acceptodds
Under review as a conference paper at ICLR 2027

From Gist to Verbatim: the Controllable Path of One Image Through a Diffusion Prior

Abstract

Modern generative models are typically studied for their ability to generalize, while memorization is treated as a failure mode. We invert this perspective and ask how a generative model can recall a single datum. We frame memorization as a control problem: a learned conditioning signal steers a frozen diffusion model toward a target image, and we follow the formation of this memory throughout training. Before converging to exact reproduction (*verbatim*), memory passes through a regime of *abstraction*: samples preserve the target's semantic content (*gist*) while remaining diverse in appearance, leaving predictable detail to the generative prior. This semantic organization emerges despite the absence of labels, captions, or any explicit semantic objective. In the abstraction regime, the learned memory approaches ordinary text conditioning and supports editing, one-shot recognition, and retrieval more effectively than in the near-perfect reconstruction regime. These results suggest that memorization can serve as a probe of the relevance structure encoded by a generative model: what must be stored about a datum, what can be reconstructed from prior knowledge, and how learned memories interpolate between reconstructive recall and literal data storage.

Primary area:
generative models
Keywords:
Diffusion Models, Reconstructive Memory, Memorization, Control Theory, Generative Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1hB7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper centers around memorization in diffusion models, but rather than the traditional lens of viewing it w.r.t its notion and ability for extraction of training data, the paper focuses on a different setting. This setting examines generated images from a model that evolves, via fine-tuning of the text condition, throughout this process. The paper discusses the different stages it defines during that process, where the memory of the model passes through the ‘abstraction’ regime, until reaching the ‘locking’ phase which generates the input image verbatim.

Critical strengths
  • The overall concept of the paper is easy to understand and follow.
  • The idea of changing the paradigm of memorization from observation at the static model to some process of optimization on its condition is novel and interesting.
  • The notion of semantic convergence before the locking stage is demonstrated effectively as can be seen in Fig. 1b.
Critical weaknesses
  • The authors discuss memorization and claim to reframe it differently. This new framing is not well justified. It is done using an optimized condition and examines the output along this process, but it does not establish how this reflects memorization, rather than a characteristic of the condition optimization. While the paper demonstrates the results of abstractions along this process, I believe this is a known phenomenon, both from the fact that gradient descent fits low frequency and high energy first [1], along with methods that showed that for personalization methods this transition happens explicitly [2].
  • The process itself (Eq. 2) resembles the personalization methods as mentioned in the paper (e.g. Textual Inversion [3]), and it is not grounded why this process with certain 'memories' is better than performing the same with such an approach.
  • The structure of the paper is not well written, I think the lack of related works in the main paper affects readability, along some preliminary background on diffusion model before the definitions of the training objective (Eq.4).
  • Mathematical framework: section 4 provides a mathematical model that is supposed to ground the idea of the abstraction regime emergence. However, these dynamics leave us with the residual and susceptibility (L216-221), and both are general and have an extra degree of freedom which does not help to establish the notion of the abstraction regime. The paper doesn’t specify ϕ\phi or measure the covariance to support this. So, for example, even if hypothetically the abstraction regime would not exist, we would still have this separation of features with larger and smaller eigenvalues, and this mathematical setting can be used to explain the exact opposite ordering (surface before semantics).

[1] On the Spectral Bias of Neural Networks, Rahaman Et. Al., ICML 2019

[2] Is This Loss Informative? Faster Text-to-Image Customization by Tracking Objective Dynamics, Voronov Et. Al., Neurips 2023

[3] An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion, Gal Et. Al., ICLR 2023

