We explored how to describe simple scenes with one character doing something in my previous article. Now let us dive into some complex scene descriptions. If you haven't already, then please also read my article about camera angle and distance, because we will be needing that here.
Let us start with a simple scene first. Batman and Spiderman together doing the same action.
Hyperrealistic photo of two-standing-on-rooftop: (batman), facing (spiderman)
Good that we see both characters standing on the rooftop, but the scene is not quiet right. Why? because we did not specify the camera angle and because of that the AI model is free to chose one of any number of poses that fit the prompt. Also notice the change in the way the prompt is worded. By saying "two-standing-on-rooftop" we don't have to say what each character is doing. This is useful if all the characters are doing the same action in the scene.
Now let us try to add a camera angle to the same scene.
Hyperrealistic side-view photo of two-standing-on-rooftop: (batman), facing (spiderman)
Great! Now batman and spider man are actually facing each other. This is because of the camera angle. But did you notice something odd in this picture and the previous one for that matter?
Spiderman's cowl has horns just like that of batman's. Also look at his nose! This is the result of attributes bleeding because of training bias. Simply put, to the AI model batman is more popular and important than Spiderman because perhaps the checkpoint was trained with more sample of batman.
How to fix attribute bleed?
Before we continue, we have to fix the attribute bleed, otherwise as we introduce different actions or more characters, the bleed will spoil the scene we are trying to create.
Let me first show you what I mean. Let us remove the standing action from the main description, add it explicitly to each character as kneeling.
Hyperrealistic side-view photo of two-on-rooftop: (batman-kneeling), facing (spiderman-kneeling)
Where is the spider man? From the previous picture, we already know that attributes from batman was bleeding into spider man. With a slight shift, it can be the pose or it can be something else like batman looking angry and spider man crying like a baby, the attributes may start to bleed heavily. In this case, it was enough turn spider man into another batman. So do we fix that? This is a good time to learn more about inline weights.
Inline Weights:
The weight of prompt or group of prompts can be changed in different ways. How to do this depends on the UI (not the checkpoint) you are using.
Multiple brackets. For ex ((spider-man-standing)). Most of the UI recognize this format.
Explicit weight: For ex (spider-man-standing:1.1). Many UIs including CivitAI recognizes this.
Using + or -: For ex (spider-man-standing)++. Early Diffusion UI recognizes this.
If you are not sure then stick to using multiple brackets or explicit weights. With that let us try to fix the attribute bleed. Let us try to fix this by adding explicit weights.
Hyperrealistic side-view photo of two-on-rooftop: (batman-kneeling:1.0), facing (spiderman-standing:1.1)
Great! Now we have batman and spider man again in the scene. But is the scene perfect? Here is a clue! Look at spider man's cowl carefully. There is a very small horn. This because the weights I chose was still off. Batman is more important!. Let us fix that!
Hyperrealistic side-view photo of two-on-rooftop: (batman-kneeling:1.0), facing (spiderman-standing:1.2)
Now the scene is perfect! Learning to tweak the weights is very important to creating scenes with multiple character without the attributes bleeding too much. By the way, did you notice the shift in who is seen on the left side? In this case spider man appears first because we gave him more weight. This is important! The AI will most always give more importance to character or object in the scene with more weight.
Let us try to add a third character to the scene. I already know that it is not going to be right.
Hyperrealistic side-view photo of three-on-rooftop: (batman-kneeling:1.0), facing (spiderman-standing:1.2), next to (hulk-lying:1.0)
So what just happened here? The scene is heavily messed up! Yes there is a hulk lying but heavily messed up, there are two spider mans kneeling (with one wearing the belt of a batman) and then there is a batman who is standing, but looking green like hulk.
Now for the first time I am going to leave you with a home work. Please fix this scene and share your image with prompt and explanations on what was wrong with my prompt.