Paper Image

Estimating the number of unique faces generative models can produce

Published on:

3 August 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Vishnu Naresh Boddeti,

Gautam Sreekumar,

Arun Ross

Bullets

Key Details

Proposes method to estimate capacity of generative face models in facial recognition feature space

Models population and identity distributions as hyperspherical caps in feature space

Derives analytical expression for capacity as ratio of population vs identity cap surface areas

Estimates capacity of StyleGAN, Latent Diffusion, and class-conditional models

Finds capacity reduces with stricter false acceptance rate thresholds

AI generated summary

Estimating the number of unique faces generative models can produce

This paper proposes a statistical approach to estimate the maximal number of unique facial identities that can be generated by different generative face models. The key idea is to model generated faces in a hyperspherical facial recognition feature space, and compute capacity as the ratio of population vs identity manifold volumes. Experiments on unconditional and class-conditional models yield reasonable capacity estimates, revealing trends across models and demographic groups.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up