Questions for authors
  • I think the core claim of the paper is confusing, as formulating memorizing in a different way is important, but the actual current formulation lacks a lot of information that would be important for such a formulation. For a memorization analysis, I would expect an analysis that can differentiate different populations: what happens when the analysis is done on something that is actually out of distribution (not just images that model did not generate as the authors name it, but actual images the model finds hard to produce)? Within the generated images, are there different behaviors according to different prompts? What about training data vs. non training data (the “traditional” memorization)?
  • As this sits closely to the personalization methods, can this analysis help improve personalization methods, e.g. when the desired concept is more/less abstract?
  • Regarding W2: What is the difference of using the learnable vectors rather than using known methods like textual inversion? Does the regime of abstraction emerge also in this scenario? If this is the case, what is the contribution of this method using the learnable vectors?
4 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

We thank the reviewer for their interesting comments. We are glad they found our work easy to read and understand, and the memorization paradigm novel and interesting. The result of a semantic-first ordering is central in our manuscript, and we are happy that the reviewer found it demonstrated in an effective way.

Before addressing all the points, we would like to stress an important dichotomy. Standard memorization in ML: a failure mode where training data are internalized and reproduced by the generative model. Our connotation of memorization: the process of learning a conditioning vector (the memory) that forces a well-trained generative model to reproduce a single new datum (e.g., think of a human adult memorizing a poem). In particular, we pay specific attention to the dynamics of this memorization process, and we show how it unveils the internal relevance structure of the generative prior. Our framework is not meant to be informative about the standard memorisation of pre-training data.

Weaknesses

Relation to previous work and novelty

Gradient descent bias for low frequency and high energy [1]: The “semantic” directions preferred during abstraction do not correspond to low-frequency/high-energy details in pixel space: Fig. 1 explicitly shows that fine, high frequency details like the eyes’ shape are fixed faster than the person’s position, their silhouette and background, which are of arguably lower frequency and higher contrast. Abstraction emerges as a byproduct of an optimization process that explicitly requires faithful reproduction of the target image.

Personalization methods [2]: We thank the reviewer for pointing out missing references (added to the revised paper). We agree with the reviewer, multiple personalization methods are based on conditioning, similarly to what is discussed in this paper. Some hints of the phenomenology that we center our work around are indeed present. The main difference lies in the scientific goal: we are not presenting a way of improving personalization in diffusion models or making it more efficient. Our framework builds on similar methods, but is aimed at probing the hidden relevance structure of the prior, how it emerges dynamically during memorization of a single image, and how different stages of this memorization process are compatible with function, i.e., how useful the memory can be beyond the reproduction task. The question is thus more general, and extends beyond SD2.1 and diffusion models, and even to different data modalities. While the late stage of training in [2] shows an undesired degradation of personalization performance, in our framework verbatim reconstruction is the only explicit training goal. Yet, the dynamics naturally pass through abstraction, and we shift the perspective to noticing how this is non-trivial and may tell us something about human memory. Moreover, this research direction is closely aligned with current results in computational neuroscience, focused on the hippocampus-neocortex interaction and on episodic memory (see Extended Related Works).

Improving clarity

We agree with the reviewer that some background knowledge is not fully covered in the current version; we will use the additional page to add more references in the introduction and to better introduce the diffusion framework.

Mathematical framework

We thank the reviewer for pointing to this limitation, already partially acknowledged in the current Limitation (iii). We agree we are not able to map back the ingredients of our framework to the experimental setup with SD2.1, given that we have no direct access to the internal features of this model. However, the simplicity of the framework allows the identification of a clear condition for the emergence of semantic-first abstraction: high susceptibility of the corresponding semantic modes. This condition is not necessarily realised in all models. In fact, it is completely possible that a different model, trained on different data and with a very different prior, will not exhibit the same semantic abstraction. For example, if one considers a model that was never trained on human faces, it should be impossible for it to enter an “identity”-preserving abstraction phase, since it lacks the necessary feature modes (see also reply to Q1). However, in our experiments, we do see a semantic-first phase with SD2.1-compatible data, and we also find a correspondence with two qualitative predictions obtained from the mathematical framework (see Sec. 5).

