跳到正文

Seedance 教程:连贯的多镜头场景与动作

把一整个场景连同其中的剪切写进一次请求,让每张图片和每段声音各司其职,从精确的首帧开始,无缝续接场景,并检查生成结果。附真实的 Seedance 场景和生成它们的提示词。

更新于

Seedance 是字节跳动的视频模型。2.5 版本能生成最长 30 秒、自带剪切、对白和声音的片段(见字节跳动旗下云平台 BytePlus 的 Seedance 2.5 提示词指南,英文),这改变了使用它的方式。一次只提示一个镜头、再费力让各个片段对上,既慢又会丢掉镜头之间的表演。更好的单位是场景:一次请求就包含铺垫、反应和结果,并写好每一个剪切点。下面每个场景都是 Bald Rhino 的智能体用 Seedance 2.5 一次生成的。画面下方的轨道就是它的各个镜头,每个镜头的时长与提示词要求一致,旁边的提示词会标出正在播放的镜头。提示词按发送时的原文(英文)展示。

自然纪录片 · 4 个镜头 · 2 张参考图

镜头 1 · 0:00

白垩纪晚期 0:00 / 0:24
提示词

A 24-second photorealistic nature-documentary scene in a misty Late Cretaceous forest. @Image 1 sets the place, framing and light: a low view at a small child's eye height, hidden behind big fern fronds, a small muddy puddle in the foreground, a path between tall conifer trunks beyond. @Image 2 sets the Tyrannosaurus rex's exact appearance: massive skull, lips covering most of the teeth, tiny two-fingered arms, horizontal body balanced by a long thick tail, dark brown and olive skin. Shot 1, 0-8s: static low view as in @Image 1. Quiet forest. A deep distant THUD: concentric ripples spread across the puddle surface. A second, stronger THUD: bigger ripples, the ferns tremble. Shot 2, 8-15s: the T. rex steps into view from the misty path far beyond the puddle, walking slowly between the trunks. It stops, lifts its head and sniffs the air. Shot 3, 15-19s: it raises its head toward the sky, turned to the side and away from the camera, and lets out a deep powerful roar; small birds flutter out of the trees. Shot 4, 19-24s: it turns and walks away deeper into the mist; its footsteps fade and the puddle slowly settles. Exciting but safe for a young child: it never notices or approaches the camera, no attack, no chasing, no blood. Sound: heavy footstep thuds, water rippling, a deep rumbling roar, wings fluttering, then quiet forest ambience. No music, no speech, no text, no people.

  1. 自然纪录片 · 4 个镜头 · 2 张参考图

    白垩纪晚期

    A 24-second photorealistic nature-documentary scene in a misty Late Cretaceous forest. @Image 1 sets the place, framing and light: a low view at a small child's eye height, hidden behind big fern fronds, a small muddy puddle in the foreground, a path between tall conifer trunks beyond. @Image 2 sets the Tyrannosaurus rex's exact appearance: massive skull, lips covering most of the teeth, tiny two-fingered arms, horizontal body balanced by a long thick tail, dark brown and olive skin. Shot 1, 0-8s: static low view as in @Image 1. Quiet forest. A deep distant THUD: concentric ripples spread across the puddle surface. A second, stronger THUD: bigger ripples, the ferns tremble. Shot 2, 8-15s: the T. rex steps into view from the misty path far beyond the puddle, walking slowly between the trunks. It stops, lifts its head and sniffs the air. Shot 3, 15-19s: it raises its head toward the sky, turned to the side and away from the camera, and lets out a deep powerful roar; small birds flutter out of the trees. Shot 4, 19-24s: it turns and walks away deeper into the mist; its footsteps fade and the puddle slowly settles. Exciting but safe for a young child: it never notices or approaches the camera, no attack, no chasing, no blood. Sound: heavy footstep thuds, water rippling, a deep rumbling roar, wings fluttering, then quiet forest ambience. No music, no speech, no text, no people.

