MiniMax H3 单次只能生成 4–15 秒。本接口把任意时长的片子拆成多个片段， 逐段生成后用「重叠续接 + 删除重复区」无缝拼成一条长视频，画面、人物、声线跨段连续。
| 项 | 普通接口 /v1/videos/generations | 长视频接口 /v1/videos/sequences |
|---|---|---|
| 单次时长 | 4–15 秒 | 最长约 52 秒（12 段 × 5 秒） |
| 执行方式 | 并行调度 | 固定串行（专用通道） |
| 耗时 | 数分钟 | 每段约 5 分钟，整体较长 |
| 跨段连续 | 无（每次独立） | 有（latent 级续接 + 去重） |
| 提示词 | 单段 | 逐段，由你自己写（本接口不做自动分段） |
edit 模式，见 7.3）| 模型 | 说明 | 能引用参考素材吗 |
|---|---|---|
| H3_T2V_loog | 纯文生。只靠文字描述角色与场景 | 不能（传了会自动改按 R2V 处理，并回一条 warnings） |
| H3_R2V_loog | 多参考。用参考图锁人物/场景，也能用参考视频借运镜 | 能（ref_images / ref_audios / ref_videos） |
怎么选？要角色长相稳定、跨段是"同一个人" → 用 H3_R2V_loog 并上传角色图；
不在意具体长相、只要文字描述 → 用 H3_T2V_loog。
本接口没有 I2V（从一张图往下拍），该需求请用普通接口的 minimax_h3_i2v（但那只出单段）。
本接口不做自动分段：段数、每段演什么、时间轴怎么排，都由你写进
segment_prompts。先读第 6 节学会六段式写法，再回来照这个例子提交。
curl -X POST http://47.111.83.139:19180/v1/videos/sequences \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <key>" \
-d '{
"model": "H3_T2V_loog",
"segment_prompts": [
"subject_definitions:\n<Subject 1> is ...\n\nsummary:\n[reference generation] ...\n\n（第 1 段全文）",
"subject_definitions:\n...\n\nsummary:\n[video continuation] ...\n\n（第 2 段全文）",
"……（第 3 段全文）"
],
"segment_seconds": 5,
"megapixels": 0.6,
"aspect_ratio": "16:9 (Widescreen)",
"seed": 20260916
}'segment_prompts 两种写法都接受：JSON 字符串数组（推荐），
或一个用独占一行的 --- SEGMENT --- 分隔的大字符串。
数组有几条就生成几段，秒级入队、无需等待。
curl http://47.111.83.139:19180/v1/videos/sequences/{id} -H "Authorization: Bearer <key>"| 字段 | 含义 |
|---|---|
status | pending → running → completed / failed |
stage | client_prompts 已读取你的提示词 ｜ generating 逐段生成中 |
segments | 本任务总段数（= segment_prompts 条数） |
data[].url | 完成后取产物地址 |
| 参数 | 类型 | 必填 | 默认 | 说明 |
|---|---|---|---|---|
model | string | 是 | — | H3_T2V_loog 或 H3_R2V_loog |
segment_prompts | array / string | 是 | — | 逐段提示词。段数 = 条数。提交前会自动做格式校验，见第 11 节 |
segment_seconds | number | 否 | 5 | 每段生成时长，5–9 秒（低于 5 秒会改按 5 秒并给出警告）。必须与你提示词里的时间轴一致 |
megapixels | number | 否 | 0.6 | 分辨率档位，范围 0.1–2.0，越界会被拒绝。见速查表，建议 0.6 |
aspect_ratio | string | 否 | 16:9 (Widescreen) | 画幅 |
seed | int | 否 | 随机 | 随机种子，全片所有段共用同一个种子（保证风格一致） |
ref_images | array | 仅 R2V | — | 参考图 URL，最多 9 张（超过会被拒绝）。顺序 = 提示词里的 <Picture 1..N> |
ref_audios | array | 否 | — | 参考音频 URL，最多 3 段（超过会被拒绝），每段 2–15 秒。顺序 = <Audio 1..N> |
steps | int | 否 | 固定 3 | 传了不生效采样步数由模型配方固定为 3 |
ref_videos | array | 否 | — | 参考视频 URL，最多 3 条，按数组顺序对应 <Video 1..N>。见 7.3 |
ref_video_mode | string | 否 | edit | 参考视频用途，目前只接受 edit；传 guide/boundary 会 400（原因见 7.3） |
ref_video_audio | bool | 否 | true | 是否使用参考视频自带原声；开启时每条视频会多占一个 <Audio N> |
prompt | string | 已废弃 | — | 早期的"给一句故事自动分段"入口，现已取消。只传它不传 segment_prompts 会收到 400 |
duration_seconds | number | 已废弃 | — | 段数由 segment_prompts 条数决定，此参数不再起作用 |
H3 每段之间要重叠一小段（固定 22 帧 ≈ 0.917 秒）用于续接，合并时这段重复内容会被删掉。 所以并不是"段数 × 段长"。
每段实际帧数 = 段长秒数 × 24，再吸附到 17k+5 网格（合法值：5, 22, 39, …, 107, 124, 141 …）
5 秒 → 120 帧 → 吸附为 124 帧（5.167 秒）
重叠 = 22 帧 = 0.917 秒
每段净增时长 = 5.167 − 0.917 = 4.25 秒
总时长 = 段帧数 + (段数 − 1) × (段帧数 − 重叠帧数)
= 124 + (段数 − 1) × 102 （帧）
= 总帧数 ÷ 24 （秒）| 段数 | 总帧数 | 总时长 |
|---|---|---|
| 3 | 328 | 13.7 秒 |
| 5 | 532 | 22.2 秒 |
| 8 | 838 | 34.9 秒 |
| 9 | 940 | 39.2 秒 |
| 10 | 1042 | 43.4 秒 |
| 12（上限） | 1246 | 51.9 秒 |
想要多长，就写多少段。本接口不再根据目标时长自动推算段数 —— 时长是段数的结果，不是输入。写 8 段就是 34.9 秒，写 9 段就是 39.2 秒。
每一段提示词都必须写成下面六个字段，顺序固定、字段名不能改、一个都不能少。
正文用英文；对白保留中文，写在 <d>[Chinese] …</d> 里。
subject_definitions:
<Subject 1> is ……（定义角色/场景/服装，见下）
<Audio 1> is ……（音色/音乐参考，仅 R2V 可用）
summary:
[任务类型] 一句话概括本段
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - ……
<Audio 1>: reference - ……（仅 R2V 可用）
detailed_description:
（风格句）[Shot 1] ……（画面/动作/机位/声音）
[Shot 2] At 00:00.917, ……（切镜后的新内容）
overall_soundscape: 环境音描述
non_diegetic_music: 非叙事音乐描述，或写 N/A| 字段 | 写什么 |
|---|---|
subject_definitions | 给每个反复出现的角色/场景起一个 <Subject N>，写清身份与外观。编号一经分配，全片不变（1=小马 就一直是小马） |
