光线
交代方向、色温、软硬和衰减。参考里有左侧窗光与暖顶灯,才使用混光;单窗光也可以成立。
TIKTOK SHOP / SEEDANCE REALISM SKILL
Seedance 真人感出海带货视频。我的方法是同时处理「一眼 AI」与「一眼广告」:让皮肤、手部和商品成立,让光线、生活痕迹与镜头符合真实使用场景。
01 / ARCHITECTURE
先核对当前生成入口,再锁人物、商品、房间和声音。四段脚本可以在同一机位完成;检查失败时,按缺陷时间点只调整一个主要变量。
箭头表示执行顺序;虚线是缺陷返修。首次结果后最多追加 3 次有依据的重试。手机上可横向查看完整架构图。
02 / FOUR PILLARS
交代方向、色温、软硬和衰减。参考里有左侧窗光与暖顶灯,才使用混光;单窗光也可以成立。
挑 2–3 件与动作有关的物件:半杯咖啡水痕、半拆快递袋、随手放下的纸巾。位置随镜头保持一致。
保留自然毛孔、局部肤色差与眼下阴影。脸、颈、手臂的曝光和软度与环境匹配,避免磨皮与统一颗粒。
轻微手持运动、动作中的少量模糊、略偏心构图、竖屏 9:16。商品展示时仍然要看得清楚。
03 / A 20-SECOND TAKE
20 秒模板。台词先朗读估时,再给操作和观察留停顿。中文可以用 4–12 字短句;出海口播按目标语言写自然短句。
| 时间 | 作用 | 执行要点 |
|---|---|---|
| 0–2 秒 | 钩子 | 从动作中段开始,用一个具体问题或人群切入;数字须有真实依据。 |
| 2–7 秒 | 使用场景 | 说清什么时候遇到什么问题。运镜如 2–5 秒间轻推至 1.05×,避免指令打架。 |
| 7–17 秒 | 商品展示 | 握住、放下、推近,展示一项可核实功能。避免复杂翻转与多指精细操作。 |
| 17–20 秒 | 下一步 | 一句话说清真实购买路径。留在自然结束动作里,不虚构折扣、库存或效果。 |
04 / COPY & ADAPT
三段锁定置顶,补充独立商品锁。模板按目标市场设置语言;中文逐音节口型只用于中文口播。替换全部方括号后使用。
[Character Lock — CREATOR_NAME]
Use [CHARACTER_REFERENCE] as the identity authority.
[Adult age range, body proportions, distinct facial features.]
[Hair, eyewear if present, clothing, accessories.]
Preserve the same facial structure and body proportions throughout.
[Allowed posture and simple gestures.]
The person is [a demonstrator / the actual authorized creator]; do not invent
customer experiences, professional credentials, affiliations or results.
[Product Lock]
Use [PRODUCT_REFERENCE / SKU] for product geometry, color, material and packaging.
Keep the real label and logo exactly as shown; do not invent lettering or claims.
Demonstrate only [VERIFIED PHYSICAL ACTION / OBSERVABLE FEATURE].
[Setting Lock]
Use [ROOM_REFERENCE]. [Room layout, camera direction, product location.]
[Window/light position, practical light temperature and soft/hard quality.]
[Two or three relevant objects with ordinary traces of use and fixed positions.]
Keep these spatial relationships consistent throughout.
[Voice Lock]
Language: [TARGET LANGUAGE]. Accent: [TARGET MARKET ACCENT].
Voice: [Age range, timbre], relaxed everyday speech at [TESTED PACE].
Speak only the quoted dialogue. No narration from a second speaker.
Mouth movement follows the exact dialogue timing in [TARGET LANGUAGE];
the mouth is closed and still during silent beats.
[If verified and needed: Brand NAME is pronounced PRONUNCIATION.]
[Capture and texture]
Vertical 9:16, casual smartphone home-video feel.
Handheld with subtle natural movement, slight motion blur during movement,
composition slightly off-center. Keep the product readable at the demonstration.
[Actual light direction, temperature and softness], with natural falloff.
Skin shows natural texture with pores and subtle imperfections,
natural local differences in skin tone, not over-smoothed plastic.
Match skin exposure and focus softness to the room and the reference.
Hands have natural anatomy with five fingers and correct proportions.
Eyes remain stable and focused on [ONE GAZE TARGET], without flickering or jitter.
[Timeline — 20-second target; adapt to the verified generation limit]
0–2s | HOOK
[Start mid-action: a simple real interaction with the product.]
