B-国外链路.md

B|国外链路:YouTube Shorts / TikTok / Instagram Reels 上的「东方仙宫 + 珠光宝气古风美人」AI 视频实操调研

调研时间点:2026-09-12(CST)。本文所有价格、模型版本均以该时点公开信息为准,AI 视频领域版本迭代以「周」计,使用前请复核原始链接。 数据性质:工具能力/价格来自厂商页面或第三方实测;账号案例来自媒体报道与创作者自述,凡属推测的部分均标注「(未证实)」。 一句话前提校正:截至 2026-09,Sora 2 已经停服(App/网页 2026-04-26 关停,API 2026-09-24 下线),Veo 4 至今未发布(2026 Google I/O 只发了 Gemini Omni)。任何还在按「Sora 2 + Veo 3.1 双主力」写方案的资料都已过期。


0. 国外链路现状速览(先看这张表)

环节2026-09 的国外实际主力备注
出图Midjourney V8.1/V8.2(alpha)、FLUX 3 / FLUX.1 Krea、Ideogram 4.0、Nano Banana Pro、Seedream 5.0、SDXL 生态(Pony / Illustrious)古风仙侠的审美天花板仍在 MJ;批量可控性在 FLUX/SDXL
出视频Veo 3.1、Kling 3.0 / 3.0 Omni、Runway Gen-4.5、Seedance 2.5、Hailuo 3(MiniMax H3)、Grok Imagine Video 1.5、Higgsfield(聚合层)详见第 2 节「各自擅长什么」
一致性Kling Elements、Runway Gen-4 References、Veo Flow Ingredients、Seedance 多参考、Higgsfield Soul ID详见第 3 节
音频模型原生音频(Kling 3.0 Omni / Veo 3.1 / Hailuo 3)、ElevenLabs(人声/旁白)、Suno v6(国风配乐)Kling 3.0 与 Seedance 已把「对白+音效+配乐」并进一次生成
后期Topaz Video AI(Starlight/Astra)、DaVinci ResolveTopaz 2025-10 起改为订阅制
分发竖屏 9:16、8–15s、前 1–2 秒定生死Shorts 主信号是 swipe-away rate

核心判断(给制作方的三条)

  1. 仙宫题材不吃「角色一致性」,吃「空间一致性」。 爆款(如「中式天庭」)是纯空镜 + 大景小人,没有正脸、没有台词,因此绕开了 AI 最弱的两项(人脸/口型),只压最擅长的两项(建筑/云雾/体积光)。这是该题材能低成本跑通的根本原因。
  2. 古风美人题材恰好相反,必须有稳定的脸与服装,必须走参考图/训练身份路线,成本大约是仙宫空镜的 3–5 倍。
  3. 两个题材的共同最优解是「静态图 → 图生视频」,而不是文生视频。先出 20–50 张图选 1 张,再让视频模型做微动(云飘、衣摆、鹤飞、光移),比 T2V 一次成片成功率高一个量级。

1. 图像侧:古风 / 仙侠风格到底怎么调

1.1 各引擎的定位与调法

Midjourney V8.1 / V8.2(alpha.midjourney.com)

FLUX 3 / FLUX.1 Krea dev / FLUX Kontext

Ideogram 4.0

Stable Diffusion / SDXL 生态(Pony、Illustrious、NoobAI)

NijiJourney(Niji 7)

1.2 可抄用的英文提示词模板(图像侧 8 条)

用法:MJ 直接把 --ar / --sref / --oref / --ow / --style raw / --v 7 拼在末尾;FLUX/SDXL 去掉参数只留正文。所有模板均已含镜头 + 光照 + 材质 + 画质四类词。

A. 仙宫场景 4 条

① 九重天宫城 建立镜头(16:9,最万能)

