Part 8 - PIX2PIX Rabbit Drawer (Done on 11th July, 2017)
The last thing we tried was to make an interactive game between human and computer. The idea comes from QuickDraw. It’s a database of quick drawings, it recorded not only the images but also the drawing paths. So we wanted to teach the computer how to draw the next stroke based on what we human draw, to make the image more like a rabbit or other things.
So the training set looks like the following images:
Rabbit_Drawer_256_Test2, 4000 images, dataset here: 28_Rabbit_Drawer_256_Test2.rar
Each original image was decomposed into 5 steps and 4 pairs of images. To save training time, we turned the images back to 512*256. The script to translate data into images is here: https://github.com/ksteinfe/drawing_with_bots/tree/master/tools/qdrawer
Test video here:
Later, we trained 4 pairs of images seperately, and got 4 models. It seems nicer comparing to one model. Dataset here: 31_Dragon_Drawer_256_SubTest1.rar 32_Dragon_Drawer_256_SubTest2.rar 33_Dragon_Drawer_256_SubTest3.rar 34_Dragon_Drawer_256_SubTest4.rar