ChatGPT
Perplexity
Claude
Gemini
最近一部用 AI 智能体流程拍摄的两分钟品牌片，从脚本到成片仅用三天，而传统同类拍摄大约需要两个月， 据 invideo 自己记录在案的制作案例. A ninety-second horror short pulled around four hundred separate video generations before it was done. Numbers like that used to be the whole story: AI got fast enough to make an entire scene, so why would you need a person in the room. The more interesting number is the other one buried in that same case study: more than 40 percent of the finished shots in one documented project weren't a single generation at all. They were stitched together from the strongest few seconds of several different attempts, a practice invideo's own team describes bluntly: "Prompt, eight tries, Frankenstein the keepers."
比起那些速度上的宣传，这个细节更能说明实际发生了什么。 一个能按指令渲染整个画面的模型，并没有让人的决策变得多余，只是把所有这些决策提前了，也让它们更快落地。 The workflow behind these AI-made shorts isn't one person typing a single magic sentence and getting a finished film back. It's a full crew of separate roles, a producer role holding the script and characters, a storyboard role visualizing shots before anything gets generated, a cinematography role taking direction like "hold the shot longer" or "track the actor through the doorway," a costume role, a production design role, each one scoped to its own job the way a real set is. The AI executes inside each of those roles. The structure of the roles themselves is still the same structure a film crew has always had, because someone still has to decide what the shot should look like before anything generates it.
最能证明瓶颈在于判断而非生成的，是素材完成之后所发生的事。有据可查的制作流程会专门设置一个环节：把拼接好的粗剪送回去做一轮审阅，专门查找节奏问题、声音问题，以及情绪基调偏差的时刻——正是这类东西 invideo 自己的工作流笔记称，这一步最容易被跳过，却能揪出人工剪辑师会遗漏的错误。这句话细想起来着实有些怪异：一次自动化审查竟能标记出训练有素的剪辑师所忽略的东西。这并不意味着 AI 更有品味，而是说品味如今被应用在了流程中的另一个环节——针对成片整体而非逐个镜头，而最终仍需有人来判断这条意见是否正确并据此行动。
So what is the creator actually making, if not the pixels. They're making the same thing a director has always made: the specific sequence of choices that turns raw coverage into a scene that lands the way it's supposed to. Which take gets used, which two generations get spliced together because neither one alone had the whole shot, which note from the critique pass actually matters and which one gets ignored, all of it decided by someone rather than generated by anything. None of that shows up if you only look at "who generated the footage," and all of it is the actual difference between a finished piece of work and four hundred generations sitting in a folder. 模型能生成整个画面。 它依然无法独立判断，哪个版本的画面才值得保留。