summary | 一个段落，必须以方括号任务类型开头：首段用 [reference generation]，第二段起用 [video continuation] |
retention_analysis | 每个 <Subject N> 一行，标注 fully_preserved / partially_preserved / attribute_transfer / weak_reference，并用一句说明保留了什么 |
detailed_description | 正文主体。先用一两句定风格，然后按 [Shot 1]、[Shot 2] At MM:SS.mmm, … 逐镜头写：构图、人物位置、环境、动作、镜头运动、声音、对白 |
overall_soundscape | 贯穿全片的环境音与物理声（风声、水声、脚步……） |
non_diegetic_music | 只有观众听得到的配乐。不需要就写 N/A |
detailed_description 的第一行是整段的风格句，
必须写出这部片子的视觉基调。第二段起可以写"风格延续"。
| ❌ 这样写等于没写 | ✅ 写一句真的 |
|---|---|
[Style sentence in English.]（方括号里的说明是给你看的提示，不是让你照抄的） |
A warm, cinematic children's-fable style with soft morning light, gentle depth of field and unhurried camerawork. |
[Continue the established style.] |
The established warm children's-fable style, character proportions, morning light and calm camerawork continue without interruption. |
同理，summary 里的 [任务类型]、时间轴里的
[OVERLAP_END] 之类，都要换成真实内容。接口会扫描方括号并把没替换的报成警告。
小马 (S1) 转头对松鼠说：<d>[Chinese] 牛伯伯说浅，松鼠说深，我该听谁的呀？</d>(S1)、(S2)… 标注，按本段发声顺序编号<d>[Chinese] …</d> 里，不翻译<d>对白是长视频最容易翻车的地方 —— 多人对话怎么排、每段能装多少字、跨段怎么说、 画外音怎么写，都在 第 9 节，建议逐条读完。
两个边界都要管：段首「重放上一段」，段尾「给下一段留一个稳定的桥接镜头」。
只做段首是新手最常犯的漏项 —— 下一段要逐字重放的，就是你这段最后 0.917 秒。
如果那 0.917 秒是"正在走路""正在说话"这种动作中间态，下一段就没有可承接的落点，接缝必然飘。
[Shot N] 是「本段内的局部编号」，不是全片编号。
每一段都从 [Shot 1] 重新数起 —— 第 5 段的第一镜也写 [Shot 1]，
不要写成 [Shot 13]。时间码同理，每段都从 00:00.000 起算。
[Shot 1] During 00:00.000-00:00.917, ← 重放上一段的结尾（首段没有这一镜）
[Shot 2] At 00:00.917, ← 新内容 + 对白，推进剧情
[Shot 3] At 00:04.250, ← 桥接：稳定、静止、可独立描述，供下一段重放00:04.250 = 段长 5.167 秒 − 重叠 0.917 秒，是「下一个分段点」所在。
段长 5 秒时这两个时间码是固定的（00:00.917 和 00:04.250）。
段长换成 9 秒则桥接点变成 00:07.333（= 8.708 − 0.917 的近似档位，实际按吸附后帧数算）。
| ❌ 动作中间态（下一段没法接） | ✅ 稳定可承接 |
|---|---|
...the little horse keeps walking toward the bank and the shot holds on the crossing.
还在走，没有落点 |
...the little horse stands still in the middle of the stream facing the far bank, the water at the middle of its legs, the wheat sack dry on its back. The camera holds a stable side shot at water level.
站定、姿势明确、机位明确 |
...she turns and starts to speak, gesturing at the door. |
...she stands facing the door with one hand resting on the frame, shoulders still, saying nothing. The camera holds steady. |
桥接镜头写清四件事就够：主体站位 / 姿态或动作相位 / 机位与景别 / 正在持续的声音。 说白了就是「定格在一个看得清楚的画面上」。
[Shot 1] 场景开场描述……
[Shot 2] At 00:00.917, 新内容 + 对白……
[Shot 3] At 00:04.250, 桥接：小马站定在河边、低头盯着水面、镜头稳住[Shot 1] During 00:00.000-00:00.917, （逐项复述上一段的桥接镜头）
[Shot 2] At 00:00.917, without a cut, 新内容自然延续……（不要为了分段人为加剪辑点）
[Shot 3] At 00:04.250, 桥接：新的稳定状态[Shot 1] During 00:00.000-00:00.917, （复述上一段的桥接镜头）
[Shot 2] At 00:00.917, the shot cuts to ……（新镜头 + 对白）
[Shot 3] At 00:04.250, 桥接：新的稳定状态怎么检查自己写对了：把每段的 [Shot 1] 和上一段的
[Shot 3] 并排读一遍 —— 如果两处描述的是同一个画面、同一个姿态、同一个机位，
就对了。对不上就说明桥接镜头没写稳。
想控制画面里有几个人，靠正向构图（谁在左、谁在右、什么景别、怎么遮挡）， 而不是靠写一堆"画面里不要有别人"。后者对模型基本无效，还会把提示词写得很乱。
| ❌ 别这样写 | ✅ 这样写 |
|---|---|
Do not create extra characters. The girl and the pig must not appear. No silhouettes or background people. |
The little witch sits alone at the centre of the frame, her face filling a tight shoulder-up close-up, the softly blurred curtains behind her. |
| 只写"小马很紧张" | <Subject 1> stands on the left of the frame, three metres from the water, ears back and head lowered toward the current. |
[Shot N] 开头先交代当前可见人物 + 各自位置（左右/前后/远近）第一镜 [Shot 1] 不带时间戳；后续镜头写 [Shot N] At MM:SS.mmm,，
时间戳必须严格递增且落在本段时长内。推荐用这些动词：
the camera cuts to … the shot cuts to … the shot transitions to …
the shot changes to … the shot switches to …交叉溶解、淡入淡出、划像（cross-dissolve / fade / wipe）只在确实需要时用。
一次切镜要带来新信息 —— 新的主体、新的空间、新的状态、新的视角或时间跳跃。 如果只是远近或轻微角度的变化，用运镜解决，不要切镜。
完整写法 = 运动类型 + 幅度 + 速度，作为自然英文动作写在镜头里， 不要在句尾堆标签。幅度中等、速度正常时通常省略不写。
The camera pushes in with small amplitude at slow speed toward the folded letter in her hands.
