Abstract
Generative Adversarial Networks (GANs) are often assumed to create novel images, but this research reveals they predictably "replicate" (memorize) training data when dataset sizes are small. We quantify this behavior, showing that replication probability decays exponentially as dataset size and complexity increase. Furthermore, we identify a "U-shaped" trend in image quality: quality is high when the model perfectly memorizes small datasets, drops during the transition to generalization, and rises again as the dataset becomes sufficiently large. We propose a mathematical model to predict replication curves, providing a framework for determining the data requirements necessary for true generative performance.
Introduction
In sensitive domains such as medical imaging or privacy-preserving face swapping, GANs are frequently employed to augment limited datasets. The utility of these models rests on the assumption that they learn the underlying data distribution to produce unique samples. However, if a GAN merely outputs exact copies of training images, it fails as an augmentation tool and introduces significant legal and ethical risks regarding copyright and privacy. This study moves beyond anecdotal evidence to establish a mathematical relationship between dataset size, complexity, and the propensity for replication.
Methodology
The study evaluates two prominent architectures, BigGAN and StyleGAN2, across three standard benchmarks: CelebA (low complexity), Oxford Flower 17 (medium complexity), and LSUN-bedroom (high complexity).
To quantify the "complexity" of these datasets, the authors utilize Intrinsic Dimensionality (ID), estimating the true number of variables required to describe the data manifold. Replication is defined using a strict Euclidean distance threshold ($\alpha$) in pixel space; a generated image is considered a replication if it falls within this distance of any training sample.
Analysis of Replication
A primary finding of this research is that replication percentage is not a binary state but follows a predictable exponential decay. As dataset size increases or the data manifold becomes more complex, the probability of the GAN "copying" training data decreases.
The authors demonstrate that the decay factor remains remarkably consistent across different architectures. By modeling replication as a function of both image count and ID, they developed a "one-shot" estimation method. This allows researchers to predict the entire replication curve for a new dataset using only a small initial subset, forecasting the data volume required to transition from memorization to generation.
The Quality Paradox
Perhaps the most counter-intuitive finding is the "U-shaped" quality curve. Perceptual image quality does not improve linearly with data. Instead, it starts high when datasets are tiny because the GAN produces perfect, memorized copies of the training data. As the dataset grows, quality drops—the model can no longer memorize the entire set but lacks sufficient data to generalize effectively. Quality only begins to rise again once the dataset is large enough for the GAN to learn the underlying manifold.
Behavioral experiments conducted via Amazon Mechanical Turk confirmed this trend, highlighting that "poor" GAN performance often occurs in the middle ground between memorization and true generalization.
Conclusion
This work proves that GAN replication is a mathematically predictable phenomenon. The proposed exponential model achieved high goodness-of-fit ($R^2 > 0.9$), and the one-shot prediction tool demonstrated a median absolute error of only 2.42%. While the use of Euclidean distance focuses on pixel-level memorization rather than semantic copying, these results provide a critical baseline for understanding the data requirements of generative models. Future work may extend these findings to newer architectures, such as Diffusion Models, to determine if they adhere to similar scaling laws.
Summary updated