Memorisation, convergence and generalisation in generative models

Antoine Maillard, Sebastian GoldtarXiv 2026 · arXiv

TL;DR

This paper reveals that generalization in generative models is not a single, monolithic phenomenon, but a multi-stage process. By analyzing linear generative models, the authors show that while independently trained models can converge to produce identical outputs at a certain data scale, this convergence only means they have matched the "bulk" of the data distribution—it does not guarantee they have captured the underlying principal latent factors, which are instead recovered via a separate, sharp phase transition.

What Problem Does This Paper Address?

When generative neural networks (like diffusion models) create highly realistic images, are they truly learning the underlying data distribution, or are they just cleverly memorizing their training sets?

A recent landmark paper by Kadkhodaie et al. (KGSM, ICLR '24) tried to answer this by training two separate diffusion models on completely disjoint halves of a dataset. Surprisingly, when both models were fed the exact same random noise vector, they generated nearly identical images. KGSM argued that this "convergence" of independent models signifies a transition from memorization to true generalization.

However, this finding left two major questions unanswered:

  1. How much data do you actually need to trigger this convergence?
  2. What does convergence actually tell us about how well the model has learned the true data distribution?

This paper tackles these questions by providing an exact mathematical characterization of the transition from memorization to convergence and latent structure recovery.

Methods

To make the problem mathematically tractable, the authors analyze the simplest possible setting that still captures these transitions: a linear generative model trained on data drawn from a multivariate Gaussian distribution. They model the data's underlying low-dimensional structure using a spiked Wishart model from random matrix theory, where the data covariance matrix consists of a standard "bulk" plus a few high-variance "spikes" representing the principal latent factors (e.g., global features like illumination or pose in images).

The authors track three key mathematical metrics (order parameters) as a function of sample complexity ($\gamma = n/d$, where $n$ is the number of samples and $d$ is the dimension):

  1. Memorisation ($m$): The maximum overlap between a generated sample and the training samples.
  2. Convergence ($q$): The overlap between samples generated by two independently trained models given the same latent noise vector.
  3. Latent recovery ($Q$): How well the principal subspaces (top eigenvectors) learned by the two independent models align with each other.

Crucially, they also map these parameters to formal statistical distances: they prove that convergence ($q$) controls the Kullback-Leibler (KL) divergence, while latent recovery ($Q$) governs the maximum-sliced (MS) distance between the true and learned distributions.

To bridge the gap between their linear theory and real-world data, they introduce a power-law input model to account for the fact that natural images have a power-law covariance spectrum, and they propose a novel rotated subspace overlap ($Q^*$) evaluated in the Fourier basis.

Datasets

The authors validate their theoretical insights using both synthetic data and real-world image datasets:

  • CelebA: Celebrity face images, downsampled to patch widths of 32, 48, and 64 pixels.
  • ImageNet: Random square patches extracted from ImageNet images, also at widths of 32, 48, and 64 pixels.

Key Insights and Ideas

  • Convergence $\neq$ Learning Latents: The most striking insight is that output convergence ($q$) is entirely blind to whether the model has learned the most important latent directions of the data. A model can easily "fit the bulk" of a distribution (achieving high convergence and low KL divergence) while completely missing its core low-dimensional structural features.
  • Two Distinct Objectives: Generalization in generative modeling actually splits into two independent goals: matching the bulk statistics of the data distribution and recovering the principal latent factors.
  • The Power-Law Illusion: In real images, a naive measurement of subspace recovery ($Q$) will look deceptively high even with very little data. This happens because independent models quickly agree on the massive low-frequency modes dictated by the power-law spectrum of natural images. To see if the model is actually learning the latent correlation structure, one must factor out this effect by looking at the correlation matrix in a rotated space, like the Fourier domain ($Q^*$).

