- Rating
- 4
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper argues that memorization in diffusion models is encoded in the geometry of the learned landscape before one-shot sampling produces copies of the training data (“latent memorization”). The authors propose to study score divergence and a basin-volume proxy to unveil these phenomena. Cyclic denoising (CD) recovers training images from checkpoints whose samples contain no copies. It also reveals “degenerate attractors” that match no training image. A simple theorem bounds residence under the exact empirical score. The approach transfers to Stable Diffusion v1.4 as a memorization detector.
Critical strengths
The question is relevant for memorization studies, and the latent-memorization picture is interesting.
The procedure and the insights gained from it are new and valuable.
The experiments are broad and consistent across datasets, architectures, optimizers and schedules.
The paper is honest about limitations, including the comparison with previous works.
Critical weaknesses
The theory part is weak but that is not a critical weakness per se. The numerical experiments is insightful and enough for the conclusions.
Questions for authors
Fig. 1C is misleading. It shows a single energy landscape, but the energy E_t is time-dependent. A CD cycle goes through the whole family of landscapes. Large γ does not necessarily mean barrier hopping; it means the training-point basins have merged at that noise level. Propose an alternative figure.
The critical amplitude is not well understood. Both the γ_c of the basin-volume sweep and the amplitude where residence disappears are treated phenomenologically. What fixes its scale and what does it mean in terms of the scenario put forward in this work? How does it evolve during training?
3 comments
I thank the authors for the reply. I find the revised figure satisfactory. I keep my good score.
Thank you!
Dear Reviewer eS5C,
Thank you so much for your fruitful suggestions and support for the manuscript.
Thank you, Authors.
Response to Reviewer eS5c(Part 1/1)
We address each point below, using W for weaknesses and Q for questions.
We propose the updated Fig. 1C in the revised manuscript, using a 3-dimensional schematic with γ on an explicit axis to remove the misleading ambiguity in the previous illustration. The additional axis makes the noise dependent landscapes explicit.
We agree with the reviewer’s interpretation that larger γ does not necessarily imply barrier crossing. We retain the paths in the figure to illustrate possible residence and transitions between basins. They are not intended to imply that a larger cycling amplitude guarantees a transition.
We agree that these amplitudes are characterized phenomenologically in our experiments. They mark the measured limits of recovery or residence, rather than theoretically established transition points. Their scale depends on how the noise schedule converts amplitude into perturbation strength and how strongly the learned reverse dynamics return perturbed states to a basin (Section 2.2, lines 212–224). Theorem 1 relates recovery and residence to noise strength relative to training-sample separation under exact empirical reversal. The new inexact-score theorem in Appendix K (Theorem 2, lines 2559–2569) instead relates residence to the margin between the neighbourhood from which the reverse pass returns and the smaller region containing the cycle endpoints. The measured thresholds also depend on the recovery tolerance and observation horizon (Section 2.1, lines 204–209; Appendix H, lines 2214–2225). The basin-volume sweep tests single-pass recovery, whereas cycling tests repeated persistence, so their thresholds need not coincide.
During training, Figure 2C shows increasing recovery-volume proxies for training samples, while held-out samples separate. Section 3.3 and Figure 3 show how residence depends on checkpoint and cycling amplitude. Together with the higher-amplitude observations in Section 3.3 (lines 423–426), these results indicate that training-associated basins become more robust and can sustain residence at higher amplitudes as training proceeds and memorization increases. A higher recovery threshold means that a sample can withstand stronger noising and still be reconstructed; residence at higher amplitudes means that repeated perturbations still return the trajectory to the same basin. The theoretical results give sufficient conditions for these behaviours; predicting how the measured thresholds evolve during training remains open.