search

anvoynov / GANLatentDiscovery

The authors official implementation of Unsupervised Discovery of Interpretable Directions in the GAN Latent Space
416 stars 52 forks source link

Awesome work! Suggest adding a video #5

Closed andrewjong closed 4 years ago

andrewjong commented 4 years ago

This work is great, your proposed approach using a matrix A and reconstructor network is quite unique. I'm wondering if it's possible to create a video such as the video created for GANSpace; GANSpace is a concurrent work for unsupervised latent space exploration, but they use PCA instead. I'm curious to know how these two concurrent works compare.

anvoynov commented 4 years ago

Hi @andrewjong, happy you like it! I have added the code for GIF creation that was utilized for the README animations generation, this certainly may improve usability, thank you for pointing. You can run it with the lines:

from visualization import gen_animation
gen_animation(G, deformator, 44, '/home/avoin/mnist.gif', r=8)  # digit rotation animation

We have not mentioned the GANSpace as it was published after our preprint, though we discuss the similar approach proposed in https://arxiv.org/abs/1812.01161 I agree that the comparison of different disentangled directions remains a challenging issue that can't be named to be fully-solved so far.