AuthorsOpenReview ↗

Rebuttal (2/2)

Answers to questions

Q1: We partially address this question in the opening remark of our rebuttal (1/2). On the experiments on strongly out-of-distribution targets: we are running new experiments with purely geometrical targets, pixelated targets, and blurry real images, instances which we don’t expect SD2.1 to be good at reproducing. It needs to be noted that the SD2.1 VAE can become a bottleneck in very out-of-distribution images, since its encoding/decoding may become lossy. Our preliminary results show that the abstraction window can shrink and even disappear if the internal dictionary of features of SD2.1 does not cover the necessary features for quickly acquiring the target. This gives further confirmation that abstraction needs the target to excite highly susceptible, semantic channels.

Q2: We thank the reviewer for the interesting suggestion. Except for some preliminary results (see also response to reviewer 668R), we have not yet devoted our study to finding direct applications of this analysis. It is plausible that the learned conditioning vector can be manipulated to deploy the wanted concept/subject in different settings, similarly to what is done in Textual Inversion. Then, the shift of the conditioning’s nature from semantic to positional can explain some of the performance deterioration in late training, as observed in [2].

Q3: We chose our conditioning scheme to be as general as possible: our setup could be extended to any model that can be conditioned through attention, via direct learning in KV space (e.g. in LLMs). In fact, we tested multiple conditioning methods: training in token space (as in Textual Inversion), training part of the empty caption conditioning, training a LoRA for the weights of the network (Fig.7 in E.3). Each one of those variations still produces the abstraction phase, which is the real object of our study.

[1] On the Spectral Bias of Neural Networks, Rahaman Et. Al., ICML 2019

[2] Is This Loss Informative? Faster Text-to-Image Customization by Tracking Objective Dynamics, Voronov Et. Al., Neurips 2023

[3] An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion, Gal Et. Al., ICLR 2023


We hope these clarifications address the reviewer's questions, and we would be glad to provide any further detail during the discussion.

Thank you for your elaborated response.

Your clarification regarding your connection to memorization is important, and I think it helped me understand that in the abstract, the text should say so explicitly. Inverting the ‘traditional’ memorization as a failure mode and asking how a model recalls a single data point plants the idea and expectation that the reader will receive information on how that specific point is memorized. However, this is not the focus of the paper, and I think this is very important for the clarity of the paper’s main idea.

Regarding the gradient descent bias, the low frequency is not in the pixel space. It is in the dominant directions of the loss being optimized, and if the authors would analyze the low frequency of the cost I believe they will experience the same phenomena. I still believe that this abstraction for fine-tuning a specific image is a result that alone, without any valid takeaways from it, is not surprising. I understand the scientific goal, but I believe its execution needs further analysis on how this can give us new information, whether for specific data memorization, “hardness” of wiping data out as done in concept erasure, or other surprising usages.

Regarding the mathematical framework, I still don’t see how the framework explains the abstraction regime emergence. Any ordering is possible depending on the prior, so the framework does not predict semantic-first behavior; this is just observed in SD2.1. The presented theory does not distinguish the high variance features, and the fact that this aligns with SD2.1 does not mean the framework is true. It needs further analysis to support some separation.

Overall, I think the idea of the paper is interesting, but it needs additional supporting analysis, more than what a revision should contain. Therefore, I’m keeping my score.

We thank the reviewer for their continued engagement. However, we believe that some of the remaining criticisms rest on an interpretation of our contribution that differs from what we actually claim.

On memorization, our goal is precisely to study how a single new datum is acquired, following which properties become reproducible along the optimization of its learned conditioning signal. We agree that the first two sentences in the abstract lead to a possible misunderstanding; we will change the opening to: "Modern generative models are typically studied for their ability to generalize, while memorization is treated as a failure mode. We invert this perspective and ask how a generative model can be made to recall a single new datum."

