I liked the weird surrealistic yuru-characters generated with the StyleGAN. At the same time, however, I also wanted to incorporate more of the artist's hand instead of relying on canny edge detection. Thus I decided to make a new dataset to train Pix2PixHD.
Using the StyleGAN trained on Yurukyara 1, I generated a 1000 mascots, from which I picked 506 to redraw in a way that's more inline with how I usually draw and animate. Trainingset A characters are made of simple shapes and sketchy lines with occasional cross-contours. Dark Blue is used for the body and black for the face. |
(I also tried a green screen version for trainingset B, because of Pix2PixHD's lack of transparency. Although it was not particularly helpful, since multiple characters also have green colors and there's not really one specific color that's entirely excluded in the set.)
|