Celestial capital above a sea of clouds, nine tiers of floating palaces cascading into the distance, vermilion lacquered eaves and white jade terraces, gold filigree ridge ornaments, colossal bronze astronomical instrument as the visual anchor, low-angle 28mm wide establishing shot, deep focus, all planes tack sharp, volumetric god rays breaking through the cloud deck, warm side-backlight at dusk, transparent cyan shadows, atmospheric perspective keeping color in the far tiers, 2% human figures as scale reference only, cinematic fantasy film still, monumental Eastern production design, physically plausible light, shot on ARRI Alexa 65, 35mm, subtle film grain, no railing, no text, 8k --ar 16:9 --style raw --sref <你的仙宫风格板> --sw 250

② 南天门 / 白玉长廊 竖屏沉浸(9:16 首图)

9:16 vertical, immense jade stairway rising from foreground to a vermilion sky-gate that extends beyond the top of frame, single-point vertical perspective, low camera, 24mm equivalent, gate opening roughly thirty adult-heights tall, two white-robed immortals walking away from camera at 3% frame height, three-layer trailing silk hems all flowing left in the same wind, transparent shallow water mirror on unfenced white jade platform, rising morning mist, warm ivory highlights, cool blue-grey shadows, one restrained starburst lens flare at the gate edge, ethereal and quiet, cinematic fantasy film still, ultra-detailed material rendering, subsurface scattering on jade, shot on 35mm anamorphic, film grain, no people facing camera, no text --ar 9:16 --style raw

③ 殿宇内部 / 藏书阁 体积光(仙宫「内景」稀缺,做差异化用)

Interior of a celestial scripture hall suspended in clouds, enormous red-lacquer columns and open lattice screens, thousands of floating scrolls slowly orbiting in mid-air, one shaft of volumetric god rays entering from a high window, dust motes and incense smoke visible in the beam, polished black stone floor with faint reflections, gold filigree lotus capitals, selective high saturation — jade green and peacock blue kept mid-low, amber gold only on lit edges, wide 32mm medium shot, symmetrical composition, deep focus, cinematic chiaroscuro, physically based rendering, octane-quality material detail, shot on ARRI Alexa, 40mm, fine film grain, 8k --ar 3:2 --style raw

④ 大景小人 + 仙鹤 + 瀑布(中式「意境」标准构图)

Astronomical scale celestial realm, no visible horizon, a giant vermilion main hall cut off by the top of frame, white jade causeway crossing empty air toward it, a transparent waterfall spilling off a floating island into the lower cloud sea, a flock of red-crowned cranes gliding across the middle ground, one lone immortal in structured flowing robes on the right third, back turned, about 3% of frame height, 45% of the frame left as breathing air holding cloud, mist and distant pavilions, low horizon, asymmetric framing with an ancient pine as foreground frame, ultra-wide 21mm, deep depth of field, high-key jade white and pale blue-grey, restrained vermilion accents, cinematic volumetric light, hyper-detailed architecture, immersive, no railing, no text, no watermark, 8k --ar 21:9 --style raw --sw 200

B. 古风美人 4 条

⑤ 珠光宝气皇后 特写(面部 + 首饰材质是重点)

Extreme close-up beauty portrait of an ancient Chinese empress, three-quarter profile, porcelain skin with visible subsurface scattering and fine peach-fuzz, jade-and-gold filigree phoenix crown with hanging pearl tassels, gold filigree hairpins, mother-of-pearl inlay, silk collar with iridescent sheen, deep red lips, kohl-lined almond eyes looking slightly off-camera, 85mm f/1.4 portrait lens, razor-thin depth of field with creamy bokeh, warm key light from camera left with a soft silver reflector, cool rim light separating the hair from a dark background, catchlights in the eyes, shot on ARRI Alexa with vintage Cooke optics, cinematic color grade, fine 35mm film grain, hyper-detailed jewelry micro-texture, 8k --ar 9:16 --style raw --oref <角色参考图> --ow 550

⑥ 汉服仙女 全身立绘(服装结构 + 比例)

