大多数用于应用广告的 AI UGC 看起来都很假。

Most AI UGC for app ads looks fake. Not because the model is bad. Seedance 2.0 is insane. because operators skip the production principles and ship default output. Default = cinematic and polished = triggers ad detection in the first second. The fix is a prompt stack: → skin texture clause on every human face ("visible pores, natural unevenness, no filter quality") → anti-polish language that tells the model what NOT to produce → lighting gets its own dedicated clause, not buried in a list → environmental clutter in every background (charger cables, smudged glasses, scattered stuff) → subtle handheld motion so nothing reads as locked-off production footage → emotional arc direction so the avatar isn't doing flat neutral expression for 15 seconds → app screen treated separately — micro-glare, natural angle, not a product shot → exclusion stack at the end ("no text overlay, no ambient noise, no simulated UI") Close it out with 3 minutes of CapCut — grain overlay, iPhone color grade, trending audio, native captions. the operators building this into their production system are converting at rates that match real creator content. Everyone else is producing technically capable content that gets scrolled past before the hook lands. Comment "Playbook" to see how we do it.

大多数用于应用广告的 AI UGC 看起来都很假。 不是因为模型不行。Seedance 2.0 非常强。 是因为运营者跳过了制作原则，直接交付默认输出。 默认 = 电影感 + 精修 = 在第一秒就触发广告识别。 解决方案是一个 prompt 堆栈： → 每张人脸都加皮肤纹理子句（"可见毛孔、自然不均、无滤镜质感"） → 用 anti-polish 语言告诉模型不要生成什么 → 灯光单独写成一条独立子句，不要埋在列表里 → 每个背景都加环境杂物（充电线、模糊的眼镜、散落的物品） → 加入轻微的手持运动，避免读起来像固定机位的制片级素材 → 加入情绪弧线指引，让数字人不会连续 15 秒做平板的中性表情 → 应用屏幕单独处理——加一点微反光、自然角度，不要做成产品图 → 末尾加排除堆栈（"无文字叠加、无环境噪音、无模拟 UI"） 最后用 3 分钟的 CapCut 收尾——颗粒叠加、iPhone 调色、热门音频、原生字幕。 把这一套嵌入生产系统的运营者，转化率已经能比肩真实创作者内容。 其他人做出来的内容技术上看没问题，却在钩子落地之前就被划走了。 评论 "Playbook" 来看我们是怎么做的。