  2. 动画喜剧 · 8 个镜头 · 1 张参考图

    漏水的温室

    Audio policy (all shots): the two characters' spoken lines, the marked sound effects and steady rain on the glass roof, plus the one light music theme marked below. No subtitles. A continuous comedy scene in one location, a rainy greenhouse workshop at dusk, in eight shots, 3D animated feature-film style. @Image 1 is the character-and-location sheet: take the two characters' designs and the greenhouse workshop from it, and take nothing else from it. On the left of @Image 1, Bolt: a small round teal tin robot with one large round cyan eye-lens, two thin antennae with red bulb tips, stubby clamp hands, rolling on a single black rubber wheel. On the right of @Image 1, Fern: a tall slender lime-green tree frog standing upright, with big golden eyes, a mustard-yellow knitted scarf, a canvas work apron and brass goggles pushed up on her forehead. The workshop: a peaked glass roof, a long wooden workbench with a row of seedlings in clay pots, warm hanging bulbs, a wooden stepladder on the right, cool blue dusk light through the glass and warm amber light inside. Music: (a light, playful pizzicato strings theme, steady medium tempo, quiet under the voices, 0-30s). Shot 1, 0-3s: wide shot of the workshop, static camera. Fern, the tall green frog in the yellow scarf, stands at the workbench on the right, watering the seedlings. A drop of water falls from a crack in the glass roof onto a seedling. <rain patters on the glass roof> <a single drip taps a clay pot> Shot 2, 3-6s: medium close-up of Fern, slight low angle. She looks up at the roof, eyes wide, and calls out. Fern, the tall green frog (a warm, quick, mid-range female voice): "Bolt! The roof is leaking on my seedlings!" Shot 3, 6-10s: medium shot, the camera tracking left to right. Bolt, the small teal robot with the cyan eye, rolls in fast from the left carrying a dented tin bucket over his head and sets it under the drip. Bolt, the small teal robot (a bright, small, slightly tinny boyish voice): "Bucket deployed!" <wheel squeak> <a drip plinks into the metal bucket> Shot 4, 10-13s: the same medium shot continues without a cut: Bolt, beside the bucket, looks up as a second and then a third drip start falling beside the first onto the bench. <drips tapping wood> Shot 5, 13-17s: two-shot at bench height, Bolt on the left and Fern on the right. Bolt's eye-lens flicks between the three drips; he holds up his one bucket. Bolt, the small teal robot (a bright, small, slightly tinny boyish voice): "Three leaks. One bucket." Fern grabs a roll of silver tape from the bench. Shot 6, 17-20s: medium shot. Fern climbs the wooden stepladder on the right, tape in hand, while Bolt rolls over and grips the ladder's legs with both clamp hands. Fern, the tall green frog (a warm, quick, mid-range female voice): "Hold the ladder steady!" <ladder creaks> Shot 7, 20-25s: the ladder shot continues without a cut, then the camera tilts up to a low angle past Bolt's antennae. At the top of the ladder, Fern tears a strip of tape and presses it over the crack; the ladder wobbles and Bolt braces it, his wheel skidding. <tape tearing> <ladder rattles> Fern (warm female voice): "Got it!" Shot 8, 25-30s: close-up of Bolt, the small teal robot, looking up proudly. One last drop falls and lands right on his cyan eye-lens; his lens shutter blinks twice. Bolt, the small teal robot (a bright, small, slightly tinny boyish voice), deadpan: "Except this one." Fern, the frog above, laughs softly off-screen. <a drip taps metal> <soft rain continues> Keep the characters' designs, sizes and screen sides consistent in every shot: Bolt reaches about Fern's hip; Bolt on the left, Fern on the right unless a shot says otherwise. The light stays cool blue dusk outside and warm amber inside throughout. Sound: only the two voices, the marked effects, steady rain on the glass and the one pizzicato theme. No subtitles, no text or logos.