Full-body vertical shot of a Chinese goddess in layered hanfu, eight-head-tall elegant proportions, structured silk outer robe with bone, sheer gauze sleeves with air, restrained ornament, one dominant color scheme, wide iridescent silk sash flowing in a single consistent wind direction, gold filigree hair ornament, pearl earrings, standing on a white jade terrace above clouds, back three-quarter view then turning, 50mm lens at eye level, full-length composition with negative space above, soft overcast skylight plus a warm rim from behind, fabric physics rendered with realistic weight and drape, subsurface scattering on skin, cinematic still, fashion-editorial lighting, shot on ARRI Alexa, 50mm, subtle film grain, ultra-detailed embroidery, no extra fingers, no text --ar 9:16 --style raw

⑦ 神女下凡 / 神性光(做「封面帧」最好用)

Medium shot of a celestial goddess descending, ethereal halo of soft golden light behind her head, translucent layered gauze robes catching volumetric god rays, iridescent silk shifting from pearl white to pale aquamarine, gold filigree diadem and pearl drops, long black hair lifting in the updraft, calm downward gaze, arms slightly open, cloud sea and distant palace spires far below in soft bokeh, low camera looking up, 35mm, slight low-angle hero framing, dramatic backlight with lens bloom, motivated practical glow from a floating lantern, cinematic fantasy key art, hyper-real skin texture, shot on ARRI Alexa 35, 35mm, film grain, 8k --ar 2:3 --style raw

⑧ 花神变装 / 回眸动态帧(用于变装类竖屏)

Half-body shot of a flower goddess turning to look back over her shoulder, hair and sheer veil caught mid-motion with believable inertia, gold thread embroidery on scarlet silk, jade and pearl hairpins, petals and pollen drifting through the air in the light, shallow depth of field, 85mm, motion energy frozen at 1/250s, warm backlight through translucent fabric showing weave and translucency, cool ambient fill, cinematic and luxurious, high-end beauty campaign look, detailed skin pores and fine facial hair, shot on ARRI Alexa with Cooke S4, subtle grain, 8k --ar 9:16 --style raw --oref <同一角色> --ow 600

镜头/光照/材质/画质词库(直接往里塞)

  • 镜头:low-angle 28mm establishing shot、single-point vertical perspective、85mm f/1.4 portrait、anamorphic 2.39:1、deep focus、razor-thin DOF
  • 光照:volumetric god rays、warm side-backlight、transparent cyan shadows、motivated practical light、rim light、atmospheric perspective
  • 材质:gold filigree、iridescent silk、mother-of-pearl inlay、subsurface scattering、translucent gauze、polished jade
  • 画质:shot on ARRI Alexa、35mm film grain、physically based rendering、cinematic color grade、8k

2. 视频侧:各模型擅长什么,仙宫题材上差异在哪

模型单次时长最强项仙宫题材实际表现
Veo 3.1(Google,2025-10/2026-01 起)4/6/8s,可续接原生音频 + 提示词遵循 + 4K 级质感 + Ingredients 参考云海/建筑质感最"贵",光很"电影";缺点是贵、单段短
Kling 3.0 / 3.0 Omni(快手,2026-02-04)3–15s,可续到 3 分钟一次生成 6 镜头故事板 + 原生 4K 60fps + Omni Audio(对白/音效/配乐同帧生成)+ Elements 角色锁定仙宫「巡游式多镜头」最强;多角色同场也能锁
Runway Gen-4.52–10s,原生 720p + 4K 放大Gen-4 References(最多 3 张参考)+ Motion Brush 逐区域控制 + Director Mode + 时间线剪辑适合「先定角色板再动」的精修流程;纯仙宫空镜性价比一般
Seedance 2.5(字节)4–30s 任意整秒首尾帧关键帧 + 最多 30 图/10 视频/10 音频参考 + 编辑式局部改720p 上限,但长镜头一次成最省事;竖屏短剧首选
Hailuo 3 / MiniMax H35–15s4K 原生 + 便宜($0.06/s @768p、$0.16/s @4K)+ 32kHz 立体声音频;H3-Base 已开源仙宫大景「清晰又便宜」,但该端点不吃关键帧/参考图
Grok Imagine Video 1.51–15s极快、便宜($0.08/s @480p),适合草稿迭代用于快速试运镜与构图,成片仍需换模型
Higgsfield聚合层Soul ID(20+ 张图训练持久角色身份)+ Cinema Studio 运镜 + LipSync做「同一仙女跨几十条视频」的账号最省事的商业化方案
Pika / Luma Dream Machine短社交特效/风格化、易用仙宫题材偏弱,更多用于「变装/转场」类效果
Sora 2—已停服不要再纳入方案

