Early Signatures of Memorization in Diffusion Models via Basin Geometry and Cyclic Denoising

Key points

• Diffusion models generalize early in training and later reproduce individual training samples.1
• Standard tests detect memorization only once one-shot generation produces near-copies, leaving a released model unaudited until its outputs fail.1
• Using score divergence and basin volume, we find that localized basins form around training samples and separate them from held-out samples before the first memorized sample appears, with an onset that follows the same $O(n)$ scaling as the memorization time.1
• These findings hold on a Gaussian mixture, CelebA, and CIFAR-10 across optimizers, architectures, noise schedules, and training-set sizes, and extend to off-the-shelf Stable Diffusion v1.4, where the cycled conditional-unconditional divergence gap separates memorized from non-memorized prompts with an AUC of 0.944 and a TPR of 0.866 at 1% FPR.1

Sources (1)

• [1]Early Signatures of Memorization in Diffusion Models via Basin Geometry and Cyclic Denoising
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 10:47 AM

  We show that memorization is encoded in the geometry of the learned energy landscape before it appears in generated samples, a state we call latent memorization.

  Diffusion models generalize early in training and later reproduce individual training samples.

Extractive summary: sentences quoted from the sources.