The developer of 'Waifu Labs' who can make his own beautiful girl illustration with AI explains 'How AI makes my daughter-in-law'
'Waifu Labs ', which automatically generates two-dimensional beautiful girl illustrations according to the user's taste, will be adopted as the character illustration of the mobile game Arrowmancer because it can create beautiful illustrations. About. Since the number of illustrations automatically generated by Waifu Labs has exceeded 20 million, developer Ruwen Riu asked himself about 'How Waifu Labs creates beautiful girl illustrations?' It is explained in the blog.
Waifu Labs --Welcome to Waifu Labs v2: How do AIs Create?
Waifu Labs is a free service that started in 2019, and you can find out what kind of beautiful girl illustrations you can actually create by reading the following article.
I tried to make 'My daughter-in-law' with 'Waifu Labs' where AI automatically generates your favorite anime beautiful girl --GIGAZINE
Waifu Labs AI employs a type of unsupervised learning called Generative Adversarial Networks (GANs). In the case of Waifu Labs, AI that learns how to draw a picture called 'generator' and AI that learns how to distinguish between 'AI illustration' and 'human illustration' called 'discriminator' compete with each other. By letting them do, we will learn to create more accurate illustrations.
An interesting point when competing with these two AIs is that 'the interesting thing is that it is essential that the AI grows at the same speed, like human rivals. One AI is the other. If you overwhelm, both AIs will stop learning. '
The illustrations below are output by AI during the learning process, and you can clearly see that the output becomes an illustration as the steps are repeated (learning time becomes longer).
Step 0: At first, the AI does not understand what the illustration is at all, so the illustration output by the AI that has not learned anything is the noise itself.
Step 1024: As the learning time increases, something like an illustration will gradually be output, and you will see something like a face from this area.
Step 4096: Hair appears like an anime illustration.
Step 13516: It is around this time that you will be able to draw human-like illustrations.
Step 23961: It is possible to express ears and shoulders.
Step 40564: Each part of the illustration, which was vague as a whole, is now drawn in more detail, and 'from this area,' texture 'and' personality 'appear in the illustration,' Riu said.
Step 43636: Even with AI that seems to have been almost completed, the learning results suddenly become unstable, and the following horror illustrations may be output.
Step 50000: Completed!
In the case of Waifu Labs AI, it took about two weeks to be able to output this level of illustration.
What is interesting is that at this point AI is not just learning by imitating the appearance of the illustration, but also learning the features such as 'shape' and 'texture' to compose the original picture. For human illustrators, the 'shape' and 'texture' of illustrations are very important elements, and beyond the realm of conscious thinking, express unique 'shape' and 'texture' in intuition. I think there are many people. Mr. Riu calls these elements 'Mental Representation' and describes them as 'the ones that illustrators mature through lifelong practice and accumulation of life experiences.'
In the case of AI, this 'mental expression' will not grow at the end of learning. Therefore, it is also possible to directly check what kind of 'mental expression' AI has formed. The development team of Waifu Labs, including Riu, calls the status that determines the 'mental expression' generated by AI 'Latent space'.
By using this 'latent functional space', you can express the same character in different poses ...
It is also possible to express the same pose of the same character with different touches.
In addition, there are strange things in the 'latent functional space', and it seems that some of them generate the following huge head illustrations. This 'latent functional space' is familiarly called 'Big Head' by the staff of Waifu Labs.
In order to identify the status that can generate attractive characters from the 'latent function space' generated by AI, Waifu Labs extracts numerical values that control poses, colors, details, etc., and users 'generate their own illustrations'. Create an interface that allows you to do so. This allowed Waifu Labs to create 'my daughter-in-law' for users.
In addition, there is FID as an index to evaluate the quality of the image output by AI, but it is said that this FID could not fully meet the needs of Waifu Labs. Therefore, Waifu Labs has created its own evaluation criteria and states that it evaluates its own AI.