On spectral bias, we agree that preferential learning directions are a generic feature of gradient-based optimization, shaped by the loss. In our setting, since there is a single datum, the central question is how the structure already encoded by the prior determines which properties are acquired first. This internal structure is what shapes the fast learning directions and creates a correspondence with semantic modes. Generic optimization bias alone does not account for the additional generative and functional signatures we observe.

On the theory, we explicitly present an interpretative framework, not a derivation of SD2.1's internal dynamics. The theoretical goal is a conditional explanation: identifying properties of a prior under which semantic-first learning can emerge, and deriving consequences that can be investigated experimentally. The fact that different priors can produce different orderings is precisely part of our perspective, rather than a contradiction of it. We acknowledge that direct verification of the proposed susceptibility mechanism remains open, and it will inspire future work.

We understand the reviewer's interest in a more complete mechanistic explanation and concrete applications, but these are extensions beyond the claims of the present paper. We believe the empirical findings and their characterization constitute a scientific contribution in their own right.

Reviewer_LYvDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how a frozen diffusion model memorises a single image through an optimised conditioning signal. It identifies an “abstraction” phase, where generations capture the target’s semantics but vary in appearance, followed by "locking", where they converge toward the image. Using SD 2.1, the authors measure this trajectory and examine its implications for generalisation, attention, editability, recognition, and retrieval, with extensions to other models and modalities. They also propose a theory based on feature residuals and covariance-driven susceptibility.

Critical strengths

This paper presents an interesting research topic.

Critical weaknesses
  1. The distinction between semantics' and surface details is not sufficiently validated. DINOv2 features and VAE latents may each encode both high-level structure and instance-specific details. Thus, the half-decay ratio of 21.4 shows different error trajectories in these representations, but does not by itself establish semantic-first learning. Please add complementary semantic, identity, perceptual, and low-level metrics; test sensitivity to feature and normalisation choices; and report uncertainty in per-target half-decay times. Clarify how the finite training endpoint affects the “total decrease'', especially for unconverged runs.

  2. The sample coverage and uncertainty analysis do not support the general claim of abstraction before locking. The main bank contains only 30 model-generated images, and the prior, optimised model, and primary evaluation representation are closely related. The external-data results remain limited, and uncertainty across optimisation restarts is unclear. Please replicate across more targets, initialisations, noise realisations, and evaluation seeds. Report within- and between-target uncertainty, and expand quantitative results on real photographs. Because Figure 2a selects seeds with the largest peak advantage, include randomly selected examples and state the selection rule.

  3. The theory does not derive the observed semantic-first ordering, and its objective differs from the experimental setup. The theoretical model describes feature dynamics under a specified distribution and objective, but the claim that semantic features have greater susceptibility is not tested against measured feature statistics or mapped to the learned conditioning vectors. In addition, the score-loss equivalence in Appendix B relies on assumptions and time weighting not used in the SD 2.1 experiments. Please clarify this gap and narrow the theoretical claims or validate the predicted ordering in a synthetic setting and against measured features in the experiments.

  4. The claimed "fidelity--functionality trade-off'' is confounded by changes in conditioning strength. Training step changes memory norms, prior-shift strength, and reconstruction error alongside the functionality probes. The reported two-order-of-magnitude norm growth is not controlled, and blurred or unrecognisable outputs may indicate failure rather than locking. Please add norm-matched or rescaled checkpoints, comparisons at matched reconstruction error, alternative training paths, and appropriate no-memory and random-memory baselines. Report output quality and limit the claim if these controls do not support generalisation.

  5. The editability experiment does not isolate whether edits are preserved. FLUX.2 edits are passed through the SD 2.1 VAE and a strongly noised U-turn procedure, so outputs may reflect model-prior effects or generation failure rather than memory overwrite. DINOv2 distances alone cannot show whether edited attributes survive. Please add no-memory and original-image controls, test multiple noise levels, and assess attributes directly or with blinded ratings. Quantify blurred or unrecognisable outputs separately from genuine reversion toward the target.

