acceptodds
Withdrawn from ICLR 2027

Analysis and Mitigation of Memorization in Multi-Modal Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-modal diffusion models, such as BLIP-Diffusion and ELITE, achieved impressive performance in generating personalized images conditioned on both a reference image and a textual prompt. Despite their success, it remains unexplored how these models memorize their training data. Existing memorization analyses for text-to-image diffusion models consider the text prompt as the only memorization factor without considering reference image input, making them inapplicable for the multi-modal setting. To address this gap, we propose a novel approach to effectively measure memorization in multi-modal diffusion models. First, our method localizes objects based on caption keywords, then masks each detected region, and measures how faithfully the model reconstructs the masked object. We show that our method provides equivalent results to the gold-standard leave-one-out memorization metric at a substantially lower computational cost. Using our new method, we find that pre-training samples for the multi-modal encoder exhibit consistently higher memorization levels than samples used to fine-tune the diffusion backbone. This indicates that the multi-modal encoder plays a more important role in driving memorization than the diffusion backbone itself. We further demonstrate that highly memorized training images are under high privacy leakage risks. This is driven by the joint memorization of vision–text modality associations, where the vision modality matters more. Finally, we show that applying memorization mitigation at the multi-modal encoder level achieves a significantly better privacy-utility tradeoff than diffusion backbone level mitigation.

Primary area:
generative models
Keywords:
Multi-Modal Diffusion Models, Memorization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vDCAOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies memorization in multi-modal diffusion models that use both reference images and text prompts. The authors propose a memorization score based on object masking and reconstruction similarity and find that multi-modal encoders contribute more to memorization than diffusion backbones. They also investigate privacy leakage through image reconstruction and compare mitigation strategies applied to the encoder and diffusion backbone.

Critical strengths

The paper addresses an important problem in multi-modal diffusion models. Existing memorization studies mainly focus on text-conditioned generation, while this work considers the additional role of reference images.

The experiments cover two different architectures, BLIP-Diffusion and ELITE. The authors compare training and non-training samples, validate their metric against a leave-one-out approach, and reverse the training datasets to examine potential distribution effects. These experiments provide useful evidence for the main observations.

Critical weaknesses

The main weakness of the paper is that the proposed memorization metric is not fully validated. Object reconstruction may reflect general predictability rather than memorization. Although Table 1 reports over 91% overlap with the LOO metric, Kendall's Tau is relatively low (0.186–0.266). Considering that the LOO metric measures representation stability rather than image reconstruction, the claim of equivalence seems too strong.

Also, the contribution of the encoder is not clear. Table 2 shows that encoder fine-tuning reduces memorization scores on both datasets. Random initialization also removes useful visual information, not just memorization. These results show that the encoder is important for reconstruction, but do not clearly demonstrate that it is the main source of memorization.

Finally, since the encoder-level and diffusion-level mitigation methods use different techniques, it is difficult to determine whether the improvement comes from the intervention location or the mitigation method itself. Without comparisons under the same conditions, it is unclear whether encoder-level mitigation is actually more effective.

Questions for authors

Q1. Could the authors conduct more controlled ablation studies to isolate the encoder's contribution to memorization? For example, selectively modifying the encoder and diffusion backbone while controlling their general reconstruction capabilities would provide stronger evidence for the main claim.

Q2. Could the authors compare encoder-level and diffusion-level mitigation under more comparable conditions?

Q3. Could the authors provide additional modality ablations for Table 3, including removing target caption information from both the encoder input and diffusion prompt, and measuring similarity between retrieved reference images and target images?

Q4. Could the authors provide a direct computational cost comparison between the proposed metric and the LOO baseline under the same evaluation setting?

2 comments
AuthorsOpenReview ↗

LOO Consistency, Encoder Contribution, and Mitigation Comparison

W1: The main weakness of the paper is that the proposed memorization metric is not fully validated. Object reconstruction may reflect general predictability rather than memorization. Although Table 1 reports over 91% overlap with the LOO metric, Kendall's Tau is relatively low (0.186–0.266). Considering that the LOO metric measures representation stability rather than image reconstruction, the claim of equivalence seems too strong.

