Are the Creations from an Artificial Entity also Artificial?

Published in

Keio SFC Interaction Design class — 2022 Spring

6 min readAug 2, 2022

Artificial minds are far from the complexity of the human mind, and creations from such artificial entitites lack in the creativity capable of human minds. This is true as of the 21st Century, and will likely continue to be true without a deeper understanding of the human brain.

With that being said, I believe that it is certainly possible for AI to be creative. This is because creativity is a non-quantitative property created by people to define the uniqueness of a person’s ideas, thoughts and creations, which is a result of a combination of that person’s experiences in life, their current environment, genetics, the combination of the billions of neurons activated in the brain at the time — all of which are the basis of many machine learning models.

Whilst the creativity of current machine learning is still a point that can be argued against, we can still use AI to assist us as a tool for us to be creative.

Theme

— A visual comparison between two common manga genres, Shoujo and Shounen

Manga as a medium of art is very broad — traditionally drawn in black and white on a paper canvas with pen, the tools used are ever changing and are part of a Mangaka’s style and a compound of the world building to tell their story.

Distinct style

Mangaka with years of experience in comic publication have art styles so distinct that even non-readers with slight exposure to the Manga can recognize the artist.

Below are some works produced by Mangaka that are not part of a manga published by them yet the artist is still easily recognizable:

Hirohiko Araki’s (Author of Jojo’s Bizarre Adventure) take on “Kenshiro” from “Fist of The North Star” {1}

Eiichiro Oda’s (Author of OnePiece) take on “Ado”, Japanese Singer {2}

Weekly Jump 40th anniversery poster by Yusuke Murata (Illustrator for Eyeshield 21, One Punch Man) {3}

Anakin and Padme from Star Wars drawn by Takeshi Obata (illustrator of Death Note, Hikaru no Go, Bakuman) {4}

Categorization

Every Mangaka and Illustrator have their own styles, so it is difficult to simply separate them into different categories. However, there are two popular manga genres that have very different art styles, and are somewhat consistent in the way characters are drawn. Those are Shoujo Manga (少女漫画) and Shounen Manga (少年漫画).

I decided to try and extract some characteristics of the two art styles using a Generative Adversarial Network (GAN).

Training Data

As for collecting the training data, I had two approaches in mind:

Collect portraits of characters drawn in their respective genres, to produce portraits in the style of the genre.
Collect any image picked up in a search engine using the keywords 少女漫画 and 少年漫画.

Approach 1 would be closer to the idea I had when I chose this theme. An advantage of this would be that portraits and character faces tend to be the most defining difference between the two genres. However, I disregarded this as it was difficult to collect a large enough set for training, and the portraits would all need to be similar in pose and dimension for the end result to produce a non-deformed portrait.

The chosen method was approach 2. Whilst this method would produce images that are uncanny to human perception, it would allow for the extraction of the genres’ style in a more general sense and could lead to some unexpected differences from my initial perspective. In addition to using the keywords of the respective genres, I mostly used the mangas’ cover art-these are in color, allowing us to observe color palette trends from the end result.

Example images in training data:

Training Results

GAN model used — mllab GAN model on runwayml.com

Collage of images produced using the Shoujo training data

Collage of images produced using the Shounen training data

Mostly human features, minimal color palette

Strong foreground character with detailed background. Broader color palette

Observations and Analysis

Even though it was expected that uncanny, non-human like characteristics would appear, it is still mildly disturbing to the human eye. That aside, here are some observations:

The images produced by the Shoujo training data is significantly less cluttered, and are composed of mostly human features such as hair and eyes — whereas the shounen image has more variation is features and the images are generally more “full”. I believe this is because Shoujo Manga as a genre is more focused on characters and their relations with each other, whereas Shounen Manga typically focus on other aspects such as world building and action.
Some text-like object was expected due to the training data, and is observed heavily from the Shounen image. Many of the images Shounen images produced even have an object towards the top of the images resembling the Weekly Shonen Jump logo. This is unsurprising as the training data did, in fact have many Weekly Shonen Jump cover images.

Color Palette Extraction

https://colordesigner.io/color-palette-from-image

Soft colors that generally use a hue between the reds and blues. Mix of warm and cool, but nothing strong or sharp. Lower saturation.

Sharp, higher saturation colors. Bold and defining, mostly composed of warm colors.

Conclusion

Although the images produced were uncanny and even slightly grotesque, I believe they managed to capture some of the key features from their respective genres.

Shoujo — mostly human features, with less clutter in the background to portray the human relationships in the story. Cover arts use lower saturation colors that are aesthetically soft.

Shounen — lots of detail to draw readers into the complex world building. Cover arts use higher saturation colors that feel bold and eye-catching.

Discussion

Are these creations by the model creative? It is difficult to say that they are, considering that these images are just generated based off of the experience given to it through the training data. Oddly however, that is also how us humans engage in creativity, just with a different level of complexity.

Whilst these images may not be creative, there certainly are some takeaways and applications. For example, data analysis of creative arts tends to be a difficult procedure due to the non-quantitative property of arts. With a model such as this however, trends and general rules of successful works become easily observable, and can directly be applied by aspiring artists looking for success in the same media form. In the particular case of manga analysis using the GAN model, we see a clear divide in color palette between the two genres of manga. Artists can then take inspiration from these successful color palettes, and analyze why these might be successful — it is not needed to only use these palettes, as such is the beauty of human creativity, the expansion of existing creations.

Reflection

Method of collecting training data was flawed — training data needed more uniformality, both within the same data set and in comparison with the other genre’s. It would allow for the GAN to produce images that are more consistent and recognizable to the human eye, thus allowing a better comparison of the two genres.