  6. The one-shot recognition experiment is too small to establish generalisable identity recognition. The study uses 10 identities and 100 edited queries, without independently validating identity preservation. Candidate denoising losses may also depend on training progress, vector norms, or target difficulty. Please expand the identities and query sources, report identity-level uncertainty and confusion matrices, and compare against empty-, random-, text-, and norm-matched memory baselines. Test score comparability and leave-one-identity-out stability.

  7. The text-memory retrieval metric lacks representational validation and controls. Equal dimensionality does not make memory vectors and text-encoder EOS vectors comparable semantic representations. Cosine similarity may reflect norm, anisotropy, or associations already present in the frozen prior. Please add random, empty-memory, label-shuffled, and within- versus between-class baselines, and evaluate retrieval with an independent ranking metric while controlling for prompt wording. If vector-space alignment cannot be validated, describe the result as a similarity correlation rather than retrieval ability.

  8. The introduction section lacks a strong motivation.

Questions for authors

Please see the weakness section.

5 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

We thank the reviewer for finding interest in our work. Below, we try to address the reviewer’s concerns point by point.

Weakness 1

Before addressing the specifics, we would like to remark that the chosen metrics and measured half-decay ratios are just a numeric quantification of a phenomenon which is clear from visual inspection. Plots analogous to those in Fig.1b-c (note that the reported curves even have a different concavity), can be obtained with different choices of metrics, and the ordering is robust.

a) We examine the VAE space for “surface detail” since this is the space where the diffusion process lives. The VAE encoder/decoder in SD is often seen as a quasi-lossless dimensionality reduction method, not displaying a particular preference for semantic features. The DINOv2 representation is considered “semantic” because it loses information about the original image: DINOv2 focuses on what objects and parts are present, while being relatively robust to changes in their appearance.

b) We have new results looking at CLIP image embeddings for the semantic level and at raw pixel level (256x256) for the surface. All four methods' embeddings (pixel, VAE, CLIP, and DINOv2), have been tested with and without unit normalisation. We report that any semantic space paired with any surface space gives the same picture: the median half-decay ratio lies between 14 and 23 across all sixteen pairings, and in every pairing the surface error is the slower of the two for all 30 targets. The same holds on the 80 real photographs of Appendix E.1 (ratio about 10, surface slower in 79 of 80 images). We test the priority of identity memorisation in the recognition experiment of Section 6, which uses no feature space at all, and shows that a memory recognises its person long before it can reproduce the picture.

c) We resampled the evaluation seeds with replacement 500 times for every target and recomputed both half-decay steps each time. For every one of the 30 targets, the 95% interval of the surface-to-semantic ratio lies entirely above 1. In 15,000 replicates, the surface error never decays first.

d) On half-decay step: in our experiments, the semantic curves converge by the end of training; the surface curves do not. A longer run would thus lower the surface curve's endpoint and its half-level, only delaying its half-decay step; it cannot move it earlier. Thus the ordering is robust with respect to chosen endpoint. We will state this explicitly in Section 3.

Weakness 2

a) We are currently working on expanding our endogenous face image bank with 30 new identities. However, we would like to point out that the paper also trains 80 single-image memories on real photographs from four classes (Appendix E.1). The ordering of semantic before surface robustly holds across targets.

b) DINOv2 was pretrained without text and independently of SD 2.1; for the rest see answer to Weakness 1.

c) Every checkpoint is already evaluated on 20 fixed noise seeds, with the target's own generating seed excluded. Figures report the between-target interquartile range. We have now added a per-target bootstrap over evaluation seeds (see also reply to Weakness 1): for every target, the 95% interval of the surface-to-semantic ratio lies above 1.

d) We retrained five face targets, from two additional initialisation seeds. The semantic-first ordering appears in each run. In the appendix of the revised manuscript, we will add a study of variability across restarts compared to variability across targets.

e) The selection rule for Fig.2a is already stated in the caption, and Appendix E.4 shows randomly selected seeds for the same targets.