We agree that “equivalent” overstates what Table 1 shows. We revise it to “consistent with the LOO metric, especially on the most memorized samples.” We clarify two points. (1) Overlap should be read against chance: with ∣DC∣|D_C| = 100K, random top-K overlap is 0.1% / 1% / 5% for K = 100 / 1000 / 5000, versus 91.0–94.2% observed. (2) We now additionally report Kendall’s test over all Candidate sets with τ = 0.684 / p = 2.21e-188.

W2: Also, the contribution of the encoder is not clear. Table 2 shows that encoder fine-tuning reduces memorization scores on both datasets. Random initialization also removes useful visual information, not just memorization. These results show that the encoder is important for reconstruction, but do not clearly demonstrate that it is the main source of memorization.

The main idea of table 2 is not that the memorization level has decreased, but rather that the gap in memorization levels between CC3M and COCO-stuff has been narrowed. We observed that fine-tuning the encoder using the training samples from the diffusion backbone (COCO-Stuff) significantly narrowed the memory gap between CC3M and COCO-Stuff (second experiments in table 2). This experiment acts as a more extreme test based on that finding: we discovered that even when the diffusion backbone has memorization of the COCO-Stuff training samples, the memorization levels for CC3M and COCO-Stuff become the same once the encoder’s memory is completely erased. Therefore, we conclude that the memorization within the encoder plays the primary role.

W3: Finally, since the encoder-level and diffusion-level mitigation methods use different techniques, it is difficult to determine whether the improvement comes from the intervention location or the mitigation method itself. Without comparisons under the same conditions, it is unclear whether encoder-level mitigation is actually more effective.

We agree that currently there is currently no direct way to quantitatively compare their mitigation strengths because these two approaches operate on different model architectures and pre-training paradigms. To make this comparison as fair as possible, we have applied the most effective mitigation method described in the original paper for each approach.

Q1: Could the authors conduct more controlled ablation studies to isolate the encoder's contribution to memorization? For example, selectively modifying the encoder and diffusion backbone while controlling their general reconstruction capabilities would provide stronger evidence for the main claim.

If the encoder is further fine-tuned, to preserve the same generalization ability of model (performance), the diffusion-backbone need to be fine-tuned as well. This is because the multi-modal encoder and the diffusion backbone within the multi-modal diffusion model are trained separately, in two distinct stages. Therefore, to evaluate the whole model’s memorization level, it’s nearly impossible to isolate any components without changing model itselves generalization ability (performance).

Q2: Could the authors compare encoder-level and diffusion-level mitigation under more comparable conditions?

Since these two approaches operate on different model architectures and pre-training paradigms, there is currently no direct way to quantitatively compare their respective strengths. To make this comparison as fair as possible, we have applied the most effective mitigation method described in the original paper for each approach.

AuthorsOpenReview ↗

Image-Retrieved Modality Analysis and Computational Cost Comparison with LOO

Q3: Could the authors provide additional modality ablations for Table 3, including removing target caption information from both the encoder input and diffusion prompt, and measuring similarity between retrieved reference images and target images?

We add extra experiments that retrieve the most related image according to the target image (instead of caption) and do the reconstruction based on these images and blank captions. To be noted that this experiment has nothing to do with the original privacy leakage risk but only used to show the modality importance during reconstruction. The results are shown in table below:

SettingBLIP-Diffusion SSIM ↑BLIP-Diffusion LPIPS ↓BLIP-Diffusion SSCD ↑ELITE SSIM ↑ELITE LPIPS ↓ELITE SSCD ↑
Most memorized samples (image retrieved by target image + blank caption)0.166 ± 0.0140.779 ± 0.0280.099 ± 0.0100.155 ± 0.0130.791 ± 0.0250.091 ± 0.008

The results achieve even better results than the original results for Retrieve Image + Blank Caption in table 3. This proves that Visual modality plays a more important role in this reconstruction.

Q4: Could the authors provide a direct computational cost comparison between the proposed metric and the LOO baseline under the same evaluation setting?