差异小结(仙宫): 建筑与云雾的「体量感」→ Veo 3.1 / Kling 3.0;长镜头一次成 → Seedance 2.5;4K 且便宜 → Hailuo 3;多镜头叙事 + 音频一体 → Kling 3.0 Omni;角色一致性 → Higgsfield Soul ID / Runway References。

2.1 可直接用的英文视频提示词(6 条)

V1 — Veo 3.1:仙宫航拍推进(含环境音)

Slow forward dolly-in through a sea of clouds toward a colossal vermilion sky-gate, camera rises slightly as the gate fills the frame, god rays sweep across the jade causeway, red-crowned cranes cross the frame left to right, silk banners lift in a steady wind. Ambient audio only: high-altitude wind, distant bronze bell, faint crane calls, no music.

V2 — Kling 3.0:6 镜头仙宫巡游(原生音频)

Multi-shot, 9:16 vertical.
Shot 1 (3s, wide, slow crane up): cloud sea at dawn, distant floating palaces resolve out of mist. Audio: wind, distant chime.
Shot 2 (2.5s, medium, dolly-in): white jade corridor, one immortal in white robes walking away. Audio: footsteps on stone, robe rustle.
Shot 3 (2s, macro CU, rack focus): gold filigree bracket detail, pearls on a tassel swinging. Audio: faint metallic shimmer.
Shot 4 (3s, wide, slow pan right): waterfall pouring off a floating island into the cloud deck. Audio: deep rushing water.
Shot 5 (2.5s, low-angle, push in): the main hall, cut off by the top of frame. Audio: low bronze bell swell.
Shot 6 (2s, wide, static): cranes gliding across the moon. Music: sparse guqin notes resolving on the final frame.

V3 — Seedance 2.5:首尾帧 30s 长镜头

First frame: wide shot of an empty jade terrace above clouds at dusk. Last frame: the same terrace, now lit by hundreds of floating lanterns, one goddess standing at the edge, back to camera. Camera: very slow continuous dolly-in with a slight upward tilt across 30 seconds, no cuts. Continuous motion: clouds drift right to left, lanterns rise one by one, robe hems lift in the same wind. Ambient audio: wind, water, distant bells.

V4 — Hailuo 3:4K 竖屏 仙鹤掠过(性价比镜头)

9:16 vertical, 10 seconds, ultra-detailed: a flock of red-crowned cranes glides past the camera in front of an enormous vermilion palace gate suspended above a cloud sea, slow lateral tracking shot, feathers backlit by warm sunset light, volumetric haze between layers, far palaces fading in atmospheric perspective. Ambience: high wind, crane calls.

V5 — Runway Gen-4.5 + References:仙女回眸(角色一致性镜头)

Reference: character plate provided. Medium close-up, a Chinese goddess in scarlet and gold hanfu turns her head over her shoulder toward camera, hair and sheer veil settle with believable inertia, gold filigree hairpins catch a warm rim light, background: cloud sea and distant palace spires in soft bokeh. Camera: locked-off 85mm, subtle handheld breathing. Keep face, hairline, jewelry, and robe exactly as in the reference. Light: warm backlight plus soft frontal fill.