AuthorsOpenReview ↗

Rebuttal (2/3)

Weakness 3

a) As explicitly stated in the main text (see Sec. 4 and limitation (iii)), the mathematical framework is introduced as an interpretative tool, not as a proof or as a mechanistic explanation of the observed phenomenology (we don’t think the setting is tractable with currently known methods). In Sec. 4 we restate the control problem entailed in the memorization process in simple mathematical language, allowing for the identification of transparent necessary conditions for the emergence of an abstraction phase. Aiming for a minimal model, we specialise the framework to a convenient time-scaling that maps the score-matching problem to a maximum-likelihood objective (as discussed in the appendix). We agree that this is not the same time-scaling setting as in SD2.1; we will make this clearer in the revision. Changing scaling in our framework would add a factor in Eq. (7), capturing a dependence on the diffusion-time window in which each feature is resolved. This per-feature rescaling creates an unnecessary complication, while contributing very little to the interpretation of the mechanism.

b) The reviewer is correct: stating conditions for the emergence of semantic abstraction does not imply they are realised in the experimental setup. Unfortunately, measuring the “feature covariance” in SD2.1 and identifying explicit susceptibility directions, as suggested by the reviewer, is not simple: the feature basis is not known, and it is not clear how to extract it from the internals of the model. The experimental evidence of the appearance of the abstraction phenomenon, and the observation of two qualitative behaviors identified through the framework (Sec 5), constitute our evidence of high-semantic susceptibility in SD2.1’s conditioning channel.

c) We have ongoing work (soon to be released) analysing a specialization of the framework of Sec. 4 to a tractable synthetic data setting, where we fully explore the described phenomenology and find a convincing parallel to the hereby presented empirical results. We believe this substantial contribution cannot find enough space in the present paper and deserves a focused separate study.

Weakness 4

Before going into the details of the answer, we would like to say we respectfully disagree with the premise of this criticism. The growing norm of the conditioning vector is not a confounding effect, but an important feature of the learning dynamics, necessary (although not sufficient) for transitioning between an initial condition where the generative model behaves like the prior and a final condition where any generative trajectory closely reconstructs the image. This growth is naturally induced by training on the memorization objective; we are not enforcing it with additional constraints.

a) We have investigated whether norm alone can explain the emergence of the two phases and, in short, the answer is that it cannot. We will expand Appendix E.2. The evidence:

  • When a single memory vector is trained (M=1), the norm grows unbounded, yet the locking phase never fully kicks in.
  • Taking a memory checkpoint in the abstraction phase and scaling it up to a norm compatible with a locking checkpoint does not result in locking. The locked state needs both the learned direction and the learned strength.
  • Similarly, scaling down a locking checkpoint to a norm similar to the abstraction phase does not result in the same abstraction behavior, although the weaker control allows for more prior-driven variation.
  • We tested the effect of keeping the norm pinned for the whole run. For example, low norm prevents the emergence of locking.

b) If the editability tests in the locking regime produce blurred images, this only corroborates the fact that the memory is losing flexibility and that denoising from a noised edit is sufficient to disrupt the reverse trajectory. This does not happen in the abstraction phase.

c) The fidelity--functionality trade-off cannot be explained away by the changing norm of the memory, since the direction of the learned vectors is crucial for the observed effects. In text-based retrieval, we measure a cosine between vectors, which is norm-insensitive. In the zero-shot experiment, we are denoising an image with same-checkpoint memories from different identities, and this means the norm is approximately the same for all of the candidate denoisers (see also norm-matched experiments mentioned below).

AuthorsOpenReview ↗

Rebuttal (3/3)

Weaknesses 5–7

We thank the reviewer for these suggestions. We expanded the controls for editability, one-shot recognition, and text–memory retrieval. For recognition, we also increased the evaluation from 10 to 30 identities, with 10 held-out pose/scene edits per identity, yielding 300 queries.