The camera pans right with large amplitude at fast speed, revealing the open doorway.
The camera holds a static shot as the runner exits the frame.| 运动类型 | 含义 |
|---|---|
Zoom In / Zoom Out | 机身不动，焦距变化 |
Push In / Pull Out | 机身向前 / 向后移动 |
Pan Left / Pan Right | 机身不动，镜头水平转 |
Truck Left / Truck Right | 机身水平横移 |
Tilt Up / Tilt Down | 机身不动，镜头垂直转 |
Pedestal Up / Pedestal Down | 整个机位上下移动 |
Arc Shot | 围绕主体做弧形运动 |
Tracking Shot | 跟随运动中的主体 |
Static Shot | 机位和镜头都不动 |
Shake Slightly / Shake Strongly | 轻微 / 强烈晃动 |
POV | 主体的主观视角 |
Roll Clockwise / Roll Counterclockwise | 绕镜头轴翻滚 |
幅度写 with small amplitude / with large amplitude；
速度写 at slow speed / at fast speed。
招牌、标语、字幕、霓虹灯等画面上真的看得见的文字，用英文双引号括起来， 原文原标点保留，不翻译：
A red neon sign reading "营业中" glows above the doorway.ref_images 的数组顺序就是 <Picture 1>、<Picture 2>…（第 1 张 = <Picture 1>），
错位就会锁错角色<Subject 1> is 小马, the little horse in <Picture 1>, with ……<Picture 3> is the first frame of [Shot 1], showing ……<Audio 1> is the voice-timbre reference for <Subject 1> (S1).<Audio 1>: reference - ……（参考音色，不复制原词）第 2 段开头那段重叠画面里已经有上一段的人物了，他们由 Latent Guide 直接继承。
如果你在 subject_definitions 里把外貌又完整写一遍，等于告诉模型"这是个新人物"，
反而容易出现换脸、换装、体型变化。
| 第 2 段起怎么写 | |
|---|---|
| ❌ 重新定义 | <Subject 1> is 小马, a small light-brown young horse with a short dark mane and a white blaze, carrying a cloth sack of wheat.（又把外貌完整写了一遍） |
| ✅ 继承 + 校准 | <Subject 1> is 小马 already carried into the opening Latent Guide; <Picture 1> calibrates its established identity, proportions and coat. |
只有这几种情况才需要补定义：
<Audio N>）参考音频只负责音色：把 <Audio N> 绑到已经在 Latent Guide 里的同一个人物即可，
不需要因此重新定义他的视觉身份。多视图人物设定图（正面/侧面/背面/特写）是同一个人的不同视角，
首次定义时要说明这一点，别当成多个角色。
新角色要在重叠区结束之后才进入画面，不要让它出现在 [Shot 1] 的重叠复述里 ——
否则它会和固定画面打架。写法就是先复述旧的，再用 [Shot 2] At 00:00.917, 把新人带进来
（见 10.5 的示例：牛是在 [Shot 2] At 00:02.600, 才入画的）。
绝对不要出现 <Picture N> / <Video N> / <Audio N> 任何标签！
没有对应素材时，这些标签会被当作普通文字处理，等于白写，还会让画面变差。接口会直接拒绝这类提示词。
subject_definitions 里写：<Subject 1> is 小马, a small light-brown young horse with a short dark mane and a white blaze, carrying a cloth sack of wheat.the same …… 把外观再确认一遍，并且不要换词（上一段写 light-brown，这段也写 light-brown）[video continuation]，不要带 audio reference| 项 | T2V | R2V |
|---|---|---|
| 角色外观来源 | 文字描述 | <Picture N> |
能否写 <Picture N> | 禁止 | 必须 |
能否写 <Audio N> | 禁止 | 可以用 |
| 正文是否重复描述外观 | 需要 | 不要 |
| 编号越界会怎样 | 400 拒绝 | 400 拒绝（<Picture 5> 但只传了 3 张图） |
ref_videos）—— 借用它的运动与风格⚠️ 目前只建议用在「1 段」的任务里。
参考视频在底层是对每一段都生效、且固定取源视频开头的，多段任务里第 2 段起画面会被拉回
参考素材的起始构图（接缝处看起来像回退）。多段任务请改用 ref_images 锁角色。
参考视频给模型提供构图、镜头运动、剪辑节奏、画面风格的参照， 但不硬锁原画面内容 —— 你可以把源视频里的人物、物件、场景换掉，只借它的"怎么拍"。
| 参数 | 必填 | 说明 |
|---|---|---|
ref_videos | 否 | 参考视频 URL，最多 3 条 |
ref_video_mode | 否 | 目前只接受 edit（也是默认值） |
ref_video_audio | 否 | 是否使用参考视频自带的原声，默认 true |
按 ref_videos 数组顺序对应 <Video 1>、<Video 2>…
在 subject_definitions 里说明它提供什么：
subject_definitions:
<Subject 1> is 小马 in <Picture 1>.
