[Sora培训内容] How does Sora keep a scene unchanged, and how do you design a consistent scene?

No story-driven content can do without scenes. The plot of the first video may look good, but as time progresses and you get to generating the second story, you will find it very hard to describe the same scene. Even with the same prompt or the same sentence, the scene is still different. So today we will share this knowledge: how to design scenes, how to use them, and how to achieve continuous scene consistency.

Designing Scenes

We can design scenes in SORA prompts, no matter how complex they are. This requires some knowledge of architecture, interior decoration, or spatial design, because such knowledge can explain more clearly the proportions, space, facilities, functions, etc., of a house or a street. Things designed with professional architectural knowledge are more realistic, more logical, and more aesthetically pleasing. If you do not know it, that is fine—you can also use GPT to help you design, or provide reference images or videos.

How to Use Scenes

The method is to separate people, objects, and scenes, then merge them. But note: these factors are all affected by style at the same time. Characters, objects, and scenes that appear to share the same style may all look different when combined. There are some techniques involved here.

Continuous Scenes


The difference between Sora and DALLE is that one is text-to-image and the other is text-to-video. In fact, both are images and both are concepts of frames, but video has an extra concept of time compared with images. It develops and changes every second. You can call it a dynamic shot; from a technical perspective, it is actually called spatiotemporal change, and the technical term is spacetime patches. Spacetime patches, and the spacetime patches of the entire sequence, are the core technology for inferring consistent composition and a consistent framework. Spacetime patches contain many elements; the following are the main ones.

• Timeline
  
• Physical environment
  
• Characters
  
• Events
  
• Actions
  

Let me give you an example. Everyone knows Along the River During the Qingming Festival, a very long scroll. After unrolling it, as we gradually look further along, the spacetime in it is constantly changing, and the picture naturally expands. The dividing line you perceive between each different image is exactly where the spacetime patches are about to change. So you can use spacetime patches to drive scenario changes; this act is called scene innovation. If you do not want to innovate, just tell it to use Along the River During the Qingming Festival as the scene and have a person walk from the left side of the painting to the right for one minute. It can express that very well. So you can have characters move in scenes you design, or in existing scenes.

How to use spacetime patches is another highly specialized skill. We also used it when making comics, but in DALLE3 its model does not support it very well; even so, it can still demonstrate consistency.




As demonstrated by the DALLE3 images above, with the two sentences before and after, when I was developing it I set the scene inference logic so that the story involving the two people always appears in the same scene.

My first sentence was "Ji Ziyue came to the supermarket," and the second was "met Ye Fan." It understood the process change in between: "came to," a movement event; "met"—this is the time change, which contains spacetime elements.

Sign Up to Learn

Regarding the entire Sora video production content, we have put together a complete course. At the same time, the moderator is also developing a prompt tool; when it is developed, it will be given together to students who sign up.




Course details: http://www.niubo1ai.com/thread-531-1-1.html


Of course, for those who do not plan to sign up, we will also gradually share more knowledge later.