acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Memorization Mitigation in Diffusion Models: Reinforcing Text Conditioning

Abstract

Text-to-image diffusion models have achieved remarkable progress in image synthesis, yet can exhibit memorization by closely reproducing individual training examples. Effective mitigation must preserve useful prompt information to guide alternative depictions. We introduce a training-free method that redistributes cross-attention with Gaussian smoothing before reinforcing content-token contributions and attenuating padding contributions, without additional denoiser evaluations. With this intervention, stronger content conditioning can improve prompt alignment at comparable training-image similarity. A local analysis identifies when reinforcement preserves shared value information while redistribution reduces localized attention mass. On Stable Diffusion v1.4 and v2.0, all evaluated smoothing widths lie on the empirical Pareto frontiers for training-image similarity versus both prompt alignment and image preference. A configuration selected on Stable Diffusion reduces template reproduction in DeepFloyd IF without further tuning. These findings support jointly controlling conditioning allocation and strength to generate prompt-consistent alternatives.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
diffusion models, memorization mitigation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_aa2SOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduced a training free method which redistributes cross-attention with Gaussian smoothing and then averaging the content-token contributions and padding contributions. The authors did a comprehensive mathematical analysis and experiments to validate the efficacy of this method (better Pareto frontier in the CLIP-SSCD, ImageReward-SSCD tradeoffs in both SD v1.4 and SDv2.0) under certain assumptions.

Critical strengths

This paper is well written and the idea is clearly presented. Comprehensive experiments are done to illustrate that this idea can alleviate the memorization issue in text to image generation and have better Pareto frontier compared with existing works. The code is clean and experiments are well shown to support the idea. Proposition 1 provides a theoretical guarantee for reinforcement with localized dilution. The theory in appendix A.3 and A.4 provide a rigorous proof of complete row preservation and localized dilution and complementary roles of redistribution and gain.

Critical weaknesses

1.The assumption seems to be too strong so I am worried about the application of it in other cases. 2.The benchmark only has 500 memorized prompts. Can this method generalize to other datasets and have the same advantages over other methods.

Minor issue: 1.I would suggest to highlight/bold the values in the tables where the proposed method beats other methods, otherwise it is hard to get information from the table at the first sight. 2.It is better to write the characteristic function 1_{i\in T} in appendix A.1.

Questions for authors

1.“ We apply smoothing during an early denoising window to establish the redistributed conditioning pattern, then let Bk become the identity for the remaining steps. We use the first 5 of 50 steps for SD v1.4 and the first 10 for SD v2.0”--- is there any reasons why s/50 for SD v1.4 and 10/50 for SD v2.0?

2 what is l in equation (5) and why the assumption that |s_j-s_i|\leq l\j-i\ can be assumed? 3.Why the transfer to DeepFloydIF seems to have worse CLIP AND ImageReward compared to the base.

4 comments
AuthorsOpenReview ↗

Response [1/2]

Thank you for recognizing the empirical trade-offs and theoretical analysis. We clarify the assumptions, their practical applicability, and the model-specific settings below.

W1. Strength and scope of the theoretical assumptions

The assumption concerns local coherence of a projected feature, without requiring neighboring tokens to have identical full value vectors. It identifies a sufficient regime in which reinforcement preserves that feature while redistribution dilutes localized attention mass.

We tested the proposition on actual SD1.4 value vectors at σ=1\sigma=1 using all 500 prompts during the first five denoising steps, across cross-attention layers and heads on a fixed query grid. Each attention row describes one image query’s attention over text tokens. For each row, we selected the content token with the largest attention weight and used the mean content-token value to define the feature direction.

The local analysis applies when the selected token has a positive projection along this direction and its entire smoothing neighborhood stays within content tokens, away from boundaries. We checked the sufficient conditions at the applied content gain, λw=2\lambda_w=2.

Diagnostic and denominatorσ=1\sigma=1
Rows eligible for local analysis / all observed rows38.29%
Rows satisfying the local sufficient conditions / eligible rows92.15%
Rows satisfying the local sufficient conditions / all observed rows35.28%

These measurements show that the proposition’s local regime occurs on actual model values during sampling: redistribution and gain can preserve the selected token’s projected contribution while reducing the fraction of its attention mass returned to its original position.

