😼 We will make a childhood image of the cat for you who adopted an abandoned cat!
Recently, pet shops have emerged as a big social problem in the concept of buying and selling living things for money.
We want to prevent this phenomenon and encourage adoption of abandoned cats.
If a cat is adopted from an abandoned cat center, there are no childhood photos of the cat, u_u
I think we can make that memory for them!
Encouraging adoption of abandoned cats
- The cat's species must not change.
- The resulting product should resemble the cat in the input image.
- The result should look young.
Considering the above problems ,
various experiments were conducted to transform cat into kitten.
- OS : Ubuntu 18.04 or Colab
- GPU : Tesla V100-32GB
Training StyleGAN2 ADA tf model with custom dataset.
Fine-tuning the StyleGAN2-ADA tf model pre-trained with afhq cat using custom dataset.
Disentangled features can be identified through the result that changes according to the layer being swapped.
PCA for Latent Space Exploration
 eyes |
 ears |
 mouth |
 face features |
Partially adjust latent vector of Images projected into the latent space, with components found through PCA.
 cat image |
 projected image |
 adjust latent vector |
We created a kitten generate model through the Stylegan2-ada model.
We explored the latent space to find a latent vector that affects a specific feature.
Using this vector, the projected latent vector was further adjusted to generate the desired images.
In conclusion, we can convert cat into a kitten by combining and applying the components.
https://github.com/NVlabs/stylegan2-ada
https://github.com/harskish/ganspace
https://github.com/orpatashnik/StyleCLIP
@BOAZ-bigdata 17th Big Data Conference