Dialogue: "[Short market-appropriate hook, with no unsupported numbers.]"
Camera: [Fixed casual angle / small motivated movement.]
2–7s | USE CONTEXT
[One concrete situation showing why the feature matters.]
Dialogue: "[Natural short sentence about the situation.]"
Camera: [For example: subtle 1.05x push-in from 2s to 5s.]
7–17s | DEMONSTRATION
[Grip the product, place it down or push it closer; show one verified feature.]
Dialogue: "[A factual observation supported by the SKU.]"
Keep hand–product contact physically coherent. Preserve the exact packaging.
[Leave a short silent beat to see the action.]
17–20s | NEXT STEP
[Finish the action naturally; retain product visibility.]
Dialogue: "[Actual purchase path or neutral next step.]"
[Quality directives]
Maintain the locked person, product, room and light-source relationships.
Hands have natural anatomy with five fingers and correct proportions.
Skin shows natural texture with pores and subtle imperfections.
Eyes remain stable and focused; synchronize the mouth to the exact dialogue.
Documentary realism throughout, not an AI-art aesthetic, 3D render or illustration.
Use ordinary practical lighting and casual framing, not polished studio staging.
No newly generated captions, floating words, fake background lettering,
unrelated logos, UI overlays, numbers or symbols.
Preserve the authentic text and branding already present on the product reference.
Unmarked background surfaces show their natural material texture.
Keep intentional subtitles out of the generated footage; add them as a separate
post-production layer only if requested.真实包装上的标签与 Logo 按参考保留。「NO text」用于限制新增乱码、字幕浮层和背景假字。重要否定配正向描述,重复数量不代表保证遵守。
05 / REPAIR LOOP
| 症状 | 下一次优先调整 |
|---|---|
| 一眼广告 | 先改实际光源或构图,再检查生活物件是否与动作有关。 |
| 塑料皮肤 / 过锐 | 保留局部肤色差、毛孔与参考软度,去掉 beauty、retouched 等冲突词。 |
| 背景漂移 / 光线假 | 绑定同一房间参考,锁窗户、灯和物件位置,保持曝光与白平衡。 |
| 跨镜头变脸 | 强化一致身份参考,减少大角度切换;入口支持时用首尾帧锚定。 |
| 手指异常 / 接触穿帮 | 简化为握住或放下,减少遮挡,明确手与商品接触位置。 |
| 眼神抖 / 口型漂移 | 眼睛固定一个注视目标;核对语言和实际读时,缩句、放慢。 |
| 乱码文字 | 锁真实包装,约束「不新增文字」,其余表面描述为自然材质。 |
| 分段色调跳变 | 先对齐光线、曝光和动作;0.6–0.8 秒转场只在需要时使用。 |
06 / BEFORE GENERATION
07 / CAPABILITIES & EVIDENCE
2026-09-09 核对。15–25 秒是这套方法的成片目标;具体模型、账号和入口支持什么,需要在执行时重新读取。
| 项目 | 当前核验结果 |
|---|---|
| 单次时长与多镜 | 官方 Seedance 2.5 已发布,支持单次最长 30 秒与场景切换。 |
| 参考素材 | 最多 30 张图片 + 10 段视频 + 10 段音频,共 50 份;不是 50 张图片。 |
| 延长 | 官方项目页列出两次延长;固定总长 90 秒未在本次核对中证实。 |
| 分辨率 | BytePlus LAS 的 2.5 接口列出 480p / 720p、24 fps、4–30 秒,不能套用原生 4K 承诺。 |
Seedance 2.5 · official release ↗ · Seedance 2.5 · project page ↗ · BytePlus LAS · video generation API ↗
原始笔记的 N=12、眼神 95、角色一致性 92、手部 89,以及 47% / 18% / 0%、结尾约 3 倍、改善 12% / 8% 等数字,以原文表述保留在英文包的 SKILL.md 中。它们尚无可复核原始样本、评分定义或对照,不作为本 Skill 的保证。官方 Seedance 2.0 提示词指南也明确提示:不能确保百分之百不生成字幕;该说明不能直接转成 2.5 的量化结论。
SKILL / ENGLISH EDITION · 2026.09.09
英文包包含 SKILL.md、README 和原始示例视频。架构图通过上方链接单独下载。
解压后,把 seedance-real-selling-video 文件夹放进 Codex 的 ~/.codex/skills/。在新任务中调用 $seedance-real-selling-video,并附商品参考、国家、语言与购买路径。
本 Skill 用于提示词与脚本编写;实际视频渲染需要另行调用生成工具。