W2. Generalization beyond the 500-prompt benchmark

Beyond the 500-prompt SD1.4 benchmark, our evaluations show favorable comparative trade-offs on SD2.0 and SD3.0, as well as suppression transfer to DeepFloyd IF.

On the SD2.0 benchmark of 219 memorized prompts, our method also exhibits a favorable empirical Pareto trade-off. The IF evaluation further covers 48 template-verbatim prompts from 11 reference groups, extending transfer to T5 conditioning and pixel-space generation.

We further evaluated the template-memorization prompts from the SD2.0 benchmark on SD3.0, which uses a diffusion-transformer backbone. We compared our method with Jain et al. [1] and random token addition (RTA) [2] under a common evaluation protocol across multiple parameter settings. The table reports all four evaluated operating points for each mitigation method.

MethodMean SSCD ↓CLIP ↑ImageReward ↑
Base0.16340.2576−0.1369
Jain et al. [1]0.15680.2552−0.1495
Jain et al. [1]0.15390.2548−0.1863
Jain et al. [1]0.14540.2527−0.2976
Jain et al. [1]0.14230.2511−0.3012
RTA [2]0.16120.2549−0.1507
RTA [2]0.16000.2548−0.1544
RTA [2]0.16000.2537−0.1589
RTA [2]0.15730.2514−0.1637
Ours0.15980.2563−0.0982
Ours0.14500.2539−0.1560
Ours0.13160.2496−0.2348
Ours0.10930.2376−0.4251

All four of our evaluated configurations lie on the empirical Pareto frontier in both the SSCD–CLIP and SSCD–ImageReward comparisons.

Minor comments

Thank you for the suggestion. Lower SSCD can come at the cost of CLIP and ImageReward, so highlighting individual metric optima could obscure the suppression–quality trade-offs represented by the different settings. We therefore leave these values unbolded; Figure 2 makes the comparison explicit by showing our method’s favorable trade-off relative to the baselines.

We have explicitly defined the indicator 1{i∈T}\mathbf{1}_{\{i\in\mathcal{T}\}} in Appendix A.1: it equals 1 when i∈Ti\in\mathcal{T} and 0 otherwise.

AuthorsOpenReview ↗

Response [2/2]

Q1. Smoothing duration in SD1.4 and SD2.0

The smoothing width σ\sigma controls the extent of attention redistribution, while the duration controls how long it is applied. Our early-window design is consistent with prior evidence that conditioning during early denoising can steer trajectories toward memorized images [1]. We restrict smoothing to a short initial window to redistribute conditioning early while limiting continued mixing of token-specific information during later denoising. We use the first five of 50 steps for SD1.4 and the first ten for SD2.0 as model-specific settings. Content reinforcement and padding attenuation continue afterward.

The table below reports SD2.0 results at σ=1\sigma=1 with smoothing applied during the first 5, 10, or 15 of the 50 denoising steps.

Initial smoothing stepsMean SSCD ↓CLIP ↑ImageReward ↑
5/500.22550.2596−0.5279
10/500.20900.2579−0.5482
15/500.19910.2575−0.5546

Q2. The meaning of ℓ\ell and the local regularity assumption

For a fixed unit direction uu, let si=u⊤vis_i=u^\top v_i denote the projection of token ii’s value vector onto that direction. The constant ℓ\ell bounds how rapidly these scalar projections vary over the kernel neighborhood: ∣sj−si∣≤ℓ∣j−i∣|s_j-s_i|\leq\ell|j-i|. It is a local discrete Lipschitz bound on the projected values. The assumption concerns a shared projected feature; it does not require neighboring tokens to have identical meanings or identical full value vectors.