Key Results

The authors prove that as the number of training samples $n$ grows relative to the input dimension $d$, the model passes through distinct regimes:

Figure 1: Memorisation, convergence, and latent recovery in generative models. Left: We illus- trate three regimes in generative models by sketching the principal components (top) and the covariance spectra (bottom) of the true data distribution (blue) and a distribution learnt from n samples (red) in d dimensions. For very small data sets, n = Θ(1), the model does not learn a smooth approximation of the data distribution, but memorises the training set. As the number of training samples becomes comparable to the input dimension, n ≍d, the spectra continuously converge and independently trained models produce similar outputs given the same latents, as described by Kadkhodaie et al. [1]. However, convergence of outputs does not imply recovery of the principal latent, which corresponds to the outlier in the true spectrum. The principal latent is learnt independently from the rest of the spectrum in a sharp transition. Right: We develop this picture mathematically by developing an asymptotically exact theory that describes the evolution of three order parameters for memorisation (m), convergence (q), and latent recovery (Q) in linear generative models. The two lines for Q correspond to high and low signal-to-noise ratio β of the principal latent (dark and light green, respectively).
  • The Memorisation Regime ($n = \Theta(1)$): When the dataset is very small, the model perfectly memorizes individual training examples ($m \approx 1$). As $n$ increases, memorization decays at a rate of $\approx \sqrt{2\log n / n}$.
  • The Convergence Regime ($n \asymp d$): As the number of samples becomes linear in the input dimension, memorization vanishes, and the convergence overlap $q$ continuously emerges and grows from 0 to 1.
  • The Latent Recovery Transition: Unlike convergence, recovering the principal latent factors ($Q$) does not happen continuously. It occurs at a sharp, separate phase transition known as the BBP (Baik-Ben Arous-Péché) transition. Depending on the signal-to-noise ratio ($\beta$) of the latent factor, this can happen either before or well after the outputs of the models have converged.
  • Diffusion Model Verification: When testing these metrics on actual deep U-Net diffusion models trained on CelebA, the authors found that while naive metrics rose smoothly, their newly proposed rotated subspace overlap ($Q^*$) successfully exposed a sharp phase transition in latent factor learning, validating their linear theory on deep architectures.

Limitations and Future Work

While powerful, the theoretical proofs in this paper rely heavily on the Gaussian approximation for both the data distribution and the generative model's behavior. Real-world data is highly non-Gaussian and contains complex, higher-order correlations that linear models cannot capture.

Additionally, this work looks at a static estimator of the model. The authors note that exploring the dynamics of how convergence and latent recovery unfold step-by-step during the iterative sampling process of a diffusion model is a promising open question. Finally, extending these concepts beyond images to auto-regressive models like transformers for text generation remains an important future milestone.

Summary updated

Abstract

Generative neural networks learn how to produce highly realistic images from a large, but finite number of examples - or do they simply memorise their training set? To settle this question, Kadkhodaie, Guth, Simoncelli and Mallat (ICLR '24) trained diffusion models independently on disjoint subsets of a dataset and showed that they converge to nearly the same density when the number of training images is large enough. This result raises two basic questions: how much data do you need for convergence, and what does convergence capture about learning the data distribution? Here, we address these questions by providing an exact analytical characterisation of the transition from memorisation to generalisation in linear generative models. We find that these models memorise at small load, while convergence emerges continuously when the number of samples is linear in the input dimension. Strikingly, we find that convergence is insensitive to recovery of the principal latent factors of the data, which are recovered in a sharp transition. After extending our approach to data with power-law spectra, we find the same distinction between convergence and latent recovery in our experiments with convolutional denoisers and in the data of Kadkhodaie et al. We thus show that generalisation in generative models decomposes into at least two distinct objectives: matching the bulk of the data distribution and recovering the principal latent factors. These objectives correspond to two different distances between true and learnt data distribution, and only the first one is captured by convergence.
BibTeX
@misc{maillard2026memorisation,
  author = {Antoine Maillard and Sebastian Goldt},
  title = {{Memorisation, convergence and generalisation in generative models}},
  year = {2026},
  eprint = {2605.21402},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.21402},
}

Loading related content...

Figure