  3. 2D 卡通 · 7 个镜头 · 无参考素材

    餐馆闹剧

    Bright, hyper-exaggerated 2D cartoon in the style of a zany Saturday-morning slapstick show: thick wobbly outlines, saturated colours, squash-and-stretch animation, rubbery limbs, huge bulging eyes. One location: a cramped, chaotic neon-lit diner kitchen with a pass-through window to the counter. Recurring cast: CHEF SPUD, a short round potato in a tall crooked chef hat with one giant unibrow and tiny arms; MORTIMER, a very tall noodle-limbed pelican waiter with a bow tie, whose beak holds a whole tray of plates; and GRUMBLE, a tiny furious lemon customer with a monocle and a huge moustache who sits on a stack of cushions at the counter. 0-4s: Shot 1, wide establishing shot of the kitchen. Chef Spud proudly flips a pancake; it sails up and sticks to the ceiling. Spud's eyes bulge to the size of dinner plates and his jaw drops to the floor with a cartoon boing. <boing, record scratch> Spud: "Nobody. Move." 4-8s: Shot 2, quick close-up on Grumble the lemon, veins popping, face turning from yellow to red, steam whistling out of his ears like a kettle. Grumble screams in a squeaky voice: "WHERE IS MY SOUP?!" His monocle pops out and spins in the air. 8-13s: Shot 3, medium shot. Mortimer the pelican stretches his neck like an accordion down the length of the kitchen, balancing a tower of twenty bowls on his beak, wobbling in a slow, panicked dance. His eyes go cross-eyed, pupils spiralling. <comic xylophone wobble> 13-18s: Shot 4, low angle whip-pan. Mortimer slips on the pancake batter, flips into a full slow-motion somersault, bowls spinning like planets around him, while Spud dives in a heroic pose and catches every bowl on his head, stacked in a perfect tower. Spud's face: smug, sweating, eyebrow wiggling. <cymbal crash> 18-23s: Shot 5, close-up on the counter window. Spud slides the tower over; Grumble peers into a bowl and finds it full of tiny googly-eyed noodles that wave at him. Grumble's face stretches comically long, eyes popping out on springs. The noodles chirp: "Hiiii!" 23-27s: Shot 6, wide shot. Grumble, trembling with fury, inhales hugely, swelling like a balloon, then burps a giant green bubble that floats up, engulfs Mortimer and Spud, and carries them out through the ceiling vent like a parade float. The ceiling pancake falls on Grumble's head like a hat. 27-30s: Shot 7, static medium shot on Grumble alone, pancake on his head, deadpan, one eyebrow raised, slowly sipping a tiny cup of tea. He says flatly: "...Needs salt." <single tiny sad trombone note> Hold the final frame. Lighting: warm, punchy diner neon, soft bounce. Camera: snappy cuts, exaggerated whip-pans, dramatic push-ins on faces. Sound: zany cartoon sound effects, boings, slide whistles, squeaky voices, no background music. Clean picture without subtitles, captions, text, logos or watermarks.

  4. 剧情 · 8 秒内 3 个镜头

    雨打挡风玻璃

    Three shots with cuts. Cinematic and realistic. Night, the warm glow of a dashboard. Inside a parked car in the rain. Shot 1 (0-2 s): a wide shot through the windscreen of a young couple in the front seats, rain running down the glass. Shot 2 (2-5 s): a close-up of the woman in the passenger seat, mid-twenties, short black hair; she says quietly: "We can't keep doing this." Shot 3 (5-8 s): a close-up of the man at the wheel, mid-twenties, stubble, grey hoodie; he grips the wheel and answers: "Then tell me what you want." Sound: rain on the roof and their voices; no music.

1. 写一个场景,而不是一个片段

把相互关联的动作组织成最长约 30 秒的场景,每个场景写成一次请求,并写明其中的镜头。一个事件的铺垫、反应和结果要放在一起。两次独立生成的每一个衔接处,都需要重新找回角色、表演和声音,所以请求越少越长,整体越连贯。短场景只需要用到它实际需要的时长。

一个场景里可以有剪切。写清楚哪里换视角、哪里是同一个镜头在继续;“一个场景”并不等于一个不间断的长镜头。让每个段落的结尾为下一个段落做铺垫:在追逐戏里,先交代是什么让角色掉头,再让他们往回跑。把远景、特写、反打和反应镜头都安排在同一次请求里,并给重要的反应留足时间,方便之后在剪辑时选择剪切点。把它们拆成几次生成就会丢掉表演:重复场景描述和角色图片,并不能把刚刚发生的事带过去。

2. 给每个镜头写明视角、动作、声音和音效

先写情境和画面风格,再写参考素材的分工,然后按顺序写镜头。每个镜头都要写明视角、关键的动作或反应、谁说话、具体说什么,以及我们应该听到的声音。移动摄像机是为了揭示或跟随某个具体的东西,不要在很短的时间里塞进好几个运镜。

以整秒划分的时间段可以组织节奏,但它们是导演意图,而不是剪辑时间码:在为字幕、音乐或替换配音对时间之前,先确认事情实际发生在哪里。每条提示词都以 “No text. No subtitles.”(“不要文字,不要字幕”)结尾:模型可能把字写错或自己加上字幕,所以标题和字幕应该放到剪辑里做。

