Diffusion Models Generalize but Not in the Way You Might Think

• URL: http://arxiv.org/abs/2603.13419v1
• Date: Thu, 12 Mar 2026 21:02:17 GMT
• Title: Diffusion Models Generalize but Not in the Way You Might Think
• Abstract summary: We show that although sufficiently large denoiser models show increasing memorization of the training set with increasing training time, the resulting denoising trajectories do not follow this trend.<n>We make use of a 2D toy diffusion model to show that overfitting at intermediate noise levels is largely determined by model error and the density of the data support.
• Score: 0.42970700836450487
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Standard evaluation metrics suggest that Denoising Diffusion Models based on U-Net or Transformer architectures generalize well in practice. However, as it can be shown that an optimal Diffusion Model fully memorizes the training data, the model error determines generalization. Here, we show that although sufficiently large denoiser models show increasing memorization of the training set with increasing training time, the resulting denoising trajectories do not follow this trend. Our experiments indicate that the reason for this observation is rooted in the fact that overfitting occurs at intermediate noise levels, but the distribution of noisy training data at these noise levels has little overlap with denoising trajectories during inference. To gain more insight, we make use of a 2D toy diffusion model to show that overfitting at intermediate noise levels is largely determined by model error and the density of the data support. While the optimal denoising flow field localizes sharply around training samples, sufficient model error or dense support on the data manifold suppresses exact recall, yielding a smooth, generalizing flow field. To further support our results, we investigate how several factors, such as training time, model size, dataset size, condition granularity, and diffusion guidance, influence generalization behavior.

Related papers

• Transformers Learn the Optimal DDPM Denoiser for Multi-Token GMMs [13.741630476895773]
  This paper provides the first convergence analysis for training transformer-based diffusion models.<n>We consider the population Denoising Diffusion Probabilistic Model (DDPM) objective for denoising data.<n>A deeper investigation reveals that trained a mean denoising module enables the trained model to approximate the oracle Minimum Mean Squared Error (MMSE) of the injected noise in the diffusion steps.
  arXiv  Detail & Related papers  (2026-04-11T07:46:15Z)
• Is Your Conditional Diffusion Model Actually Denoising? [35.496429577832416]
  We study the inductive biases of diffusion models with a conditioning-variable.<n>We demonstrate that the deviation from an idealized denoising process occurs irrespective of the model capacity or amount of training data.
  arXiv  Detail & Related papers  (2025-12-21T13:54:27Z)
• Diffusion models under low-noise regime [3.729242965449096]
  We show that diffusion models are effective denoisers when the corruption level is small.<n>We quantify how training set size, data geometry, and model objective choice shape denoising trajectories.<n>This work starts to address gaps in our understanding of generative model reliability in practical applications.
  arXiv  Detail & Related papers  (2025-06-09T15:07:16Z)
• Consistent World Models via Foresight Diffusion [56.45012929930605]
  We argue that a key bottleneck in learning consistent diffusion-based world models lies in the suboptimal predictive ability.<n>We propose Foresight Diffusion (ForeDiff), a diffusion-based world modeling framework that enhances consistency by decoupling condition understanding from target denoising.
  arXiv  Detail & Related papers  (2025-05-22T10:01:59Z)
• Understanding Generalizability of Diffusion Models Requires Rethinking the Hidden Gaussian Structure [8.320632531909682]
  We study the generalizability of diffusion models by looking into the hidden properties of the learned score functions.<n>As diffusion models transition from memorization to generalization, their corresponding nonlinear diffusion denoisers exhibit increasing linearity.
  arXiv  Detail & Related papers  (2024-10-31T15:57:04Z)
• Impact of Noisy Supervision in Foundation Model Learning [91.56591923244943]
  This paper is the first work to comprehensively understand and analyze the nature of noise in pre-training datasets.<n>We propose a tuning method (NMTune) to affine the feature space to mitigate the malignant effect of noise and improve generalization.
  arXiv  Detail & Related papers  (2024-03-11T16:22:41Z)
• Data Attribution for Diffusion Models: Timestep-induced Bias in Influence Estimation [53.27596811146316]
  Diffusion models operate over a sequence of timesteps instead of instantaneous input-output relationships in previous contexts. We present Diffusion-TracIn that incorporates this temporal dynamics and observe that samples' loss gradient norms are highly dependent on timestep. We introduce Diffusion-ReTrac as a re-normalized adaptation that enables the retrieval of training samples more targeted to the test sample of interest.
  arXiv  Detail & Related papers  (2024-01-17T07:58:18Z)
• Adaptive Training Meets Progressive Scaling: Elevating Efficiency in Diffusion Models [52.1809084559048]
  We propose a novel two-stage divide-and-conquer training strategy termed TDC Training.<n>It groups timesteps based on task similarity and difficulty, assigning highly customized denoising models to each group, thereby enhancing the performance of diffusion models.<n>While two-stage training avoids the need to train each model separately, the total training cost is even lower than training a single unified denoising model.
  arXiv  Detail & Related papers  (2023-12-20T03:32:58Z)
• Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks [91.15120211190519]
  This paper aims to understand the nature of noise in pre-training datasets and to mitigate its impact on downstream tasks. We propose a light-weight black-box tuning method (NMTune) to affine the feature space to mitigate the malignant effect of noise.
  arXiv  Detail & Related papers  (2023-09-29T06:18:15Z)
• How Much is Enough? A Study on Diffusion Times in Score-based Generative Models [76.76860707897413]
  Current best practice advocates for a large T to ensure that the forward dynamics brings the diffusion sufficiently close to a known and simple noise distribution. We show how an auxiliary model can be used to bridge the gap between the ideal and the simulated forward dynamics, followed by a standard reverse diffusion process.
  arXiv  Detail & Related papers  (2022-06-10T15:09:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.