The art style that AI has
Before starting to discuss with what kind of image AI can create, have you ever heard of art-style called “Dreamcore”? Dreamcore is a type of art that illustrates surreal scene that cannot be in reality, or include contradiction between the characteristics of objects drawn in the image. There are similar types of art style such as Wierdcore and Traumacore, but all these styles create chaotic image that is dreamlike.
AI’s art style
Image created by AI using machine learning can create what the AI learned from their data used for training. Meaning, that these AI are expected to illustrate image, understanding the feature of the data given. If AI was trained with image data of train, the vehicle, we expect the AI to understand what is “train” as appearance and be able to create new image. However, the reality does not go that simple.
When we see image created by AI, we see that they are learning a pattern from given data. Training using data of human face will make AI able to create various face that does not exist, or is not in the data. They would have skin color, hair, eye, nose, mouth, and more multiple elements that we human would recognize as “human face”. However, that is possible because that AI is speciality in human face. The wider the usage, the vague the image become.
Once AI starts to create image in multiple genre, the elements starts to become vague, leaving only the silhouettes. The image above is created by diffusion model website DALLE mini. The man in front of screen has hair and correct silhouettes as human, but does not have any eye, mouth, nose, nor ears. The reason is simple; AI can’t understand the human has those elements as nature. Since the usage of DALLE mini is not only creating human face, its understanding to one field became neglect.
If the user expected realistic picture, this image isn’t enough. Though from art perspective, this image can be said as Dreamcore like art. It may be possible to create realistic picture by adding more data for training, but thinking about the number of data needed, it would not be a pragmatic idea. Therefore, at least for now, this can be said as one art style that AI has.
Bias in Training
When the usage widens, AI’s illustration does not always become vague in a way by losing some characteristic or elements from image. There are trend that can be seen that AI learn pattern from data in a way we don’t do. Since AI learns from pattern that can be found in given data, there are bias that AI learns multiple elements as one set, despite if those have connection in real. For example, there tend to be mountain or forest illustrated when word “train” are inputted in diffusion model.
We often talk about bias in data (such as color, value, culture) when we say “Bias in AI”, but there might be bias or pattern in a way that AI learn from data given. Since they can understand the “pattern: train”, but not the “train” as single object itself. This can be said as way of thinking that AI only can have.
Can this be Creativity?
This Dreamcore like art is created by unique perspective that AI has in its learning logic. It includes idea, thought that we human do not have. Yes, from perspective of technology, this specification is unwilling and shows need of improvement to create more accurate AI. If time passes by, AI technology will develop even more, creating more accurate image, and deeper understanding to human society by AI. Therefore, this “art style” that AI has is something that will be gone in future. The fault itself that AI has cant be said as creativity, but using the fault in some kind of art can be said as one type of transformative creativity.
