Anime Generative Model

2018 · Machine Learning

Github Repo

Task Description

• Use a GAN model to generate random anime characters

GAN

• Generator
  • Conv2DTranspose
  • BatchNormalization
  • LeakyReLU
• Discriminator
  • Conv2D
  • BatchNormalization
  • LeakyReLU

Result

Conditional GAN

• Use a conditional GAN model to generate a given text anime characters

• Generator
  • input: 128-dim noise, 12-dim (one-hot) hair color, 11-dim (one-hot) eyes color
  • 16 * residual block (input=image, output=bn(conv(relu(bn(conv(image, 64, 3))), 64, 3))+image)
  • 3 * upsampler (input=image, output = relu(bn(pixel_shuffler(conv(image, 256, 3)))))
• Discriminator
  • conv
  • leaky_relu
  • residual block = (input=image, output =leaky_relu(seperable_conv(leaky_relu(seperable_conv(image)))+image))

Result