analiza aiqbench
This study establishes a theoretical link between generalization and balanced representation spaces in diffusion models, distinguishing them from memorization which manifests as localized spiky representations. The authors validate these findings across real-world unconditional and text-to-image models, offering practical tools for detecting memorization and performing training-free editing via representation steering. Ultimately, the work underscores that achieving meaningful generative modeling hinges on learning robust representations rather than merely optimizing the training objective.
Generalization of Diffusion Models Arises with a Balanced Representation Space
arXiv:2512.20963v3 Announce Type: replace Abstract: Diffusion models excel at generating high-quality, diverse samples, yet they risk memorizing training data when overfit to the training objective. We analyze the distinctions between memorization and generalization in diffusion models through the lens of representation learning. By investigating a two-layer ReLU denoising autoencoder (DAE), we prove that (i) memorization corresponds to the model storing raw training samples in the learned weights for encoding and decoding, yielding localized spiky representations, whereas (ii) generalization arises when the model captures local data statistics, producing balanced representations. Furthermore, we validate these theoretical findings on real-world unconditional and text-to-image diffusion models, demonstrating that the same representation structures emerge in deep generative models with significant practical implications. Building on these insights, we propose a representation-based method for detecting memorization and a training-free editing technique that allows precise control via representation steering. Together, our results highlight that learning good representations is central to novel and meaningful generative modeling.
Abstract:Diffusion models excel at generating high-quality, diverse samples, yet they risk memorizing training data when overfit to the training objective. We analyze the distinctions between memorization and generalization in diffusion models through the lens of representation learning. By investigating a two-layer ReLU denoising autoencoder (DAE), we prove that (i) memorization corresponds to the model storing raw training samples in the learned weights for encoding and decoding, yielding localized spiky representations, whereas (ii) generalization arises when the model captures local data statistics, producing balanced representations. Furthermore, we validate these theoretical findings on real-world unconditional and text-to-image diffusion models, demonstrating that the same representation structures emerge in deep generative models with significant practical implications. Building on these insights, we propose a representation-based method for detecting memorization and a training-free editing technique that allows precise control via representation steering. Together, our results highlight that learning good representations is central to novel and meaningful generative modeling.
Accepted at ICLR 2026. 40 pages, 19 figures. The first two authors contributed equally
Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
arXiv:2512.20963 [cs.LG]
(or arXiv:2512.20963v3 [cs.LG] for this version)
Text integral sub licența deschisă a publicației · 183 cuvinte
Subiecte conexe
Deconfusing Corrigibility
AI Alignment Forum · 17 oct. 2026
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Latent Space · 13 oct. 2026
[AINews] The Future of Latent Space
Latent Space · 25 sept. 2026
Does DiffusionGemma do latent reasoning?
AI Alignment Forum · 16 aug. 2026