[Sora培训内容] Sora’s principles of consistent creation: what settings are required to achieve consistency?

When you learn about consistent creation, you enter a relatively complex new field. Many people find that creating one plot with a single sentence is easy, but when you use another sentence to create the subsequent plot, you will find that the characters have changed, the style has changed, and the scenery has also changed. So how can you create a consistent plot? Today in the NIUBOYIAI community you can learn this knowledge. Now let us begin to explain it.

In the entire image, what we need to control is not only the characters. If only character consistency is achieved while other elements remain unchanged, the image will still look terrible. So first we need to understand compositional elements. A complete story image consists of three main elements, several secondary elements, and several supplementary elements.

• Characters
  
• Scene
  
• Style
  
• Visuals
  
• Camera
  

Referring to the official example, without changing the character, we change other elements, such as the scene. You will find that not only has the scene changed, but the character has changed too. The reason is that you did not specify what the character must look like. Their height and weight, age, skin color, hairstyle, face shape, etc. were not clearly defined. Therefore, when we say "a person," if there is no very complete description, it is a random person.





We can regard the prompt in the demonstration case as a template and analyze its element completeness.

Extract elements

• A woman
  
• Green dress and sun hat
  
• Antarctica
  
• Winter storm
  

Element analysis

• Main element. Character {woman, clothing, missing, missing, missing, missing}
  
• Main element. Character. Supplementary elements {behavior, action}
  
• Main element. Scene {Antarctica, winter storm}
  
• Main element. Style {missing}
  

Through analysis, we can clearly find that the description in the example is incomplete and missing a lot of content. For example, the character only says [a woman, green dress and sun hat], and there is no more description about her, such as age, skin color, hairstyle, face shape, eyes, nose, body proportions, etc. This means that the next time [a woman, green dress and sun hat] is used, there will be great randomness, as shown in Figure 2: her skin color, nationality, age, and clothing have all changed.

Over the past year or more, we have encountered this problem when using dalle to draw images. It cannot fully understand what we have in mind. You must tell it a very detailed and complete description to make the same thing more similar. If you say "use the same person," that also does not work; it will still make mistakes. Therefore, to use it well, deeper research is needed. It is not just a matter of prompts; you also need to understand its receiving and feedback habits and the logic of the model. If you have just come into contact with it, solving this problem may take months or even one or two years to understand.

Our image consistency research results
Video demonstration: http://www.niubo1ai.com/thread-451-1-1.html
Tool test: https://chat.openai.com/g/g-KDJWWgCHd-niuboyiai-comic-assistant

We spent several months specifically researching consistent comics, and finally figured out its processing procedure.






So currently there are two methods for Sora consistency:

1. When making a video, you need to use a way that the robot can understand to fully tell it the characters, style, and scene, and keep them unchanged.
2. Use our comic assistant to generate comics, and then use Sora to realize image-to-video.

To master this technology like the moderator, you need to master the following knowledge. This is practical experience we have summarized over half a year.



Course introduction: http://www.niubo1ai.com/fidtype-53-type-8-1.html

This cannot be clearly explained by just a few articles and some examples. In response to this demand, we have created a set of teaching knowledge. Friends who want to learn this content and quickly master video control capabilities can choose to sign up for the course.
For this set of tutorials, we have spent a lot of time, manpower, and material resources, so it is paid content.

Of course, in the follow-up time we will gradually share the chapter knowledge from the course, so that everyone can learn it here.