We report the evaluation for 1M samples and training cost for these 2 methods (on BLIP-Diffusion) in table below:

Extra Training TimeEvaluation TimeTotal overhead
Our method0154.1 h154.1 h
LOO118.6 h147.9 h266.5 h

The training is done on 4 H100 GPU and evaluation is done on 4 A100 GPU. Therefore, the actual additional overhead incurred by LOO is far greater than that of our proposed method.

Reviewer_L7scOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies memorization in multimodal diffusion models. It proposes an object-level memorization metric based on masked reconstruction, and validates it through training/non-training comparisons and a leave-one-out-inspired metric. The authors analyze memorization across model components and input modalities, investigate privacy leakage risks, and compare encoder-level and diffusion-level mitigation strategies.

Critical strengths

The paper studies memorization in image-conditioned diffusion models, which is a relatively underexplored problem. The proposed object-level evaluation approach has practical potential. The experiments across two architectures provide some encouraging evidence, particularly the separation between training and non-training samples.

Critical weaknesses
  1. My major concern is the validation of the proposed memorization metric.
  • In section 4, the authors argue that reconstructing masked objects relies more on memorization than generalization, because the model cannot generalize from background information that is ignored during training. I find this argument unconvincing, as background context and the object keywords provided in the prompts can provide useful information for reconstructing objects through generalization. Moreover, object and background reconstruction may have inherently different difficulties.
  • The claimed equivalence to the gold-standard LOO metric is a bit questionable. Although the top-K overlap is high, Kendall's τ is relatively low, which suggests relatively weak agreement in the rankings.
  • The metric also relies on Grounding DINO for object localization, but i wonder how robust this choice is for different object detectors or localization errors.
  1. Another concern is on the main conclusion that encoder memorization contributes more to memorization than the diffusion backbone, but i think the evidence is not enough to support this claim.
  • The authors conclude this based on comparisons between data examples used to train different components. However, these comparisons involve many different confounding factors (e.g. different training objectives, datasets, and potentially different numbers of training exposures). Although the authors reverse the datasets in an additional experiment, these other factors remain uncontrolled.
  • The component ablations are also a bit questionable. For instance, replacing the encoder with random weights inevitably disrupts conditioning and generation quality, which makes it difficult to attribute the reduced reconstruction scores specifically to reduced memorization.
  • The modality analysis is also confounded by retrieving reference images using the target captions. This means that the visual conditioning already contains information derived from the text modality.
Questions for authors
  1. Can the authors provide other evidence that the proposed metric distinguishes memorization from generalization, and evaluate its robustness to different object localization models?

  2. Can the authors provide more controlled comparisons to isolate the encoder's contribution to memorization?

2 comments
AuthorsOpenReview ↗

Memorization vs. Generalization, LOO Consistency, Detector Robustness, and Training-Exposure Confound

W1: In section 4, the authors argue that reconstructing masked objects relies more on memorization than generalization, because the model cannot generalize from background information that is ignored during training. I find this argument unconvincing, as background context and the object keywords provided in the prompts can provide useful information for reconstructing objects through generalization. Moreover, object and background reconstruction may have inherently different difficulties.

We show in Figures 2 & 4 that, when masking objects, our method can effectively distinguish between training samples, non-training IID samples, and OOD samples. In particular, non-training IID samples share the same distribution as the training samples and should therefore benefit from a level of model generalization capability similar to that of the training samples. This indicates that our object-masking method captures the model's memorization of training samples, in addition to its generalization capability.

W2: The claimed equivalence to the gold-standard LOO metric is a bit questionable. Although the top-K overlap is high, Kendall's τ is relatively low, which suggests relatively weak agreement in the rankings.

We agree that “equivalent” overstates what Table 1 shows. We revise it to “consistent with the LOO metric, especially on the most memorized samples.” We clarify two points. (1) Overlap should be read against chance: with ∣DC∣|D_C| = 100K, random top-K overlap is 0.1% / 1% / 5% for K = 100 / 1000 / 5000, versus 91.0–94.2% observed. (2) We now additionally report Kendall’s test over all Candidate sets with τ = 0.684 / p = 2.21e-188.

W3: The metric also relies on Grounding DINO for object localization, but I wonder how robust this choice is for different object detectors or localization errors.

