第 11 课,共 21 课 · 第 5 单元: 生成影片

如何为每个镜头选择最合适的 AI 视频模型

AI 电影制作课程 · 5.1 5:18 视频免费

学会用六项标准和一次快速的并排测试,为每个镜头选对 AI 视频模型。然后降低成本:先做草稿,只给胜出的素材做成品。

章节(10)
在 YouTube 的课程播放列表中观看

视频为英语讲解。

在这一课中你将学到

  • 选择 AI 视频模型的六项标准:一致性、运动、人脸、声音、时长与分辨率限制、成本与速度
  • 如何对两三个模型做一次公平的并排测试并给结果打分
  • 录制时 ScreenWeaver 提供哪些视频模型,以及叙事场景推荐的默认模型
  • 如何用 Seedance 2.5 的草稿模式省钱:以 480p 做草稿,只把胜出的那一条完成为 1080p
  • 什么时候为整场戏生成一个片段,什么时候逐个镜头生成
  • 为什么要在点击生成之前,逐个镜头决定是否生成声音

选择 AI 视频模型,不是要找出市面上最好的模型。模型每隔几个月就会更新换代,所以这一课给你一套比模型活得更久的方法。你要为自己的影片给六项标准排出优先级(一致性、运动、人脸、声音、时长与分辨率限制、成本与速度),然后用相同的参考图和提示词,在两三个模型上跑同一个镜头,并排给结果打分。以 The Photo 最后的特写为例,你会明白为什么最贵的模型值得用在撑起结局的那个镜头上,却不值得用在一个简单的定场镜头上。你还会了解录制时 ScreenWeaver 提供的视频模型,包括 Seedance 2.5、Veo 3.1 和 Kling 3.0,学到一套控制 AI 视频成本的 “先草稿后成品” 策略,以及一条关于每场戏生成一个片段还是每个镜头生成一个片段的实用规则。

完整文字稿

英语旁白的文字稿。

0:04There is no best video model. There's the best model for a given shot on a given day. Models change every few months. So in this lecture, you'll learn a method that outlives them. The criteria that matter, a quick test to compare models, and a strategy to keep your costs under control.

0:23Here are the criteria. First, consistency. Does the model respect your reference sheets, shot after shot? Second, motion. Are bodies, hands, and objects physically believable? Third, faces. Some models refuse photorealistic faces as references. Fourth, sound. Some generate dialogue and ambiance, others don't. Fifth, duration and resolution limits. And sixth, cost and speed. Rank these for your film before you look at a single demo reel.

0:56Then, test. The same way every time. Take one representative shot with the same references and the same prompt and run it on two or three models. Put the results side by side and score each one on your criteria. Write the scores down. After a few films, that notebook will be worth more than any ranking you'll find online.

1:17Here's what a test can look like for the close -up that ends the photo. Model A keeps her face, but the tears look painted. Model B gets the emotion, but changes her scoff. Model C nails both, but costs twice as much. For the one close -up that carries the whole ending, Model C wins. For a simple establishing shot, it probably wouldn't be worth it.

1:42Here's what you'll find in Screenweaver at the time of this recording. Seedance 2 .5, the default, references, start and end frames, sound, and clips from 4 to 15 seconds, longer in some modes. Veo 3 .1, up to 4K and shorter clips. Cling 3 .0, in standard and pro versions, plus a motion control variant. Minimax H3, then Grok Imagine Video, both with sound always on, and Seedance 2 .0. My recommendation for narrative scenes today is Seedance 2 .5. But run your own test. That's the whole point of this lecture.

2:20Now the money strategy. Draft first, finish later. Seedance 2 .5 has a draft mode, at 480p, for quick and cheaper previews. You generate your takes as drafts, pick the winner, and then finish that one draft to 1080p. For even higher resolution, you can upscale a finished video with topaz, up to 4K. You pay full price only for the shots that made the cut.

2:49And check the limits before you plan. Some models cap the length of a clip, some only offer their highest resolution in short clips, and some always generate sound. Read the options of the video node before you write your shot list, not after.

3:06Next decision. One clip per shot, or one clip per scene. Some models, like Seedance 2 .5, can generate several shots in a single clip, with real cuts between them. The four shots of the photo were generated that way, as one continuous clip, so the light, the rain, and the actors stay identical across the cuts. Generating shot by shot gives you more control over each one, but more chances for things to drift between them. For short, intimate scenes, one clip per scene is often the safest choice.

3:43A practical rule, if a scene has only a few shots, one location, and the same characters throughout, try one clip for the whole scene first. If it has many shots, several locations, or complex action, generate shot by shot, or in small groups of shots.

4:00And think about sound before you click generate. When a model generates audio, it can deliver the dialogue and the ambience in the same take. In the photo, Thomas' line and the rain on the glass were generated with the picture. But if you plan to design the sound yourself, switch the audio off and you'll get a clean picture to build on. Decide shot by shot, and write it down in your shot list.

4:26One last tip. Match the model to the shot, not to the film. A dialogue scene, a wide establishing shot, and a fast action beat don't have the same needs. It's perfectly fine to use one model for most of your film, and a second one for the two shots where it struggles. Just keep your references and your art direction identical across both, so the audience never feels the switch.

4:51Your exercise, list your six criteria in order of importance for your film. Pick one representative shot, test it on three models with the same references and prompt, and score the results. Then decide your default model and your draft and finish strategy. In the next lecture, we'll write the prompt itself, the way a director briefs a crew.

在你自己的项目中试试这一课

免费打开 ScreenWeaver