Fréchet Video Distance

This repository contains an easy to use implementation of the Fréchet Video Distance (FVD) metric, developed by Unterthiner et al., for PyTorch. The implementation is largely based on the StyleGAN-V repository but was modified to work with two Tensors representing sets of videos.

Usage

To use the FVD metric copy the fvd_metric folder to your project. Then, import the metric using from fvd_metric import compute_fvd.

Afterwards, you can use the compute_fvd function. Here is an example:

fvd = compute_fvd(y_true, y_pred, num_samples, device, batch_size=4)

y_true and y_pred should have the same shape and the shape should be num_videos x channels x num_frames x width x height.

As computing the FVD metric for a large number of videos can be expensice, the num_samples parameter can be used to calculte the FVD on a subset.

Citation

If used please cite the original publication and the StyleGAN-V paper

@misc{
unterthiner2019fvd,
title={{FVD}: A new Metric for Video Generation},
author={Thomas Unterthiner and Sjoerd van Steenkiste and Karol Kurach and Rapha{\"e}l Marinier and Marcin Michalski and Sylvain Gelly},
year={2019},
url={https://openreview.net/forum?id=rylgEULtdN}
}

@misc{stylegan_v,
    title={StyleGAN-V: A Continuous Video Generator with the Price, Image Quality and Perks of StyleGAN2},
    author={Ivan Skorokhodov and Sergey Tulyakov and Mohamed Elhoseiny},
    journal={arXiv preprint arXiv:2112.14683},
    year={2021}
}

@inproceedings{digan,
    title={Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks},
    author={Sihyun Yu and Jihoon Tack and Sangwoo Mo and Hyunsu Kim and Junho Kim and Jung-Woo Ha and Jinwoo Shin},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=Czsdv-S4-w9}
}