温室工作间里,一个独轮的青绿色小机器人和一只围黄围巾的高个子绿青蛙。
一张参考图,八个镜头
温室场景背后的提示词

Audio policy (all shots): the two characters' spoken lines, the marked sound effects and steady rain on the glass roof, plus the one light music theme marked below. No subtitles. A continuous comedy scene in one location, a rainy greenhouse workshop at dusk, in eight shots, 3D animated feature-film style. @Image 1 is the character-and-location sheet: take the two characters' designs and the greenhouse workshop from it, and take nothing else from it. On the left of @Image 1, Bolt: a small round teal tin robot with one large round cyan eye-lens, two thin antennae with red bulb tips, stubby clamp hands, rolling on a single black rubber wheel. On the right of @Image 1, Fern: a tall slender lime-green tree frog standing upright, with big golden eyes, a mustard-yellow knitted scarf, a canvas work apron and brass goggles pushed up on her forehead. The workshop: a peaked glass roof, a long wooden workbench with a row of seedlings in clay pots, warm hanging bulbs, a wooden stepladder on the right, cool blue dusk light through the glass and warm amber light inside. Music: (a light, playful pizzicato strings theme, steady medium tempo, quiet under the voices, 0-30s). Shot 1, 0-3s: wide shot of the workshop, static camera. Fern, the tall green frog in the yellow scarf, stands at the workbench on the right, watering the seedlings. A drop of water falls from a crack in the glass roof onto a seedling. <rain patters on the glass roof> <a single drip taps a clay pot> Shot 2, 3-6s: medium close-up of Fern, slight low angle. She looks up at the roof, eyes wide, and calls out. Fern, the tall green frog (a warm, quick, mid-range female voice): "Bolt! The roof is leaking on my seedlings!" Shot 3, 6-10s: medium shot, the camera tracking left to right. Bolt, the small teal robot with the cyan eye, rolls in fast from the left carrying a dented tin bucket over his head and sets it under the drip. Bolt, the small teal robot (a bright, small, slightly tinny boyish voice): "Bucket deployed!" <wheel squeak> <a drip plinks into the metal bucket> Shot 4, 10-13s: the same medium shot continues without a cut: Bolt, beside the bucket, looks up as a second and then a third drip start falling beside the first onto the bench. <drips tapping wood> Shot 5, 13-17s: two-shot at bench height, Bolt on the left and Fern on the right. Bolt's eye-lens flicks between the three drips; he holds up his one bucket. Bolt, the small teal robot (a bright, small, slightly tinny boyish voice): "Three leaks. One bucket." Fern grabs a roll of silver tape from the bench. Shot 6, 17-20s: medium shot. Fern climbs the wooden stepladder on the right, tape in hand, while Bolt rolls over and grips the ladder's legs with both clamp hands. Fern, the tall green frog (a warm, quick, mid-range female voice): "Hold the ladder steady!" <ladder creaks> Shot 7, 20-25s: the ladder shot continues without a cut, then the camera tilts up to a low angle past Bolt's antennae. At the top of the ladder, Fern tears a strip of tape and presses it over the crack; the ladder wobbles and Bolt braces it, his wheel skidding. <tape tearing> <ladder rattles> Fern (warm female voice): "Got it!" Shot 8, 25-30s: close-up of Bolt, the small teal robot, looking up proudly. One last drop falls and lands right on his cyan eye-lens; his lens shutter blinks twice. Bolt, the small teal robot (a bright, small, slightly tinny boyish voice), deadpan: "Except this one." Fern, the frog above, laughs softly off-screen. <a drip taps metal> <soft rain continues> Keep the characters' designs, sizes and screen sides consistent in every shot: Bolt reaches about Fern's hip; Bolt on the left, Fern on the right unless a shot says otherwise. The light stays cool blue dusk outside and warm amber inside throughout. Sound: only the two voices, the marked effects, steady rain on the glass and the one pizzicato theme. No subtitles, no text or logos.

  • take the two characters' designs and the greenhouse workshop from it, and take nothing else from it

    唯一的参考图有了明确的任务和边界。

  • Fern, the tall green frog in the yellow scarf

    每个动作都用外貌而不只是名字来指明主体,让模型把它绑定到正确的角色上。

  • the same medium shot continues without a cut

    摄像机不切换的地方也要写出来。

  • <a drip plinks into the metal bucket>

    每个声音都和画面上发出它的东西绑定在一起。

  • No subtitles, no text or logos.