Editability. Across all target–edit pairs, we added no-memory and original-image controls, and tested four noise levels. Without memory, edit similarity remains constant across checkpoints. With learned memory, it improves early and deteriorates late. Meanwhile, reconstruction of the original target can see two possible outcomes: if the edit is sufficiently close to the original image, the locking memory successfully redirects to the original target. If the edit is instead far, the reconstruction leads to broken blurry images. The loss of edit similarity also persists under milder noising. We will remark in the revision how the D_orig allows for discriminating between these cases.

One-shot recognition. The expanded evaluation reproduces strong early recognition and its subsequent drop: accuracy decreases from approximately 0.77 to 0.12, against chance accuracy of 0.033. Empty and norm-matched random memories remain near chance, while text conditioning achieves 0.193. Equalising learned-memory norms across identities preserves the trajectory, arguing against unequal conditioning strength across candidates as its explanation. Identity-level results, confusion matrices, and leave-one-identity-out analyses further show that performance extends across the identity pool rather than depending on a few successful cases.

Text–memory retrieval. We additionally performed norm-matched random and untrained-memory baselines, shuffled-label controls, centering and standardization, ranking metrics, prompt rewording, and separate matching- versus nonmatching-identity similarities. Learned-memory margins exceed the shuffled thresholds and remain positive after standardization; candidate rankings are well above chance, and the tested wording changes leave margins unchanged.

Weakness 8

We motivate our study with the following questions:

  • How does the interaction between new information and prior shape a memory as it forms? If a strong generative prior is deliberately made to memorize one image, what does it learn along the way?
  • How is the information in a single datum ordered by relevance? What is committed to memory first, and what can initially remain implicit in the prior?

We believe answering these questions is a sensible scientific programme, and we tried to address them in the present manuscript. We will try to make our scientific goal clearer in the revised version.


We hope these clarifications address the reviewer's questions, and we would be glad to provide any further detail during the discussion.

Reviewer_668ROpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how a frozen diffusion model learns to gradually memorize a single image by optimizing a small set of conditioning vectors which are injected into the frozen model by cross-attention modules. Early in training, generated samples capture the target’s semantic content while varying in details such as pose, appearance, and background. With further training, this diversity disappears and the model approaches exact reproduction. The authors provide a simplified theoretical framework to interpret this progression. They also show that editing, recognition, and text-based retrieval work better during the intermediate abstraction stage than during near-exact recall. Experiments on real photographs and additional generative models suggest that the pattern extends beyond the main setup.

Critical strengths
  • The paper is quite interesting to read. It offers an alternative view of the memorization/generalization issue that has been extensively studied in litarature: rather than treating memorization as only a failure, the paper uses it to study what a pretrained generative model captures first and showcases an interesting observation.
  • The usage of DINOv2 and VAE latents to study semantic and surface attributes is also quite interesting.
  • The paper extends its scope to study editability, one-shot recognition, etc, makes the study more comprehensive.
Critical weaknesses
  • The reviewer's concern of the paper is mainly its practical utility. Although the paper is interesting, the reviewer found it a bit hard to think of possible implications of the paper. Can the authors comment on this?
  • The model was pretrained with text supervision, and the memory vectors use its frozen attention weights. Could this be related to the observation that semantic content is learned first? Would the same pattern hold if the memory branch’s key/value projections were also trained?
Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Rebuttal (1/1)

We thank the reviewer for their positive assessment. We are glad that the reviewer found the paper interesting to read, that the reframing of memorization as a probe of what a pretrained model captures first rather than as a failure mode came across as the central idea, and that the use of DINOv2 and VAE latents to separate semantic from surface content, together with the editing, recognition, and retrieval tests, made the study comprehensive. These are the contributions we hoped to convey. Below we answer the two questions.

