On Training Sample Memorization: Lessons from Benchmarking Generative Modeling with a Large-scale Competition

TL;DR

This paper critically evaluates the reliability of common metrics for generative models by analyzing a large-scale competition, revealing widespread unintentional memorization and proposing a new metric, MiFID, to better assess genuine generative quality.

Contribution

It introduces MiFID, a memorization-aware metric, and provides a comprehensive analysis of memorization issues in generative models through a large competition and manual inspection.

Findings

Unintentional memorization is prevalent in popular generative models.

MiFID effectively detects memorization in generative outputs.

Analysis of top models reveals common forms of memorization.

Abstract

Many recent developments on generative models for natural images have relied on heuristically-motivated metrics that can be easily gamed by memorizing a small sample from the true distribution or training a model directly to improve the metric. In this work, we critically evaluate the gameability of these metrics by designing and deploying a generative modeling competition. Our competition received over 11000 submitted models. The competitiveness between participants allowed us to investigate both intentional and unintentional memorization in generative modeling. To detect intentional memorization, we propose the ``Memorization-Informed Fr\'echet Inception Distance'' (MiFID) as a new memorization-aware metric and design benchmark procedures to ensure that winning submissions made genuine improvements in perceptual quality. Furthermore, we manually inspect the code for the 1000 top-performing models to understand and label different forms of memorization. Our analysis reveals that unintentional memorization is a serious and common issue in popular generative models. The generated images and our memorization labels of those models as well as code to compute MiFID are released to facilitate future studies on benchmarking generative models.