 | 常见原因 | 改法 |
|---|---|---|
| 口型对不上 | 说话人离得远、侧着身,或台词太长 | 用中景或更近的景别,说话时脸朝镜头,台词缩短 |
| 语言变了 | 没写语言,或在括号里附了翻译 | 每句台词前写明语言,删掉翻译 |
| 台词跑到别人嘴里 | 引号前没写说话人,角色长得太像 | 每句都标明说话人,让每个人外形和位置都不同 |
| 两个人同时说话 | 台词时间挨得太近 | 每个时间段只放一句,台词之间加一个动作 |
| 台词被截断 | 台词量超过了时长 | 减少字数,或把时长调长 |
| 音乐盖过人声 | 要了音乐,或没排除音乐 | 加上「no music」或「music very low under the dialogue」 |
口型对不上
大多数口型问题出在景别,而不是措辞。远景或侧脸的说话人,模型能参考的信息太少。把镜头调到中景或近景,说话时让脸朝向镜头,再把台词缩短:两句短话比一句长话更容易对上。如果某一版已经接近了,先用同样的提示词再生成一次,再考虑改写,因为每次生成的结果都会不同。
语言变了
如果写的是 She says, "Bonjour",却没有写明 French,模型可能把这场戏当成英语,改掉台词或口音。每句台词前都写明语言,台词用它本来的文字,不要在旁边附英文翻译。如果角色是故意切换语言,就给每一句分别写明语言。
多人说话串了
两个人长得像时,模型分不清该动谁的嘴。让他们容易区分(年龄、服装、在画面里的位置),开头各描述一次,之后每一句都标明说话人。再给听的那个人安排点事,比如点头、移开视线、端起杯子,画面就能清楚地看出一个人在说、一个人在听。
常见问题
Seedance 2.5 能说任何语言吗?
字节跳动没有公布语言列表。官方演示里有英语、法语和中文对白。你可以用其他语言写台词,比如西班牙语或俄语,并在提示词里写明语言,但正式使用前先用短片测试。
口型同步准确吗?
人声和嘴型是一起生成的,这有帮助,但不是每次都分毫不差。台词短、脸朝镜头、用中景或近景时效果最好。发布前逐个检查近景镜头。
一句台词可以多长?
按每秒两到三个英文单词来规划,并留出反应时间。10 秒的视频放一问一答很从容。超过四句台词,就用 30 秒时长加分段时间。
可以用真实演员的声音或经典电影台词吗?
不行。台词请自己写,人物用虚构角色。点名真实人物或照搬电影、歌曲台词的提示词可能被拒绝,而且你本来也没有使用这些内容的权利。
可以让我自己的图片里的人物开口说话吗?
可以。用图生视频生成器,把图片设为首帧,对白照同样的方法写。请使用虚构角色的图片,或已经同意使用的人的图片,因为含有真实人物的上传可能被拦截。
想比较其他模型拍对话戏的表现,可以看 Seedance 2.5 和 Veo 3.1 的对比;更多提示词合集在提示词指南。准备好了就打开 Yovid 的文生视频生成器,选择 Seedance 2.5,保持 Sound 开启,先把你最想拍的那场戏出一条 10 秒的测试片。