<Video 1> is the reference for the camera movement and pacing of [Shot 1] and [Shot 2].
summary:
[reference generation] ...
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - identity follows <Picture 1>.
<Video 1> (camera movement and pacing): weak_reference - only the camera work is borrowed;
its subjects are not copied.视觉类参考标签可用的关系词：fully_preserved / partially_preserved /
attribute_transfer / weak_reference。
借用运镜风格属于 weak_reference。
参考视频开了原声（ref_video_audio 为 true）时，每条视频会多占一个 <Audio N>。
编号顺序是：独立 ref_audios 在前，视频原声配对在后。
ref_audios 给 1 段 + ref_videos 给 1 条(带原声)
→ <Audio 1> = 那段独立音频
→ <Audio 2> = 那条视频的原声只要不想用视频自带的声音，传 "ref_video_audio": false 即可，
编号里就不会多出配对的那几条。
ref_images 锁角色插件本身还提供两种模式，但在长视频的分段链路里不能安全使用，所以暂未开放：
| 模式 | 用途 | 为什么暂不开放 |
|---|---|---|
guide | 固定 Guide —— 拿现成视频当起点，逐帧锁定它的结尾状态 | 它俩会在第 2 段起往同一帧（frame_idx=0）追加硬关键帧，
与段间续接用的 latent 关键帧叠在一起；插件对此不去重，也从未测试过这个组合 |
boundary | 仅固定边界 —— 只要求首尾构图一致，中间自由 |
传这两个值会返回 400 并说明原因。想要"从一段现有视频接着往下拍"，现在请用普通接口
/v1/videos/generations 的 R2V + ref_videos（那条路是单段的，没有分段续接，不受此限制）。
写 [Shot 1] During 00:00.000-00:00.917, … 时，下面 6 项都要写进去。
只靠"文字回忆"上一段，接缝一定会飘。
| # | 复述什么 |
|---|---|
| 1 | 景别、机位高度、焦段观感、相机朝向 |
| 2 | 每个人物在画面中的左右、前后、远近位置 |
| 3 | 人物姿态、视线方向、表情、当前动作进行到哪一步 |
| 4 | 关键道具的位置与谁拿着 |
| 5 | 场景布局、光线方向、色温、时间状态 |
| 6 | 上一段结尾正在持续的环境声 / 对白 / 音乐状态 |
写的时候有个诀窍：先把上一段的结尾画面写成文字，再把它抄进下一段的
[Shot 1] During …，然后才写新内容。这样接缝几乎不会飘。
对白是最容易写错、也最影响成片质量的部分。规则分三块： 谁在说（说话人编号）、怎么说（写法格式）、跨段怎么办（续接与音色）。
(S1)、(S2) 编号(S1)，第二个是 (S2)[Shot 1] 里是 (S1)，到 [Shot 2] 还是 (S1)(Sx)(S1,S2)(Sx) 是「本段内」的编号，每段重新数。
第 3 段第一个开口的角色也叫 (S1)，不是 (S5)。
这和 <Subject N>（全片稳定、跨段不变）是两套完全不同的编号体系，千万别混：
<Subject N> 回答"这是谁"，(Sx) 回答"这句是谁说的"。
公式：「谁 + 什么声音 + 什么动作 + 说」写在 <d> 外面，
<d> 里面只放「语言标签 + 逐字台词」。
The young woman with a quiet, breathy voice (S1) says: <d>[English] I get off at the next station.</d>| 元素 | 写在哪 | 说明 |
|---|---|---|
| 说话人身份 + 声音特征 | <d> 外 | 角色类型、年龄、性别、音色、语速、口音。首次出现时给足，同段后面再说话可以省略 |
(S1) | <d> 外 | 紧跟说话人 |
| 动作 / 表情 / 语气 | <d> 外 | 如 turns and shouts、says softly |
| 语言标签 | <d> 内 | [Chinese] / [English] / [Japanese]… |
| 逐字台词 | <d> 内 | 原文原标点，不翻译、不改写 |
| ❌ 抽象描述（模型不知道说了什么） | ✅ 逐字台词 |
|---|---|
The squirrel warns him that the water is deep. |
<Subject 3> (S2) cries out: <d>[Chinese] 小马，别过河！这水可深啦！</d> |
He says a few reassuring words. |
<Subject 4> (S1) says in a low unhurried voice: <d>[Chinese] 水刚没过我的小腿，你过来就是了。</d> |
听不清的话用猜测填上 |
<d>[Chinese] [unclear]</d> —— 听不清就写 [unclear]，不要猜 |
一个镜头里几个人说话，按画面里的实际发言顺序连续写，每句自带自己的 (Sx)。
不要为每一句话单独开一个 Shot —— 切镜是镜头语言，不是台词分隔符。
[Shot 2] At 00:02.600, the camera cuts to a wide shot across the river.