Practical utility. Thank you for raising this point. The main objective of the paper is to characterize a phenomenon, the ordered formation of a memory relative to a generative prior, and to establish its functional consequences; it does not aim at proposing a specific application or technique. The findings nevertheless suggest more concrete directions, some of which we are already pursuing in ongoing work: retrieval-augmented generation over a database of learned memories; memories as compact representations of video clips, including as a conditioning signal for coherence in long video generation. Moreover, access to the relevance structure of the generative model (through abstraction) allows the exploration of novel monitorability/interpretability and unlearning methods.

Developing and benchmarking these applications is beyond the scope of this study. Our experiments establish the underlying phenomenon and its functional consequences, providing a basis for investigating these directions.

Finally, from a more theoretical perspective, we are exploring the addition of a second learning stage where the generative model itself is updated given the compressed memory, which we believe closely mimics the hippocampus-neocortex interaction during memory consolidation. We believe our numerical and analytical frameworks provide a concrete first step in understanding this important mechanism for human episodic memory.

Text supervision. We agree that text supervision during pretraining may contribute to the observed semantic organization, and we state this in Limitation (i). We claim that semantic-first learning emerges without semantic supervision during memory formation, not that it is independent of the model’s pretraining. Note that, even if one excludes the direct role of the cross-attention text conditioning, the internal representations of SD2.1 are shaped through the interaction with language during pretraining, and one cannot fully remove this bias.

On the specific question of the frozen key/value projections, we believe the LoRA experiment of Appendix E.3 (Fig. 7) provides useful evidence. There, no conditioning vector is learned at all, the text cross-attention layers are excluded from adaptation, and the image is memorized through low-rank corrections to the remaining weights of the UNet. Nothing learned passes through the frozen text-pretrained projections, yet the same two phases appear: abstraction followed by locking. The ordering is therefore not produced by routing a memory through those projections. The FLUX.2 and CogVideoX results (Sec. 7) further broaden the tested interfaces.

The variant the reviewer describes, a memory whose projections co-adapt with it, is a very interesting direction. In particular, one could explore a setup where a set of images sharing a specific semantic similarity structure are trained simultaneously with a fresh cross-attention channel. In this case, one could observe a reshaping of the relevance structure emerging during the abstraction phase. However, this protocol requires multiple non-trivial design choices and would likely also require large-scale training; thus, we cannot add it during this rebuttal phase.


We hope these clarifications address the reviewer's questions, and we would be glad to provide any further detail during the discussion.

I thank the authors for their response and apologize for this late reply.

I find the discussion regarding text supervision quite convincing, and I appreciate the authors' further discussion of the practical utility of their approach.

Overall, I find the paper interesting and appreciate its novel perspective on the problem. I therefore remain positive about the work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Revision uploaded

Dear Reviewers and Area Chair,

We have uploaded a substantially revised version of our manuscript, incorporating the clarifications and additional experiments discussed in our rebuttal responses. We thank the reviewers for their constructive feedback, which has helped us strengthen both the presentation and the empirical support for our findings.

In particular, we have:

  • Improved the presentation and theoretical framing, expanding the introduction and diffusion-model background, strengthening the discussion of related work, and clarifying the assumptions, aims, and limitations of our theoretical framework.
  • Strengthened the robustness analysis, introducing alternative semantic and surface metrics, additional training initializations, experiments on strongly out-of-distribution targets, and controls for the growth of memory norm.
  • Expanded the functional evaluation, increasing the one-shot recognition experiment to 30 identities and 300 held-out queries, and introducing additional baselines and controls for editability, recognition, and text–memory retrieval.
  • Added further qualitative evidence, including randomly selected noise realizations and additional examples across models and modalities, and clarified the broader implications and limitations of our findings.

We believe these additions substantially strengthen the evidence for our central results and address the main concerns raised during review.

We would be happy to discuss any remaining questions or provide further clarifications. We hope the revised manuscript will help resolve your reservations, and we would be grateful if you would consider updating your assessments in light of the new evidence.

Thank you again for your time and thoughtful feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.