We implemented additional experiments on Florence-2 and YOLO-World. Then, we tested the Kendall Tau test on all samples (with memorization score sorted from high to low). The results are shown in table below:

Florence-2YOLO-World
Tau / p0.914 / 00.907 / 0
Time usage (to G-Dino)119.3%78.2%

Both Florence-2 and YOLO-World show very high consistency with the current using Grounding-Dino. This shows that the metric is robust to the choice of object detectors.

W4: The authors conclude this based on comparisons between data examples used to train different components. However, these comparisons involve many different confounding factors (e.g. different training objectives, datasets, and potentially different numbers of training exposures). Although the authors reverse the datasets in an additional experiment, these other factors remain uncontrolled.

We acknowledge that, due to the different dataset size between CC3M and COCO-stuff (and in an effort to ensure the total training volume remains same to that reported in the original paper of Blip-diffusion and Elite), individual training samples are exposed to the encoder and the diffusion backbone a different number of times. However, we believe that swapping the training sets effectively addresses this issue: samples that were previously exposed more frequently to the diffusion model are now exposed more frequently to the encoder. Our experimental results show that samples trained on the encoder still exhibit a higher level of memorization, rather than diffusion backbone due to the dataset exchange.

W5: The component ablations are also a bit questionable. For instance, replacing the encoder with random weights inevitably disrupts conditioning and generation quality, which makes it difficult to attribute the reduced reconstruction scores specifically to reduced memorization.

In table 2, we are not using randomly initialized weights to demonstrate a decline in memorization, but rather to verify whether the CC3M and COCO-stuff datasets exhibit similar levels of memory retention when the encoder itself contains no memory. We observed that fine-tuning the encoder using the training samples from the diffusion backbone (COCO-Stuff) significantly narrowed the memorization gap between CC3M and COCO-Stuff (second experiments in table 2). This experiment acts as a more extreme test based on that finding: we discovered that even when the diffusion backbone has memorization of the COCO-Stuff training samples, the memorization levels for CC3M and COCO-Stuff become the same once the encoder’s memory is completely erased. Therefore, we conclude that the memorization within the encoder plays the primary role.

AuthorsOpenReview ↗

Encoder Ablation Clarification and Image-Retrieved Modality Analysis

W6: The modality analysis is also confounded by retrieving reference images using the target captions. This means that the visual conditioning already contains information derived from the text modality.

We agree that this is insufficient. We add extra experiments that retrieve the most related image according to the target image (instead of caption) and do the reconstruction based on these images and blank captions. To be noted that this experiment has nothing to do with the original privacy leakage risk but only used to show the modality importance during reconstruction. The results are shown in table below:

SettingBLIP-Diffusion SSIM ↑BLIP-Diffusion LPIPS ↓BLIP-Diffusion SSCD ↑ELITE SSIM ↑ELITE LPIPS ↓ELITE SSCD ↑
Most memorized samples (image retrieved by target image + blank caption)0.166 ± 0.0140.779 ± 0.0280.099 ± 0.0100.155 ± 0.0130.791 ± 0.0250.091 ± 0.008

We achieve even better results than the original ones for Retrieve Image + Blank Caption in Table 3. This shows that the visual modality plays a more important role in this reconstruction.

Q1: Can the authors provide other evidence that the proposed metric distinguishes memorization from generalization, and evaluate its robustness to different object localization models?

In previous research, the standard approach for evaluating memorization is to distinguish training samples and IID non-training samples, and then verifying whether the measured most memorized samples are under higher risk of privacy leakage [1][2][3]. We evaluate the robustness with Florence-2 and YOLO-World, and results already discussed in W3.

Q2: Can the authors provide more controlled comparisons to isolate the encoder's contribution to memorization?

We provide extra experiments with retrieving images directly with the target image instead of caption. The results are already discussed in W6. Again, this experiment has nothing to do with the original privacy leakage risk but is used only to show the modality importance during reconstruction.

References:

[1] Feldman, Vitaly. "Does learning require memorization? a short tale about a long tail." Proceedings of the 52nd annual ACM SIGACT symposium on theory of computing. 2020.

