SRVP‑FD

SRVP‑FD computes Fréchet distance between images or videos using the encoder of the
Stochastic Latent Residual Video Prediction (SRVP) model.

Installation

pip install srvp-fd          # via pip
uv pip install srvp-fd       # via uv (faster resolver)

Usage

import torch
from srvp_fd import frechet_distance, FrechetDistanceCalculator

# Images – shape: [batch, channels, height, width]
images1 = torch.randn(512, 1, 64, 64)
images2 = torch.randn(512, 1, 64, 64)

# Basic image comparison
fd = frechet_distance(images1, images2)
print(fd)

# Using a different pretrained encoder
fd_bair = frechet_distance(images1, images2, dataset="bair")

# Videos – shape: [batch, seq_len, channels, height, width]
videos1 = torch.randn(512, 16, 1, 64, 64)
videos2 = torch.randn(512, 16, 1, 64, 64)

# Different comparison types
fd_frame    = frechet_distance(videos1[:, 0], videos2[:, 0], comparison_type="frame")
fd_static   = frechet_distance(videos1, videos2, comparison_type="static_content")
fd_dynamics = frechet_distance(videos1, videos2, comparison_type="dynamics")

# Class-based API (weights loaded only once at initialization - more efficient for multiple calculations)
calc = FrechetDistanceCalculator(dataset="mmnist_stochastic")

fd1 = calc(images1, images2)                           # frame by default
fd2 = calc(videos1, videos2, comparison_type="static_content")
fd3 = calc(videos1, videos2, comparison_type="dynamics")

Comparison Types

For precise mathematical definitions consult SRVP paper (Franceschi et al., 2020) and the official implementation.

Gaussian‑mixture approximation: For the "dynamics" comparison we collapse the batch of per‑sample Gaussians into a single Gaussian using the closed‑form mean and covariance of an equal‑weighted mixture for the Fréchet distance computation.

Features

• Multiple comparison modes (frame / static content / dynamics)
• Pre‑trained weights downloaded automatically from HuggingFace Hub
• Supported datasets: Moving MNIST, BAIR, KTH, Human3.6M
• Single‑load, class‑based API for high‑throughput evaluation
• CPU and GPU compatible
• Covariance maths guarded for numerical stability

Fréchet distance formula

$$ d^2\bigl((m_1,,C_1),(m_2,,C_2)\bigr) = \lVert m_1 - m_2\rVert^2 + \mathrm{Tr}\bigl(C_1 + C_2 - 2\sqrt{C_1C_2}\bigr) $$

where $m$ are the feature means and $C$ the covariances.

Citation

If you use SRVP‑FD in your research or publications, please cite the original SRVP paper, which provides the foundation for this tool's encoder and latent representations:

This package builds on the encoder from the original SRVP model to define a set of Fréchet-based distances for evaluating image and video generation models.

@inproceedings{franceschi2020stochastic,
  title   = {Stochastic Latent Residual Video Prediction},
  author  = {Franceschi, Jean-Yves and Delasalles, Edouard and Chen, Mickael
             and Lamprier, Sylvain and Gallinari, Patrick},
  booktitle = {International Conference on Machine Learning},
  pages   = {3233--3246},
  year    = {2020},
  organization = {PMLR}
}

License

Apache License 2.0 – same as the original SRVP implementation.

Acknowledgements

Built on the excellent work of the SRVP authors:
Jean‑Yves Franceschi · Edouard Delasalles · Mickaël Chen · Sylvain Lamprier · Patrick Gallinari

• https://github.com/edouardelasalles/srvp
• https://sites.google.com/view/srvp/