Together with a positive projection on the localized component, this condition controls the feature variation introduced by smoothing. The local variation must be small enough for a gain to compensate for it while still diluting localized attention mass: ε+κT<1\varepsilon+\kappa_{\mathcal{T}}<1 gives the admissible interval 11−ε≤λw<1κT\frac{1}{1-\varepsilon}\leq\lambda_w<\frac{1}{\kappa_{\mathcal{T}}}. The actual-value measurements reported under W1 quantify how often this sufficient local condition holds in the model.

Q3. Quality trade-offs in DeepFloyd IF

Our objective is to improve the suppression–quality trade-off. As Figure 2 shows, our method achieves a favorable trade-off relative to competing mitigation methods on SD1.4, although CLIP and ImageReward can decrease relative to the unmitigated Base model. The IF results exhibit a similar pattern: template matches decrease from 56/144 to 29/144, accompanied by reductions in CLIP and ImageReward. Here, we directly transferred the SD-selected configuration without IF-specific tuning. These results demonstrate that the suppression effect extends to T5-conditioned, pixel-space generation, while the associated quality cost can vary with the model and intervention setting.

References

[1] Jain et al. Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization. CVPR, 2025.

[2] Somepalli et al. Understanding and Mitigating Copying in Diffusion Models. NeurIPS, 2023.

Thank you. That answered my question.

AuthorsOpenReview ↗

Response 2

We are glad to hear that our response answered your question. Thank you for taking the time to engage with our work and for your thoughtful feedback.

Reviewer_79j8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to mitigate the memory of training images in text-to-image diffusion models while preserving alignment with the input prompt. The authors propose a training-free method to intervent cross-attention, where Gaussian smoothing redistributes attention along text-token positions, and where content-token contributions are amplified and padding-token contributions after EOS token are attenuated. The method is compared with several inference-time mitigation approaches on Stable Diffusion v1.4 and v2.0, with an additional transfer to DeepFloyd IF. The experiments use SSCD for similarity to training reference images, CLIP for prompt alignment, and ImageReward for image preference.

Critical strengths
  • The method is simple and achieves a competitive trade-off between reducing training-image similarity and preserving image quality. It requires neither pre-training nor fine-tuning, and its attention-level operations do not add additional inference cost.
  • The ablation studies examine the roles of content amplification and attention smoothing. These controlled comparisons help explain what each component contributes to the overall similarity-quality trade-off.
Critical weaknesses
  • The practical relevance of the findings to larger and more recent text-to-image models remains insufficiently explored. The evaluation focuses on earlier-generation models and prompts already known to be associated with memorization. Although this is useful for testing mitigation, it remains unclear whether newer models still exhibit substantial memorization of training images.
  • Generalization across model architectures and text encoders is also not sufficiently demonstrated. The main experiments use SD v1.4 and v2.0, which adopt U-Net-based architectures and CLIP text encoders. The transfer to T5-based DeepFloyd IF is a useful step beyond CLIP, but it does not establish whether the method remains effective with more recent diffusion transformers or text encoders. Because the intervention on token positions and explicitly distinguishes content, EOS, and padding, its behavior under these architectural changes deserves further validation.
Questions for authors
  • It would be helpful to show more qualitative comparisons, as the paper currently includes only a few sets of images, making it difficult for readers to get a clear sense of the effects.
  • How does the method compare with meaning-preserving prompt rewriting or expansion? The proposed intervention changes how text features are mixed and weighted through cross-attention, while rewriting the prompt offers a straightforward alternative way to change the conditioning supplied to the model. Could a modest rewrite achieve a similar reduction in training-image similarity while preserving alignment with the original request? Since many current text-to-image models already incorporate prompt rewriting or expansion, this comparison would help clarify the additional value of the proposed method.
5 comments

Thank you for the questions about recent models, architectural generalization, and prompt rewriting. We extend the comparative evaluation to SD3.0 and directly compare our method with a modest rewriting baseline.

W1. Relevance to recent text-to-image models

Template memorization persists in recent text-to-image models. Yarkoni et al. [1] report template-memorized images in SD3.5 Medium and FLUX.1-schnell, showing that the issue extends beyond earlier Stable Diffusion models. These findings motivate evaluating mitigation on diffusion-transformer architectures.