<Subject 4> (S1) wades into the shallows and says in a low unhurried voice:
<d>[Chinese] 水刚没过我的小腿，你过来就是了。</d>
<Subject 3> (S2) stays on the stone and flicks its tail without speaking,
while <Subject 1> looks from one to the other.常见情况的写法：
| 情况 | 怎么写 |
|---|---|
| 一问一答 | 两句连着写，各自带 (S1) / (S2)，中间用动作衔接 |
| 三人轮流 | 按画面顺序写三句，编号按本段发声顺序给；没开口的那位不要给编号 |
| 打断 / 抢话 | <Subject 2> (S2) cuts him off: <d>[Chinese] …</d>，或写 their voices overlap |
| 齐声 | 用复合编号：The two children (S1,S2) shout together, <d>[English] Wait for us!</d> |
| 某人沉默但被提到 | 写成动作，不写台词也不给编号，如 <Subject 3> stays on the stone and flicks its tail |
每段 5 秒里，前 0.917 秒是上一段的重复，真正能推进新内容的只有约 4.25 秒。 中文正常语速约 4–5 字/秒，所以：
| 每段时长 | 净新增 | 建议对白总量 |
|---|---|---|
| 5 秒（默认） | 约 4.25 秒 | 不超过 2 句、合计约 20 字 |
| 8 秒 | 约 7.25 秒 | 不超过 3 句、合计约 35 字 |
| 9 秒（上限） | 约 8.25 秒 | 不超过 3–4 句、合计约 40 字 |
塞太多对白是长视频最常见的翻车原因： 嘴型对不上、语速被压缩得非常赶、台词被吞掉。宁可少写两句、把每个字说清楚。 不要求每个 Shot 都有人说话 —— 动作、反应、环境空镜、道具细节，都是有效的镜头内容。
人物在画面外说话（旁白、电话另一头、回忆），用固定短语
says in an off-screen voiceover，并且紧跟在那句 <d> 后面说明画面里的人嘴是闭着的：
<Subject 1> (S1) says in an off-screen voiceover: <d>[Chinese] 我到现在还记得那条路。</d> while its lips remain completely closed.旁白仍然占 (Sx) 编号，只是注明是画外音。不要因为"看不到人"就不给编号。
最好的做法：让分段点落在一句话说完之后。这样最省事，接缝也最稳。
如果一句话必须横跨两段，注意四件事：
<scenetrans> 标记这句话的接续点，并明确说明音频跨段连续(Sx)（这是唯一需要跨段保持编号的情况）可用的连续性表述：continues seamlessly across the cut、
continues uninterrupted into the next shot、
carries over from the previous shot、remains audible across the transition。
如果一句话被影片结尾生生截断（话说一半就到点了），用 <cutoff> 标记。
── 第 2 段的结尾（最后 0.917 秒）──
<Subject 1> (S1) is midway through a sentence, mouth open and voice
continuing: <d>[Chinese] 可是我一直以为<scenetrans></d> The line is not
finished when the segment ends.
── 第 3 段的重叠复述 ──
[Shot 1] During 00:00.000-00:00.917, <Subject 1> holds the exact ending
state: the same standing pose, mouth still open, the unfinished sentence
<d><scenetrans>这水很深</d> continuing seamlessly across the cut. From
00:00.917, the sentence finishes and the little horse turns away.不要重复已经说完的台词。 如果第 2 段里那句话已经说完了，第 3 段就不能再写一遍 —— 合并时虽然会删掉重叠画面， 但音频重复听起来非常明显。
ref_audios（≤ 3 段，每段 2–15 秒），在 subject_definitions 里定义职责：
<Audio 1> is the voice-timbre reference for <Subject 1> (S1).retention_analysis 里用官方关系词：
fully_copy（整段复制）、partially_copy（部分复制）、
reference（只借特征）、weak_reference（仅氛围相近）its words are not copied<Audio N> 绑到已经在
Latent Guide 里的同一个 <Subject N> 即可<Audio N>；
但编号始终跟着角色本人走，不因换段而变（见 10.7 的对应表）audio reuse、只借音色 audio reference、
或仅延续环境声 / 节奏[Shot 1] During … 里明确写出"上一段的环境声、动作声、说话尾音、音乐相位在延续"用 1–4 句英文概括整段的环境音、物理动作声、非语言人声（风、雨、脚步、布料摩擦、撞击、呼吸、笑声、喘息）。
overall_soundscape: Birdsong, a light breeze in the grass, the river's steady
gurgle, and slow splashing as the ox wades into the shallows.detailed_description 里了non_diegetic_musicN/ADuring 00:00.000-00:00.917, the previous room tone and the voice tail continue seamlessly.用 1–3 句英文描述只有观众听得到的配乐，聚焦乐器、速度、节奏、强弱变化。
non_diegetic_music: Sparse piano notes at a slow tempo, joined by sustained low
strings that gradually increase in volume before fading out.N/A想让成片有统一的轻音乐底子，最省事的做法是每段都用同一句话，只在最后一段写"收束"。
下面两份都是可复制的完整全文（每份 3 段，成片 13.7 秒）。每段都是三镜头结构（重放 → 新内容 → 桥接 00:04.250）：
10.1–10.3 是 T2V（纯文字描述外观），10.4–10.7 是 R2V（用参考图锁外观、带音色参考）。
两份的六字段都齐全，第二段起都写了正确的重叠复述 —— 对着改就行。
subject_definitions:
<Subject 1> is 小马, a small light-brown young horse with a short dark mane
and a white blaze on its forehead, carrying a small cloth sack of wheat.
<Subject 2> is the forest riverbank: a shallow clear brook with smooth pebbles,
mossy grey stones, tall pines and dappled morning sunlight.
summary:
[reference generation] Xiaoma arrives at the river and hesitates.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the same
light-brown coat, dark mane, white blaze and wheat sack.
<Subject 2> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the same
brook, pebbles, pines and morning light.
detailed_description:
A warm, cinematic children's-fable style with soft morning light and calm camerawork.
[Shot 1] A sunny meadow beside a shallow forest river. The little horse trots into
frame along the bank with the cloth sack of wheat on its back.
[Shot 2] At 00:00.917, the little horse slows to a stop at the water's edge, lowers
its head and looks at the flowing water; its ears flick forward and back. The camera
is a medium-wide shot that slowly pushes in.
[Shot 3] At 00:04.250, the little horse stands at the water's edge with its weight
settled on all four hooves, head lowered toward the current and body angled slightly
across the frame; the wheat sack rests on its back. The brook keeps running past the
stones, sunlight keeps shifting on the surface and its mane drifts in the light
breeze; the camera eases forward in a very slow, steady push.
overall_soundscape: Birdsong, a light breeze in the grass, and the steady
gurgle of the shallow river.
non_diegetic_music: A soft, playful light orchestral bed of piano and strings
in a warm children's-fable mood; no lyrics.subject_definitions:
<Subject 1> is the same 小马, a small light-brown young horse with a short dark
mane, a white blaze and the cloth sack of wheat.
<Subject 2> is the same forest riverbank: shallow clear brook, smooth pebbles,
mossy grey stones, tall pines, dappled morning sunlight.
summary:
[video continuation] The shot opens on the previous segment's ending state at the
water's edge, then the little horse steps toward the water and pulls back.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the same
light-brown coat, dark mane, white blaze and wheat sack.