    文字在剪辑时添加,从不交给模型。

3. 把每个参考素材绑定到一项任务

参考素材按类型分别编号,顺序就是附加的顺序:@Image 1、@Image 2、@Audio 1、@Video 1。写下一个标记并不会附加任何东西,所以要核对实际的素材列表,输入一变就更新提示词。然后给每个参考素材一项任务,并说明它提供什么:

  • 图片用来确定身份、当前外观、地点、构图、风格或开场姿势。说明哪些特征要保留,哪些由这个场景改变。
  • 音频提供说话人的声音、需要跟随的具体表演,或者音乐。把每个说话的角色绑定到各自的样本,并说明只取声音,不取原来的台词或背景。
  • 视频是要续接的瞬间、要跟随的动作或表演,或者要编辑的素材。说明是哪一种。

除了名字,还要把动作和台词绑定到角色看得见的外貌和位置上:写“青绿色的小机器人提着水桶”,而不只是“Bolt”。这个宫廷场景有五个人物和一个地点需要分清,所以提示词绑定了七张图片和两段声音样本,每一份对应一个人。把鼠标移到卡片上,就能看到提示词在哪里用到了它。

大殿

大殿:场景
@Image 7 场景

武丁王

武丁王:面部
@Image 1 面部
武丁王:全身
@Image 2 全身

占卜师 Que

占卜师 Que:面部
@Image 3 面部
占卜师 Que:全身
@Image 4 全身

Jing 夫人

Jing 夫人:全身
@Image 5 全身

Ji 王子

Ji 王子:全身
@Image 6 全身
提示词

A 30-second scene at dusk in the Shang capital of Yin, about 1250 BCE, in a torchlit open-fronted thatched hall on a rammed-earth platform. Cinematic live-action historical drama, 35 mm anamorphic look, deep amber torchlight and blue dusk, ochre, indigo and charcoal tones with cinnabar-red accents. Use @Image 7 for the setting: the thatched hall, red wooden pillars, throne mat, bronze ding vessels and torch stands. Courtiers in muted hemp and indigo robes line the walls. Use @Image 1 and @Image 2 for KING WU DING: a tall broad-shouldered man of about forty, severe calm face, short square-cut black beard, dark indigo layered robe with cinnabar-red border, bronze belt plate, jade pendant, topknot with a bone hairpin, no crown. Use @Image 3 and @Image 4 for QUE: a narrow man of about fifty-five, thin grey topknot, long thin grey beard, undyed hemp robe with cinnabar stains on the sleeves. Use @Image 5 for LADY JING: an elegant woman of about forty, cinnabar-painted lips, hair in a high coil with bone and turquoise pins, layered black-and-red silk robe, still hands. Use @Image 6 for PRINCE JI: a slim serious boy of fifteen, indigo robe, topknot, a white jade ear ornament. Keep all faces and costumes exactly as in the references, as real people, not paintings. @Audio 1 is the King's deep, low, heavy voice: use it whenever the King speaks. @Audio 2 is Que's raspy, gentle, ironic older voice: use it whenever Que speaks. Take voice identity only, not the old lines or background sounds. Shot 1, 0-5s: Wide establishing view down the hall, slow push in. Torches flicker; courtiers line both walls in silence. King Wu Ding enters without a word from the far end and kneels on the throne mat. Lady Jing and Prince Ji stand beside the mat. The court goes quiet. Shot 2, 5-12s: Medium on Que kneeling before the king. Que offers a bronze rod with both hands and says, formally and clearly, "On Gway-sih day, Chweh divines: will the rain come within ten days?" Que bows his head. Shot 3, 12-20s: Over the King's shoulder: the king takes the rod and presses it himself into a drilled pit on a flat pale turtle plastron (the flat belly shell of a turtle, with a central seam, never a round or pearly shell) lying on a cloth. Slow macro on the shell. A thin curl of smoke. A sharp tick; a crack runs across the shell. Sound: tick, breath. Shot 4, 20-26s: Close on King Wu Ding's severe face, staring at the crack. A long pause; everyone waits. He says, low, slowly, in one clear word: "Rain." Shot 5, 26-30s: Slow wide of the court as it breathes out. Prince Ji smiles with relief; Lady Jing watches the king's face, not the shell. A single deep bronze bell sounds. Period guardrails: bronze, wood, bone, hemp and clay only; no iron, no modern objects, no tiled roofs, no stone walls, no visible writing. Only the two quoted lines, spoken by those speakers. No music. No subtitles, text or logos.