W2. Generalization across architectures and text encoders

To test generalization beyond the U-Net architectures used in SD1.4 and SD2.0, we evaluated our method on SD3.0, which uses a multimodal diffusion transformer. Models in the SD3 family combine CLIP- and T5-based text encoders with their respective tokenizers. For this comparison, we selected the method of Jain et al. [2] and random token addition (RTA) [3], which can be applied to SD3.0 while retaining their original mitigation mechanisms. We used the template-memorization prompts from our SD2.0 benchmark and evaluated all methods under a common protocol across multiple parameter settings. The table reports all four evaluated operating points for each mitigation method.

MethodMean SSCD ↓CLIP ↑ImageReward ↑
Base0.16340.2576−0.1369
Jain et al. [2]0.15680.2552−0.1495
Jain et al. [2]0.15390.2548−0.1863
Jain et al. [2]0.14540.2527−0.2976
Jain et al. [2]0.14230.2511−0.3012
RTA [3]0.16120.2549−0.1507
RTA [3]0.16000.2548−0.1544
RTA [3]0.16000.2537−0.1589
RTA [3]0.15730.2514−0.1637
Ours0.15980.2563−0.0982
Ours0.14500.2539−0.1560
Ours0.13160.2496−0.2348
Ours0.10930.2376−0.4251

All four of our evaluated configurations lie on the empirical Pareto frontier in both the SSCD–CLIP and SSCD–ImageReward comparisons.

Q1. Additional qualitative comparisons

We have added SD1.4 qualitative comparisons in Figure 10 of Appendix K.1. In the examples involving blond hair and a green kitchen, our method generates compositions distinct from the training references while preserving the requested visual attributes. In the "Perform Slough'' example, our output depicts a singing performance, reflecting the action specified by the prompt.

Q2. Meaning-preserving prompt rewriting

Our method achieves lower mean SSCD than prompt rewriting at comparable CLIP and ImageReward scores.

Prompt modification has also been explored for memorization mitigation through random token addition (RTA), proposed by Somepalli et al. [3] and evaluated in our Figure 2. Following the reviewer’s suggestion, we additionally evaluated LLM-based rewriting designed to preserve the original meaning.

We compared rewriting and our method on the same 500 SD1.4 benchmark prompts with ten images per prompt. GPT-6-Luna received the following fixed instruction:

Rewrite each caption using different wording or word order. Preserve its meaning, named entities, numbers, product codes, quoted text, and original language. Add no visual or stylistic details. Return one rewritten caption per input, without alternatives or revision.

All 500 rewritten captions differed from their originals and were fixed before generation without image-score selection. CLIP and ImageReward were measured against the original captions to evaluate alignment with the original request.

MethodMean SSCD ↓CLIP ↑ImageReward ↑
RTA, 2 tokens (Table 14)0.41710.2778−0.1581
Prompt rewriting0.44680.2856−0.0627
Ours, σ=4\sigma=4, λp=0.4\lambda_p=0.40.30480.2875−0.0607

Our method achieves 31.8% lower mean SSCD than rewriting at comparable observed quality, and also improves all three reported metrics over the two-token RTA setting. Rewriting still leaves 47.52% of generated images above SSCD 0.5, showing that the tested rewrite does not eliminate high-similarity outputs. Our method operates on the original prompt and requires neither an external LLM nor additional denoiser evaluations.

References

[1] Yarkoni, Sharif, and Livni. Reconstructing Template-Memorized Images from Natural Prompts. ICML, 2026.

[2] Jain et al. Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization. CVPR, 2025.

[3] Somepalli et al. Understanding and Mitigating Copying in Diffusion Models. NeurIPS, 2023.

Thank you for adding the SD3.0 experiments and the prompt-rewriting comparison. I would appreciate clarification on whether the selected prompts produce clear memorization cases in SD3.0 before intervention, as the baseline mean SSCD is already relatively low. Without such evidence, it is difficult to interpret the further reduction as successful memorization mitigation.