<Subject 2> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the same
brook, pebbles, pines and morning light.
detailed_description:
The established warm children's-fable style, morning light and calm camerawork
continue without interruption.
[Shot 1] During 00:00.000-00:00.917, the little horse stands at the water's edge with
all four hooves on the smooth stones, head lowered toward the current and body angled
slightly across the frame, the wheat sack on its back; the brook keeps running past
the stones and its mane drifts in the breeze.
[Shot 2] At 00:00.917, without a cut, the little horse takes half a step toward the
water, then pulls back and tosses its head with a worried snort.
[Shot 3] At 00:04.250, the little horse stands one pace back from the water facing the
river with its weight settled on all four hooves and its ears tipped back; the brook
keeps flowing past the stones and the grass sways behind it. The camera eases back in
a very slow, steady pull.
overall_soundscape: Birdsong, a light breeze in the grass, the steady gurgle of
the shallow river, and a soft anxious whinny.
non_diegetic_music: The same soft, playful piano-and-strings fable bed,
unhurried, with no lyrics.subject_definitions:
<Subject 1> is the same 小马, a small light-brown young horse with a short dark
mane, a white blaze and the cloth sack of wheat.
<Subject 2> is the same forest riverbank: shallow clear brook, smooth pebbles,
mossy grey stones, tall pines, dappled morning sunlight.
summary:
[video continuation] The little horse finally steps into the shallow water and
crosses to the far bank, ending on a calm wide shot of the empty river.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the same
light-brown coat, dark mane, white blaze and wheat sack.
<Subject 2> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the same
brook, pebbles, pines and morning light.
detailed_description:
The established warm children's-fable style, morning light and calm camerawork
continue without interruption.
[Shot 1] During 00:00.000-00:00.917, the little horse stands one pace back from the
water facing the river with its weight settled on all four hooves and its ears tipped
back; the brook keeps flowing past the stones and the grass sways behind it.
[Shot 2] At 00:00.917, the little horse steps forward into the shallow water and wades
across, the current breaking around its legs. It neither sinks nor struggles.
[Shot 3] At 00:04.250, the little horse stands on the far bank with its weight on all
four hooves and its head turned back to look across the river, the wheat sack dry on
its back. The water keeps sliding past the shallows and its mane drifts in the breeze;
the camera eases back into a wide shot of the two banks and the quiet river.
overall_soundscape: Birdsong, a light breeze in the grass, water splashing
around hooves, and the river's steady gurgle settling back to stillness.
non_diegetic_music: The same soft piano-and-strings fable bed, resolving into
one gentle sustained chord as the little horse reaches the far bank.subject_definitions:
<Subject 1> is 小马 in <Picture 1>.
<Subject 2> is the forest riverbank in <Picture 2>.
<Subject 3> is the little red squirrel in <Picture 3>.
<Audio 1> is the voice-timbre reference for <Subject 1> (S1).
<Audio 2> is the voice-timbre reference for <Subject 3> (S2).
summary:
[reference generation + audio reference] Xiaoma reaches the river and meets a
squirrel who warns him that the water is deep.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - identity follows <Picture 1>.
<Subject 2> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - identity follows <Picture 2>.
<Subject 3> (appears in [Shot 2], [Shot 3]): fully_preserved - identity follows <Picture 3>.
<Audio 1>: reference - <Subject 1> (S1) speaks with <Audio 1>'s timbre; its words are not copied.
<Audio 2>: reference - <Subject 3> (S2) speaks with <Audio 2>'s timbre; its words are not copied.
detailed_description:
A warm, cinematic children's-fable style with soft morning light and calm camerawork.
[Shot 1] A sunny meadow beside a shallow forest river. <Subject 1> walks in from
the left and stops at the water's edge, lowering its head toward the current.
<Subject 2> fills the background with tall pines and dappled morning light.
[Shot 2] At 00:00.917, the camera cuts to a low angle as <Subject 3> darts up onto
a mossy stone in the foreground. <Subject 3> (S2) cries out: <d>[Chinese] 小马，
别过河！这水可深啦！</d> <Subject 1> (S1), using <Audio 1>'s timbre, backs
half a step from the water and answers without turning away: <d>[Chinese] 可是牛
伯伯说，这水很浅呀。</d>
[Shot 3] At 00:04.250, <Subject 1> stands one pace back from the water with its
weight settled on all four hooves and its body still facing the river, ears tipped
back; <Subject 3> stays on the mossy stone just in front of it with its tail curled
up. The brook keeps running past the stones and the squirrel's tail twitches lightly;
the camera holds a steady medium shot.
overall_soundscape: Birdsong, a light breeze in the grass, and the steady
gurgle of the shallow river.
non_diegetic_music: A soft, playful light orchestral bed of piano and strings
in a warm children's-fable mood; no lyrics.注意三件事：① 既有人物不再重复描述长相，改写成
already carried into the opening Latent Guide; <Picture N> calibrates …；
② 新角色在重叠区结束之后才登场，并把它的 <Picture N> 和
<Audio N> 一起补进来；③ 段尾写 [Shot 3] 桥接。
subject_definitions:
<Subject 1> is 小马 already carried into the opening Latent Guide; <Picture 1>
calibrates its established identity, proportions and coat.
<Subject 2> is the forest riverbank already carried into the opening Latent
Guide; <Picture 2> calibrates the brook, stones and pines.
<Subject 3> is the little red squirrel already carried into the opening Latent
Guide; <Picture 3> calibrates its established identity.
<Subject 4> is the old ox in <Picture 4>.
<Audio 3> is the voice-timbre reference for <Subject 4> (S3), preserving his
deep, unhurried delivery.
summary:
[video continuation + reference generation + audio reference] An old ox wades
into the shallows on the far bank and tells Xiaoma the water is shallow,
contradicting the squirrel.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - <Picture 1>
recalibrates it while the opening Latent Guide preserves its pose.
<Subject 2> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - <Picture 2>
recalibrates the brook, stones and pines.
<Subject 3> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - <Picture 3>
recalibrates it while the Latent Guide preserves its pose on the stone.