场景成片,含原生声音

这两段声音样本是从同一制作项目前几个场景中截取的干净片段,分别是国王和 Que 的声音。只靠文字描述声音,指导作用较弱;一段只有一个说话人、词句完整、背景声很少的短样本,才能把声音从一个场景带到下一个场景。

4. 从精确的首帧开始

当开场构图必须精确时,要把图片作为首帧提供,而不是作为参考图。在文字里把参考图称为“首帧”,只能要求一个相似的开场。首帧带有姿势和位置,但不带动作、声音或环境氛围,而且不能在同一次请求中与参考图一起使用(BytePlus 视频生成 API 参考,英文)。

晨光中一杯冒着热气的黑咖啡,旁边是一袋 NORTE 咖啡。

首帧 Seedance

咖啡广告

Shot 1, 0-3s: start exactly on the opening frame; a very slow push-in toward the black ceramic cup as the backlit steam curls and swirls in graceful spirals in the sunbeam.

蓝调时刻,穿黄色雨衣的女孩在屋顶上提着一盏发光的纸灯笼。

首帧 Seedance

城市上空的灯笼

Shot 1, 0-4s: medium shot from the opening frame; the wind lifts her red scarf, she smiles with wonder, raises the glowing lantern in both hands and gently lets it go; it floats up past her face, warm light sweeping across her cheeks.

5. 无缝续接一个场景

当一个场景超出一次生成的长度时,先选出剪辑中保留的最后一个瞬间,再决定下一次请求如何使用之前的素材。要用已确认的素材本身,而不是再描述一遍:

你想要的 怎么请求 提示词怎么写
延续同一个动作 延长模式,把你保留的最后几秒作为 @Video 1 从 @Video 1 的结尾继续:写新的动作和剪切,保持位置、运动、光线和声音
从一个精确的姿势继续 把保留的最后一帧作为首帧 从这一帧开始,指导接下来发生什么
沿用之前画面风格或动作的新场景 参考模式,把有用的片段作为 @Video 1 @Video 1 提供哪些特征,以及新场景是什么
修改现有片段 编辑模式,附上整段要修改的内容 精确的修改,以及要保留的动作、节奏、人物和声音

要延长时,先剪下保留片段的最后几秒,并检查它的最后几帧:结尾处的一个剪切或不想要的特写会把续接带偏。然后检查每个衔接处的两侧,用正常速度看,也要凑近看:重复的瞬间、身份或位置的跳变、失去的动势、光线变化,或者重新开始的声音。

这个衔接来自一个要求 “continue the scene of @Video 1 without a visible cut”(“无可见剪切地续接 @Video 1 的场景”)的制作项目。请求把上一个片段的最后八秒作为 @Video 1 附上,但没有选择延长模式,结果续接片段一开始就换了新的机位。这项检查只需十秒,就能在剪辑之前发现问题:

@Video 1 的尾帧

续接片段的首帧

衔接处

@Video 1 的尾帧
续接片段的首帧

6. 让声音和画面一起被导演

当对白、环境声或音效属于这个场景时,保留原生音频。把环境声描述成这个地点持续存在的声音(雨打玻璃、房间底噪),并在仍处于同一地点的剪切之间保持一致。把每个音效绑定到画面上看得见的起因,然后检查同步:动作可能落在与你要求不同的那一秒。如果音乐需要在整个制作项目中保持连续,就单独生成配乐,让 Seedance 只生成人声、环境声和音效,不带音乐。视频配乐与音效教程讲解了这类配乐。

Seedance 2.5 的原生对白支持 11 种语言:中文、英语、西班牙语、印尼语、马来语、泰语、阿拉伯语、葡萄牙语、越南语、日语和韩语(BytePlus 上的 Seedance 2.5 教程,英文)。Bald Rhino 的智能体在英语和中文对白中使用它;其他语言,以及台词必须一字不差的时候,它会先用语音模型准备好台词,作为说话人的音频参考。无论哪种情况,都要逐句听,检查有没有说错的词、发音问题、说话人不对或漏掉的句子。第一个片段没问题,并不代表后面的片段也没问题。

7. 先看完整个场景,再从根源修复

在评判细节之前,先看完整的动作、听完整的声音。检查每个参考素材是否保持了身份、每句台词是否出自正确的说话人、关键动作是否发生、衔接处是否成立,以及有没有混进字幕或音乐。只要镜头服务于场景,即便模型改了一个无关紧要的构图,也可以保留。当关键之处出错时(缺了一次退场、角色互换、该有的结果始终没出现),要修正它的根源:参考素材的对应关系、过于拥挤的调度、提示词漏掉的动作,或者剪辑本身。然后再检查修复处两侧的衔接。