Unsupervised Panoptic Interpretation of Latent Spaces in GANs Using Space Filling Vector Quantization

This repository contains the PyTorch implementation of the paper entitled "Unsupervised Panoptic Interpretation of Latent Spaces in GANs Using Space-Filling Vector Quantization", accepted at Transactions on Machine Learning Research (TMLR) in June 2025.

Requirements

Please create the conda environment to use this repository using the following lines in your terminal window:

conda create --name interp_sfvq python=3.9

conda activate interp_sfvq

pip install -r requirements.txt

To use CUDA, you need GCC 7 or later (Linux) or Visual Studio (Windows) compilers.

Demo

This directory contains the demo to test and compare interpretable directions found by our proposed method, GANSpace, and LatentCLR methods in intermediate latent space (W) of pretrained StyleGAN2-FFHQ.

Contents of this directory:

• comparison.pdf: Comparison of our method over 20 random vectors with GANSpace and LatentCLR
• demo.py: Code to create "comparison.pdf" file. You only need to change the num_random_samples and sigma_list in the code.
• demo_one_direction.py: Code to compare only one direction for one random sample. You only need to change the direction_name and sigma_list in the code.
• files.zip: Required files to run the codes.

To use the demo, please follow the steps below:

Please create the conda environment and its dependencies that mentioned above.

Also, please download the StyleGAN2-FFHQ pretrained model under this link. Or you can download the StyleGAN2 pretrained model named "stylegan2-ffhq-1024x1024.pkl" directly from NVIDIA website.

In addition, please extract the files existing in 'files.zip'. Please keep the pretrained model and extracted files in the same directory as demo.py.

In demo.py code, you only need to change the num_random_samples and sigma_list variables to test all interpretable directions over different random vectors and shift values ($\sigma$). The results will be saved in comparison.pdf file.

Interpretable Directions

• StyleGAN2-AFHQ: Discovered directions in pretrained StyleGAN2 on the AFHQ dataset.
• StyleGAN2-FFHQ: Discovered directions in pretrained StyleGAN2 on the FFHQ dataset.
• StyleGAN2-LSUNCAR: Discovered directions in pretrained StyleGAN2 on the LSUN Cars dataset.

In each directory, there is a code named manipulate.py that can be used to test the discovered directions. You can download the pretrained models from NVIDIA website.

We are still updating the repository!