[2] Wang, Wenhao, et al. "Memorization in self-supervised learning improves downstream generalization." International Conference on Learning Representations. Vol. 2024. 2024.

[3] Wang, Wenhao, et al. "Captured by captions: On memorization and its mitigation in CLIP models." International Conference on Learning Representations. Vol. 2025. 2025.

Reviewer_NSyQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates memorization and privacy leakage in multi-modal diffusion models conditioned on both reference images and text prompts. The authors propose a memorization metric based on caption-grounded object masking and reconstruction, and validate it against a leave-one-out (LOO) representation-based measure. Experiments on BLIP-Diffusion and ELITE suggest that multi-modal encoders contribute more to memorization than diffusion backbones. The paper further examines privacy leakage through training-image reconstruction and reports that encoder-level mitigation achieves a better privacy-utility tradeoff than diffusion-level mitigation.

Critical strengths
  1. Important problem and practical insights. The paper addresses memorization in multi-modal diffusion models, where both visual and textual inputs may expose training information. The investigation of different architectural components provides useful insights into potential sources of memorization and possible mitigation strategies.
  2. Comprehensive experimental investigation. The authors evaluate two representative architectures, compare training and non-training samples, and include LOO-based validation and reversed-dataset experiments. These analyses strengthen the empirical evidence and provide useful support for the proposed findings.
Critical weaknesses
  1. Insufficient evidence for equivalence to the LOO memorization metric. Although Table 1 reports over 91% top-K overlap, the Kendall's Tau correlations remain relatively low (0.186–0.266), suggesting limited agreement in sample rankings. Moreover, the validation uses a large candidate-set replacement rather than conventional sample-level leave-one-out training. Therefore, the claim of equivalence to the gold-standard LOO metric appears stronger than the current evidence supports.
  2. Limited isolation of memorization in privacy leakage and mitigation evaluations. In Table 3, reconstruction relies on semantically retrieved reference images and target captions, making it difficult to determine how much similarity results from memorization rather than semantic conditioning or generalization. Additionally, the encoder-level and diffusion-level mitigation strategies use different intervention mechanisms, while their utility is evaluated mainly through CLIP and aesthetic scores. More controlled comparisons are needed to substantiate the claimed privacy-utility advantage.
Questions for authors
  1. Given the relatively low Kendall's Tau values in Table 1, what evidence supports the claimed equivalence between the proposed metric and LOO memorization beyond identifying overlapping top-K samples?
  2. Can the authors provide more controlled privacy reconstruction experiments that distinguish instance-specific memorization from reconstruction enabled by semantically similar reference images and captions?
  3. Can the encoder-level and diffusion-level mitigation strategies be compared under matched utility degradation or computational budgets, with additional evaluations of subject preservation and privacy attack success rates?
1 comment
AuthorsOpenReview ↗

Clarified LOO Consistency, Model-g Control Experiment for Instance-Specific Memorization, and Mitigation Comparison

W1: Insufficient evidence for equivalence to the LOO memorization metric. Although Table 1 reports over 91% top-K overlap, the Kendall's Tau correlations remain relatively low (0.186–0.266), suggesting limited agreement in sample rankings. Moreover, the validation uses a large candidate-set replacement rather than conventional sample-level leave-one-out training. Therefore, the claim of equivalence to the gold-standard LOO metric appears stronger than the current evidence supports.

We agree that “equivalent” overstates what Table 1 shows. We revise it to “consistent with the LOO metric, especially on the most memorized samples.” We clarify two points. (1) Overlap should be read against chance: with ∣DC∣|D_C| = 100K, random top-K overlap is 0.1% / 1% / 5% for K = 100 / 1000 / 5000, versus 91.0% / 92.1% / 94.2% observed. (2) We now additionally report Kendall’s test over all Candidate sets with τ = 0.684 / p = 2.21e-188.

W2: Limited isolation of memorization in privacy leakage and mitigation evaluations. In Table 3, reconstruction relies on semantically retrieved reference images and target captions, making it difficult to determine how much similarity results from memorization rather than semantic conditioning or generalization. Additionally, the encoder-level and diffusion-level mitigation strategies use different intervention mechanisms, while their utility is evaluated mainly through CLIP and aesthetic scores. More controlled comparisons are needed to substantiate the claimed privacy-utility advantage.

