mlsanityData to 14 Oct 2026

Memorization and Malign Generalization in Conditional Diffusion Models with Random Features

arXiv 2610.11288 ↗First posted 8 Oct 2026cs.LGcited by 0

On the map: Diffusion models → Diffusion foundations and sampling

Abstract (arXiv)

Conditional diffusion models generate diverse, novel, and high-quality samples under prescribed conditions. However, theoretical understanding of their memorization and generalization remains limited, while recent works have characterized these behaviors primarily in unconditional settings. In this work, we analyze a random-feature conditional score model in the high-dimensional proportional limit, deriving asymptotic expressions for training and test losses. By decomposing the test loss, we show that in the overparameterized regime, increasing model width improves prediction of the condition-dependent mean while reducing within-condition prediction variance, a phenomenon we term "malign generalization." Furthermore, analyzing the training loss reveals that more informative conditions lead to memorization of training samples at smaller widths. These theoretical findings are supported by experiments with U-Net architectures on realistic data.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Diffusion foundations and sampling, one of 14 directions in Diffusion models. Builds on 14 of its 66 references that are in the map. First posted on arXiv 8 Oct 2026.

Builds on

14 of 66 references.

Closest ancestors · 11

Same direction, newest first

  1. 2026Double Descent and Malign Overfitting in Diffusion Models22 Sep
  2. 2026Generalization, memorization, and overfitting for diffusion models trained in the lazy high-dimensional regime25 Aug
  3. 2026Benign Overfitting Does Not Occur in Diffusion Models2 Jul
  4. 2026Asymptotic Learning Curves for Diffusion Models with Random Features Score and Manifold Data24 Mar
  5. 2025Generalization Dynamics of Linear Diffusion Models30 May
  6. 2025Taking a Big Step: Large Learning Rates in Denoising Score Matching Prevent Memorization5 Feb
  7. 2024Convergence of Diffusion Models Under the Manifold Hypothesis in High-Dimensions27 Sep
  8. 2024Unveil Conditional Diffusion Models with Classifier-free Guidance: A Sharp Statistical Theory18 Mar
  9. 2023The Emergence of Reproducibility and Generalizability in Diffusion Models8 Oct
  10. 2022Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions22 Sep
  11. 2022Classifier-Free Diffusion Guidance26 Jul

Other references · 3

  1. 2023A Survey on Audio Diffusion Models: Text To Speech Synthesis and Enhancement in Generative AIcited by 68
  2. 2023A Reproducible Extraction of Training Images from Diffusion Modelscited by 61
  3. 2026ELF: Embedded Language Flowscited by 45

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Denoising Score Matching with Random Features: Insights on Diffusion Models from Precise Learning Curves
  2. 2026On the Memorization of Consistency Distillation for Diffusion Models
  3. 2025Provable Separations between Memorization and Generalization in Diffusion Models
  4. 2023On Memorization in Diffusion Models
  5. 2024Detecting, Explaining, and Mitigating Memorization in Diffusion Models
  6. 2026Local Coverage Governs Memorization in Diffusion Models
  7. 2025On the Edge of Memorization in Diffusion Models
  8. 2026Understanding diffusion models requires rethinking (again) generalization