pytorch-everybody-dance-now

paper : Everybody Dance Now - Caroline Chan, Shiry Ginosar, Tinghui Zhou, Alexei A.Efros

link : https://arxiv.org/pdf/1808.07371.pdf

refer source

• pix2pixHD : https://github.com/NVIDIA/pix2pixHD
• Pix2pix : https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix

Result

• result

• source

Input

• source : 512x512x3
• target : 512x512x3

---

Model

Discriminator

• multi scale discriminator (n_layer(scale) : 3, num_D : 3 )
  • Down sampling : average pooling
  • instanceNorm, spectralNorm

Generator

• 9 block of resnet
  • instanceNorm

VGG16

• for vgg loss

---

Loss

• PathchGAN Loss
• VGG Loss
• Feature Match Loss