<Subject 4> (appears in [Shot 2], [Shot 3]): fully_preserved - identity follows <Picture 4>.
<Audio 3>: reference - <Subject 4> (S3) speaks with <Audio 3>'s timbre; its words
are not copied.
detailed_description:
The established warm children's-fable style, morning light and calm camerawork
continue without interruption.
[Shot 1] During 00:00.000-00:00.917, the little horse stands one pace back from the
water facing the river with its weight on all four hooves and ears tipped back;
<Subject 3> is perched on the mossy stone just in front of it with its tail curled
up; the brook keeps running past the stones.
[Shot 2] At 00:00.917, a heavy footfall sounds and both heads turn; the camera swings
across the river to reveal <Subject 4> wading slowly into the shallows on the far
bank, the water barely reaching his shins. <Subject 4> (S3), using <Audio 3>'s
timbre, says in a low unhurried voice: <d>[Chinese] 水才到我的小腿，过得去。</d>
[Shot 3] At 00:04.250, <Subject 1> stands at the water's edge on the near bank with
its head turned toward the far side and its weight settled on all four hooves;
<Subject 4> stands in the shallows across the river with the water at his shins, and
<Subject 3> remains on the mossy stone in the foreground. The current keeps sliding
past the ox's legs and the grass sways in the breeze; the camera holds a steady wide
shot across the water.
overall_soundscape: Birdsong, a light breeze in the grass, the river's steady
gurgle, and slow splashing as the ox wades into the shallows.
non_diegetic_music: The same soft, playful piano-and-strings fable bed,
unhurried, with no lyrics.subject_definitions:
<Subject 1> is 小马 already carried into the opening Latent Guide; <Picture 1>
calibrates its established identity, proportions and coat.
<Subject 2> is the forest riverbank already carried into the opening Latent
Guide; <Picture 2> calibrates the brook, stones and pines.
<Subject 3> is the little red squirrel already carried into the opening Latent
Guide; <Picture 3> calibrates its established identity.
<Subject 4> is the old ox already carried into the opening Latent Guide;
<Picture 4> calibrates its established identity and build.
<Audio 1> is the voice-timbre reference for <Subject 1> (S1).
<Audio 2> is the voice-timbre reference for <Subject 3> (S2).
summary:
[video continuation + audio reference] Xiaoma stops asking and tests the water
himself, crossing the shallow river to the far bank while the squirrel calls
after him and the ox waits.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - <Picture 1>
recalibrates it while the Latent Guide preserves its stance.
<Subject 2> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - <Picture 2>
recalibrates the brook, stones and pines.
<Subject 3> (appears in [Shot 1], [Shot 2]): fully_preserved - <Picture 3>
recalibrates it while the Latent Guide preserves its pose on the stone.
<Subject 4> (appears in [Shot 1], [Shot 3]): fully_preserved - <Picture 4>
recalibrates it while the Latent Guide preserves its position in the shallows.
<Audio 1>: reference - <Subject 1> (S1) speaks with <Audio 1>'s timbre; its words are not copied.
<Audio 2>: reference - <Subject 3> (S2) speaks with <Audio 2>'s timbre; its words are not copied.
detailed_description:
The established warm children's-fable style, morning light and calm camerawork
continue without interruption.
[Shot 1] During 00:00.000-00:00.917, <Subject 1> stands at the water's edge on the
near bank with its head turned toward the far side and its weight on all four hooves;
<Subject 4> stands in the shallows across the river and <Subject 3> remains on the
mossy stone in the foreground; the current keeps sliding past the ox's legs.
[Shot 2] At 00:00.917, <Subject 1> turns away from both of them and faces the water
squarely. <Subject 1> (S1) says quietly, without turning around:
<d>[Chinese] 我自己试试就知道了。</d> It steps off the bank into the stream and
wades forward, the current breaking around its legs.
[Shot 3] At 00:04.250, <Subject 1> stands in the middle of the stream facing the far
bank with its weight settled on all four legs, the water at the middle of its legs and
the wheat sack dry on its back; <Subject 4> waits in the shallows ahead. The current
keeps breaking in small ripples around its legs and its mane drifts in the breeze; the
camera eases back into a steady wide shot of the river.
overall_soundscape: Birdsong, a light breeze, water splashing around hooves,
and the river's steady gurgle.
