The authors study the problem of estimating the squared 2-Wasserstein distance between two multivariate Gaussians, which is exactly the FID when image embeddings are Gaussian.
New estimator achieves optimal sample complexity for FID evaluation
Chen et al. introduce Relative Taylor Debiasing (RTD), which estimates Fréchet Inception Distance with the theoretically minimal number of samples.
Top University
Ziyun Chen · Jerry Li · Kevin Tian · Yusong Zhu
University of Washington · University of Texas at Austin
Research Digest··3 min read
The authors provide a comprehensive theoretical analysis of the sample complexity of estimating the Fréchet Inception Distance (FID), a standard metric for evaluating generative models.
Why this paper
From University of Washington and University of Texas at Austin
In one line
Relative Taylor Debiasing estimates Gaussian FID with optimal O(d/ε²) sample complexity, avoiding the plug-in estimator’s quadratic dimension dependence.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§