V6 — Grok Imagine 1.5:480p 草稿试拍(构图与运镜试验)

Draft, 480p, 8s, 9:16. Fast low-angle push-in on a jade stairway leading to a giant sky-gate, god rays sweeping, mist rolling down the steps, two tiny robed figures ascending. Text-to-video, no dialogue, ambient wind.

V7 — Higgsfield(Soul ID 锁定 + Cinema Studio 运镜)

Soul ID: <仙女身份>. Shot: crane-up revealing a celestial terrace, then slow orbital move around the subject as her gauze sleeves lift in the wind; camera executes a 30-degree orbit at constant speed, no cut. Lighting: golden hour rim light, cool ambient fill. Keep identity, costume silhouette and jewelry consistent with Soul ID across all shots.

3. 一致性方案(国外链路的四套主流做法)

  1. 参考图锁脸(Reference / Ingredients / Elements)
    • Kling Elements:角色用 2–4 张不同角度图;3.0 Omni 还能上传 3–8 秒视频同时锁「脸 + 声音」。官方建议 5–10 张、含正面/3-4/侧面/不同光照/一张特写。
    • Runway Gen-4 Image References:单次最多 3 张生效,官方建议用中性均匀打光的素材;先出「character plate」审批,再进 Gen-4.5 出视频。
    • Veo 3.1 Flow Ingredients / Vertex subject reference:1–3 张参考图 → 8 秒片段;Flow 的 Characters 支持 1–2 张外观图并绑定语音。注意:Quality 档不支持 Ingredients,只有 Lite/Fast 支持。
    • Seedance 2.5:参考预算最大(30 图 + 10 视频 + 10 音频),适合把「身份/服装/场景/运镜/音色」拆成不同参考各司其职。
  2. 角色表(character sheet) 标准做法:一次性出「正面全身 + 3/4 侧 + 侧面 + 特写 + 服装平铺」四联图,作为项目固定资产。国外短剧团队普遍把它当「选角照」用,任何镜头都从这张表出发,而不是从文字描述出发。
  3. 首尾帧插值(first/last frame)
    • Seedance 2.5、Kling(首尾帧控制)、Veo 3.1(首帧/尾帧)均支持。做仙宫最实用的组合是:同一张图的「未加光」与「加光」两个版本做首尾帧,让光在镜头里生长。
  4. 身份训练(Soul ID 类)
    • Higgsfield Soul ID:上传 20+ 张参考图训练一个持久身份,之后在平台内所有模型(Seedance 2.0 / Kling 3.0 / WAN 2.6 / Veo 3.1 / Gemini Omni Flash)复用,不必每次重传参考。这是目前国外「AI 古风美人账号」做规模化最省事的一条路,属于商业化的关键差异点。

多镜头同一角色的实操纪律(来自 2026-09 的第三方对照测试):


4. 音画


5. 后期与分发

后期

分发与算法偏好(2026)


6. 成本:订阅价格与单条视频折算(信息时间点 2026-08 ~ 2026-09)

