离谱！打死我都想不到，有公司愿意把这种东西公开出来。 AI 视频圈最没用的就是"只放成片"，成片可以随便给人看，但工程文件是命根子。 Higgsfield 这次把 Originals 的工程文件全开源了： 每个镜头的完整 prompt、角色参考图、音频母带、生成参数，点开项目全能看。 我翻了一圈，学到了很多： 角色先做成一整套参考图锁一致性，每个镜头单独一个文件夹，哪个镜头坏了只重跑那一个。 看完你就明白了，大片是把 AI 当真剧组管出来的，写一条 prompt 等奇迹这种事从来不存在。 挑了个我最喜欢的镜头，提示词原文长这样： Seedance 2.0 on Higgsfield AI Full open-sourced prompts & assets below: Prompt: Slow-motion handheld close-up, 16:9. Shaky organic handheld shot shoving in close on an older grey-haired man in a soaked white shirt and tie, chest-deep in churning green-grey flood water clutching a glass bottle, mouth open mid-shout — while all around him a dense crowd churns: a woman shoves past his shoulder, two men wrestle a floating chair behind him, someone falls backward into the water beside him sending up a slow sheet of spray, more panicking bodies wade and thrash in the background. In slow motion water explodes across the whole frame, droplets suspended everywhere. Dynamic jostling handheld movement at water level in the thick of the crowd. Sickly institutional palette — seafoam green, cream, muted teal, cold reflective water. All faces anatomically correct, stable, realistic, genuine emotion, no warping. Heavy coarse analog 35mm film grain baked into every frame, thick and consistent, grain crawling over highlights and shadows, like a scanned celluloid film print, never clean, never digital-looking. Flat overhead fluorescent light, dramatic slow-motion intensity, cinematic photorealistic, ultra-detailed.