Anime, a form of animation which originated in Japan, is heavily focused on Japan. The word “anime” comes from the English “animation,” which is why it’s used in Japan to describe any type of animation. Outside of Japan, the term is usually used to describe a particular type of animation.
Art styles in anime vary widely depending on the artist or genre and audience. There are some stylistic elements that are common. Characters tend to be vibrant, colorful and drawn with great detail. Anime features are often exaggerated. The eyes of anime characters are often large and expressive. Hairstyles and colors are often extravagant, as is clothing.
It can be difficult to draw anime characters, and they are underpaid. A new tool that makes it easier to draw anime characters is on its way. However, the effects of this tool are not yet clear.
AI, meet anime
In the last year, generative AIs dominated the world. There is the ChatGPT, which is remarkably good at writing texts. And there are about half a dozen AIs that generate images. There was no surprise that anime also received some attention.
The new tool is not intended to create any new AI images. It is designed to enhance and complement the human abilities.
This paper discusses how AI can be used by general users to create professional portraits.
The researchers wrote in their study that the key to achieving high-quality animes is to convert rough sketches during the sketching stage.
Zhengyu Huang, a scientist from Waseda University and Japan’s Japan Advanced Institute of Science and Technology(JAIST), led the research team. Researchers focused on a single task: how to turn rough sketches into anime portraits.
The training arc
Abstract imagery is a specialty of many AI image-generators. They most commonly use generative adversarial network (or GAN). GANs are machine learning models that create data samples similar to input data. In the context images, GANs are able to generate new images which look as though they were generated from the same distribution of training images. The model’s two main components, the discriminator and the generator, are what make up the ‘adversarial part’.
- Generator This network transforms a random noise vector into an image. The images are initially random but, over time, the generator begins to produce images that are similar to those in the training data.
- Discriminator : This network accepts an image as input (either a real image from the training data or one produced by the generator), and then outputs the probability that the image input is real (i.e. from the learning dataset). Its primary role is to evaluate the quality of images generated by the generator.
Researchers used an adversarial network to create new images using a Style Generative Adversarial (StyleGAN), a model of the highest level that is based on a generative model.
They used an unsupervised strategy. It means that AI matched the features directly, without labeling them.
“We trained a styleGAN encoder as a teacher first, using an already-trained StyleGAN. We then simulated the drawing of images generated without any additional data in order to train the sketch coder for incomplete progressive sketches. This allowed us to generate portrait images of high quality that are aligned with the teacher encoder’s disentangled representations.
This allows the user to have more control over the generated images by allowing them to adjust the parameters.