non_diegetic_music: The same soft piano-and-strings fable bed, resolving into
one gentle sustained chord as the little horse reaches the far bank.{
"model": "H3_R2V_loog",
"segment_prompts": ["第1段全文", "第2段全文", "第3段全文"],
"ref_images": [
"https://.../小马.png",
"https://.../河边.png",
"https://.../松鼠.png",
"https://.../牛.png"
],
"ref_audios": [
"https://.../小马音色.wav",
"https://.../松鼠音色.wav",
"https://.../牛音色.wav"
],
"segment_seconds": 5,
"megapixels": 0.6,
"aspect_ratio": "16:9 (Widescreen)",
"seed": 20260916
}ref_images / ref_audios 的数组顺序就是提示词里的编号，
必须一一对应：小马=<Picture 1>，河边=<Picture 2>，
松鼠=<Picture 3>，牛=<Picture 4>；
音色同理，小马=<Audio 1>，松鼠=<Audio 2>，牛=<Audio 3>。
编号是对应关系，不是出场顺序。某个角色在哪几段说话，就只在那几段里引用它的
<Audio N>；但编号始终跟着它本人走，不会因为换段而变。
提交后接口会先做一次格式校验（纯文本层面：字段、时间码、标签编号）， 不合格直接返回 400 并指出是第几段的哪个问题，不会白白占用算力。
| 检查项 | 说明 |
|---|---|
| 六字段缺失 / 重复 / 顺序错 | subject_definitions → summary → retention_analysis → detailed_description → overall_soundscape → non_diegetic_music，一个不少、顺序不变 |
| 第 2 段起没有重叠开场 | 必须在 detailed_description 里写 [Shot 1] During 00:00.000-00:00.917, …。写在别处、或写成别的时间码都不行 |
| 重叠时间码不对 | 固定是 00:00.917（22 帧 ÷ 24fps）。写成 00:01.000 之类会被拒绝 |
| T2V 里出现 ref 标签 | 没有 ref_images 时写 <Picture N>、没有 ref_audios 时写 <Audio N>，都会被拒绝 |
| 引用编号越界 | 传了 3 张图却写 <Picture 5>；传了 1 段音频却写 <Audio 2> |
<Video N> 越界 | 没传 ref_videos 却写了 <Video N>，或编号超过 ref_videos 条数 |
传了 guide / boundary | 这两种模式在分段链路里未经证伪，暂未开放（见 7.3 末尾） |
| 镜头时间码超出段长 | [Shot 2] At 00:07.000 但 segment_seconds 只给了 5 —— 时间码用本段局部时间，要从 00:00.000 起算并落在段长内 |
| 段数超过 12 | 单批上限 12 段 |
ref_images 超过 9 张 | H3 模型硬限制，报错信息会写清「最多 9 张，收到 N 张」 |
ref_audios 超过 3 段 | 同上，上限 3 段 |
megapixels 越界 | 范围 0.1–2.0。超上限会被直接拒绝，不让你白等 |
| 请求体过大 | 上限 2 MB（逐段提示词正常只有几十 KB） |
下面这些不算错，会正常受理，但响应里会带一个 warnings 数组提醒你：
| 检查项 | 说明 |
|---|---|
| 方括号里像是模板占位符 | 例如 [Style sentence in English.]、[Continue the established style.] 没替换成真实内容 |
summary 没有 [任务类型] 开头 | 首段应为 [reference generation]，第二段起 [video continuation] |
用了参考素材但 retention_analysis 是空的 | 每个实际用到的参考标签应写一行 |
model 写 T2V 却带了参考素材 | 已自动改按 R2V 处理，但会告诉你 —— 免得你以为跑的是 T2V |
model 写 R2V 却一个参考都没给 | 多半是忘了传 ref_images；确实不需要就改用 H3_T2V_loog |
aspect_ratio 传了不认识的值 | 已按 16:9 处理，并列出可用值 |
seed / segment_seconds 不是数字 | 已按默认值处理（随机种子 / 5 秒） |
segment_seconds 超过 9 | 已按 9 秒处理。想要更长请增加段数，不是加长单段 |
segment_seconds 低于 5 | 已按 5 秒处理 —— 本接口单段最短 5 秒。要更短的成片请减少段数，不要缩短单段 |
{
"error": "segment_prompts 格式校验未通过",
"issues": [
{ "segment": 2, "field": "detailed_description",
"problem": "没有找到 \"During 00:00.000-00:00.917\"",
"hint": "第 2 段起必须先写 [Shot 1] During 00:00.000-00:00.917, ..." }
],
"docs": "http://47.111.83.139:19180/longvideo"
}segment 是第几段（从 1 开始），field 是哪个字段，
problem 是具体问题，hint 是改法。照着改完重新提交即可。
{ "id": "seq_1401250001", "object": "video.sequence",
"model": "H3_T2V_loog", "status": "pending",
"segments": 3 }{ "id": "seq_1401250001", "status": "completed", "segments": 3,
"data": [{ "url": "http://47.111.83.139:19180/videos/video/H3_finite_segments/seq_1401250001_00001_.mp4" }] }产物是单个 MP4（H.264 + AAC 立体声 32kHz），已是合并好的完整长视频， 不需要你再拼接。下载地址带签名/时效，请及时转存。
{ "status": "failed", "error": "参考素材下载失败" }两种失败要分开看：格式问题在提交时就会 400（见第 11 节）；
status: failed 是入队之后才发生的（素材下载失败、资源不足、服务异常等）。
每段生成后请检查：
如果音频切口有"爆音"，可以做极短的等功率交叉淡化； 但不能靠保留重复音频来掩盖接缝。
| ❌ 错误做法 | ✅ 正确做法 |
|---|---|
忽略上一段画面，把新剧情直接写成 [Shot 1] | [Shot 1] 先复述重叠状态，新内容放 [Shot 2] |
T2V 里写 <Picture 1>（或任何 ref 标签） | 用文字写全外观；R2V 才能用 <Picture N> |
R2V 里 ref_images 顺序与 <Picture N> 对不上 | 严格一一对应：第 1 张 = <Picture 1> |
| R2V 正文里又重复描述一遍角色长相 | 外观锁在图上，正文只写动作/环境/机位/声音 |
风格句留着 [Style sentence in English.] 没替换 | 写成真的英文句子，如 The target video uses warm cinematic... |
segment_seconds 与提示词里的时间轴不一致 | 写 5 秒段就传 segment_seconds: 5，时间码落在段长内 |
| 第 2 段复制第 1 段改几个字 | 第 2 段起必须承接直接上一段的结尾状态，不是回到第一段 |
| 把整段长片塞进一段提示词 | 拆成多段，每段 5–9 秒 |
| 每段都堆上全部历史参考素材 | 只带当前段真正需要的素材 |
| 把台词写成"它警告了一句" | 写出逐字台词 <d>[Chinese] …</d> |
| 把角色/场景写成剧情摘要，没有逐镜头信息 | 每个镜头都写构图、位置、动作、镜头、声音 |
| 只听清一半就随便编台词 | 听不清写 [unclear] |
| megapixels | 实际输出 | 说明 |
|---|---|---|
| 0.4 | 864×480 | 最快 |
| 0.6 | 1056×608 | 推荐，速度与画质平衡 |
| 1.0 | 1376×768 | 更清晰，明显更慢 |
| 环节 | 耗时 |
|---|---|
| 提交 + 格式校验 | 秒级 |
| 每段生成 | 约 5 分钟 |
| 3 段整体（含合并） | 约 16 分钟 |
| 9 段整体（含合并） | 约 50 分钟 |
| 参数 | 默认 |
|---|---|
segment_seconds | 5 |
megapixels | 0.6 |
steps | 固定 3（不可改） |
| 重叠帧数 | 22 帧（0.917 秒，不可改） |
| 单批段数上限 | 12 |