In Table 3, we applied the same reconstruction process to both the "most memorized" and "least memorized" samples. It is evident that the results for the "least memorized" samples are significantly worse than those for the "most memorized" ones. Furthermore, we conducted separate experiments with blank textual and visual semantic information; in both cases, the quality of the reconstructed images declined markedly. We believe these findings provide sufficient support for our conclusion. To further address this, we reconstruct the same 1000 most images with LOO paired model g (has very similar generalization ability with model f) and the results are:

SettingBLIP-Diffusion SSIM ↑BLIP-Diffusion LPIPS ↓BLIP-Diffusion SSCD ↑ELITE SSIM ↑ELITE LPIPS ↓ELITE SSCD ↑
Top-1000 most memorized samples of ff, reconstructed by gg (RI + TC)0.1140.8250.0790.1010.8360.072

For model g with a similar generalization ability, the reconstruction results for the model f's 1,000 most memorized samples are lower than those for the samples least memorized by model f itself. This is sufficient to rule out the influence of generalization ability.

Q1: Given the relatively low Kendall's Tau values in Table 1, what evidence supports the claimed equivalence between the proposed metric and LOO memorization beyond identifying overlapping top-K samples?

Beyond top-K overlap, we provide three additional results with full-candidate-set rank correlations. We believe this supports the claim that our metric’s results are aligned with highly memorized samples as LOO.

Q2: Can the authors provide more controlled privacy reconstruction experiments that distinguish instance-specific memorization from reconstruction enabled by semantically similar reference images and captions?

We ran additional experiments that reconstruct the model f’s most memorized samples with similar-performance model g and results show that semantic conditioning affects both equally and the difference isolates instance-specific memorization. Besides, our experiments in table 3 that provide blank text or image input also show that this is not driven by semantically similar reference.

Q3: Can the encoder-level and diffusion-level mitigation strategies be compared under matched utility degradation or computational budgets, with additional evaluations of subject preservation and privacy attack success rates?

Since these two approaches operate on different model architectures and pre-training paradigms, there is currently no direct way to quantitatively compare their respective strengths. To make this comparison as fair as possible, we have applied the most effective mitigation method described in the original paper for each approach.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal

We thank the Reviewers for their thoughtful and encouraging feedback, which greatly helped us further improve our submission. We have replied to each Reviewer's weaknesses and questions individually under their respective reviews. Based on these comments, we have revised our submission, and all changes are highlighted in blue in the updated PDF. The main revisions are:

  1. Claim on LOO consistency: We now describe our metric as consistent with the leave-one-out (LOO) metric, especially on the most memorized samples, rather than equivalent to it. We also report the chance-level top-KK overlap (0.1%0.1\% / 1%1\% / 5%5\% vs. 91.0%91.0\%–94.2%94.2\% observed) and Kendall's τ\tau over the full candidate set (τ=0.684\tau = 0.684, p=2.21×10−188p = 2.21 \times 10^{-188}).
  2. Robustness of the metric: We added results using Florence-2 and YOLO-World as alternative object detectors (Kendall's τ=0.914\tau = 0.914 and 0.9070.907 against Grounding DINO). We also explain more clearly why our metric captures memorization beyond generalization.
  3. Controlled privacy experiments: We added two experiments. In the first, the LOO reference model gg reconstructs the most memorized samples of model ff, which isolates instance-specific memorization from semantic conditioning. In the second, the reference image is retrieved by the target image together with a blank caption, which isolates the role of each modality.
  4. Encoder contribution and mitigation: We clarified how to read the Table 2 ablation and the effect of different numbers of training exposures. We also explain how the encoder-level and diffusion-level mitigation strategies were compared.
  5. Computational cost: We added a direct cost comparison with the LOO baseline (154.1 h for our method vs. 266.5 h for LOO on 1M samples).

We hope that our responses and revisions address the Reviewers' concerns, and we would be happy to discuss any remaining questions or suggestions further.

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

Related papers

Discussion (0)

Sign in to comment.