订阅/额度

平台价格(时点)额度说明
MidjourneyBasic $10/月(年付 $8)、Standard $30($24)、Pro $60($48)、Mega $120($96)按 GPU 时间计费:Basic ~3.3h 快速、Standard 15h + Relax 无限、Pro 30h + Stealth、Mega 60h;额外快速时长约 $4/小时(2026-04 口径)
Kling AI首月 Standard $6.99→续费 $8.80(660 credits);Pro $25.99→$32.56(3,000);Premier $64.99→$80.96(8,000);Ultra $127.99→$159.99(26,000)消耗:720p 无音频 6 credits/s、1080p 无音频 8、1080p + 原生音频 12、4K 直接 30 credits/s。月 credits 不滚存
Kling API1 unit=$0.14;Kling 3.0:720p $0.084/s、1080p $0.112/s、1080p+音频 $0.168/s、4K $0.42/sAPI 失败任务不扣 unit(订阅侧失败仍扣)
RunwayStandard ~$15/月(625 credits)、Pro ~$35(2,250)、Max ~$95(9,500)Gen-4.5 约 12 credits/秒(10s 1080p ≈ 100–150 credits)
Veo 3.1(Vertex)720p 无音频:Lite $0.03/s、Fast $0.08/s、Standard $0.20/s;带音频 $0.05 / $0.10 / $0.40Flow 免费档每天 50 credits(Lite 10 credits/次 → 每天最多 5 次 Lite)
MiniMax H3(Hailuo 3)768p $0.08/s、2K $0.13/s;fal 口径 480p $0.05 / 768p $0.06 / 2K $0.13 / 4K $0.16参考图前 5 张免费,之后 $0.04/张
Grok Imagine Video 1.5480p $0.08/s、720p $0.14/s、1080p $0.25/s;图片输入 $0.01/张1–15s,默认 480p
HiggsfieldBasic $9/月(120 credits)、Plus $49(1,000)、Ultra $129(3,000);也可买 credit pack(80 credits $5 起,需有效订阅)共享模型单价:Kling 3.0 约 $1/条(720p 10s)、Seedance 2.0 约 $3/条、Veo 3.1 8s 约 $2.9
ElevenLabsStarter $5、Creator $22、Pro $99、Scale $330、Business $1,320Free 无商用权
SunoPro $8/月、Premier $24/月(免费档 v6-mini)2026-09-11 v6 发布
Topaz Video AIPersonal $299/年、Pro $699/年2025-10 起订阅制

单条竖屏视频折算成本(15 秒、5 镜头、含 4 次重试)

⚠️ 成本口径提醒:不同来源的 Veo 3.1「每有效秒」差异极大($0.17 ~ $5.5/秒),差异来自是否把重试、是否用 Vertex 计费、Resolution/音频档位混在一起。报价前务必用「有效秒 × 重试系数」自算,不要直接引用博客的单秒数字。


7. 真实账号 / 案例(3 个)

案例 1|「栖光」(陈爱军,成都眼镜店老板)——「中式天庭」,海外 500 万+ 播放

案例 2|@灵筠仙逸(完美世界,抖音,广州)——国风 AI 美人账号

案例 3|Hanfu AI 变装(TikTok,2026-01 起全球扩散,印尼/东南亚媒体跟进教学)

另一个可参考的邻近案例(非仙宫题材,但方法同构):日本一位创作者用 AI 做「AI 女神」账号,7 天冲到 120 万点阅,港媒报道其成本约 $156、月收入达 10 万(报道未说明币种,且未经独立核实)。来源:ezone 报道(未证实)


8. 落地建议:一条最小可行的国外链路 SOP

  1. 定审美:MJ V8.1(alpha)跑 30–50 张仙宫图,建 moodboard,固定 --sref --sw;确定「一个主色 + 一个点缀色」(例:月白 + 朱红鎏金)。
  2. 建资产:仙宫类做「5 个固定机位模板」;美人类做「角色四联表 + 2–4 张角度图」。
  3. 出图:MJ 定稿 →(可选)训练轻量 SDXL/FLUX LoRA → ComfyUI 批量出分镜。
  4. 出视频:仙宫空镜用 Hailuo 3 / Seedance 2.5(便宜 + 长镜头);多镜头叙事用 Kling 3.0 Omni;角色一致性镜头用 Higgsfield Soul ID / Runway Gen-4.5 + References;草稿用 Grok Imagine 480p。
  5. 音画:优先把 audio block 写进视频提示词;需要旁白用 ElevenLabs(≥Starter);配乐用 Suno v6(guqin/guzheng/bianzhong/pentatonic)。
  6. 后期:Topaz 放大 → DaVinci 降全局暖调、分离色相 → 9:16 裁切 → 8–15s、4–6 镜头、前 2 秒给最大景别。
  7. 分发:先发 Reels/TikTok 测 hook,再把高完播的搬运到 Shorts;标题用「提问式」+ #ChineseHeaven / #Chinamaxxing。

9. 不确定项与注意事项

主要来源

下载此文件