AuthorsOpenReview ↗

Response 2

Thank you for raising this point. We have added Figure 13 to Appendix L to present clear cases of template memorization in SD3.0 before intervention, alongside the corresponding outputs of our method. Several outputs generated without intervention retain the relative placement of objects and surrounding structures seen in the reference images, even when the printed designs change. Each column within a panel pairs outputs generated with the same prompt and random seed, with and without our intervention. Template reproduction is not equally apparent across seeds. Since the reported mean SSCD aggregates all evaluated outputs, a relatively low mean can coexist with clear cases of template memorization in individual samples.

The paired examples show that the reference-associated arrangements become less apparent after applying our method. This sample-level evidence complements the aggregate SSCD–CLIP and SSCD–ImageReward trade-offs in Figure 14 and supports interpreting the SSCD reduction as mitigation of template memorization.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Response [1/2]

Thank you for recognizing our method’s favorable suppression–quality trade-off. Our new experiments address the mechanistic concerns using the deployed operator and actual model values, and quantify residual high-similarity events across baselines.

W1. Mechanistic controls for the deployed operator

Our method’s suppression cannot be explained by guidance magnitude attenuation alone. We repeated the direction/norm controls with the deployed unnormalized operator at σ=4\sigma=4 on all 500 SD1.4 prompts, generating four images per prompt. Both guidance fields are evaluated at the same current latent, with norms matched per image.

ConfigurationMean SSCD ↓CLIP ↑ImageReward ↑
Base0.54590.30210.0165
Base direction / Our method’s norm0.45450.2906−0.0863
Ours0.23110.2712−0.2645
Our method’s direction / Base norm0.22110.2640−0.3330

Matching our method’s norm while retaining the base direction yields mean SSCD 0.4545, compared with 0.2311 for our method. Conversely, retaining our method’s direction while restoring the base norm preserves substantial suppression, with mean SSCD 0.2211. These controls support a contribution from guidance direction beyond norm attenuation.

W2. Empirical verification of the theoretical conditions

At σ=1\sigma=1, the local preservation conditions hold for 92.15% of eligible attention rows, corresponding to 35.28% of all observed rows.

We tested the proposition on actual SD1.4 value vectors using all 500 prompts during the first five denoising steps, across cross-attention layers and heads on a fixed query grid. Each attention row describes one image query’s attention over text tokens. For each row, we selected the content token with the largest attention weight and used the mean content-token value to define the feature direction.

The local analysis applies when the selected token has a positive projection along this direction and its entire smoothing neighborhood stays within content tokens, away from boundaries. We checked the sufficient conditions at the applied content gain, λw=2\lambda_w=2.

Diagnostic and denominatorσ=1\sigma=1
Rows eligible for local analysis / all observed rows38.29%
Rows satisfying the local sufficient conditions / eligible rows92.15%
Rows satisfying the local sufficient conditions / all observed rows35.28%

These measurements show that the proposition’s local regime occurs on actual model values during sampling: redistribution and gain can preserve the selected token’s projected contribution while reducing the fraction of its attention mass returned to its original position.

W3. Residual copying risk and quality trade-offs

Our method at σ=1\sigma=1 achieves higher CLIP and ImageReward than both CS 1.30 [1] and AIN 1.1 [2], while yielding a larger reduction in the high-similarity event rate relative to Base.

We evaluated 500 SD1.4 prompts with ten images per prompt under a common protocol. The event rate is the fraction of images with SSCD>0.5\mathrm{SSCD}>0.5; Δrate\Delta\mathrm{rate} is the difference from Base in percentage points. The 95% CIs use 10,000 paired prompt-bootstrap resamples, retaining each prompt’s ten images.

Method and settingMean SSCD ↓CLIP ↑ImageReward ↑Per-image rate ↓Δrate\Delta\mathrm{rate} vs Base [95% CI], pp
Base0.54590.30210.016561.44%—
Ours σ=1\sigma=10.24100.2769−0.199915.42%−46.02 [−49.32, −42.62]
Ours σ=4\sigma=40.23110.2712−0.264512.96%−48.48 [−51.80, −45.08]
CS 1.150.30140.2718−0.135724.50%−36.94 [−40.18, −33.72]
CS 1.300.26010.2612−0.255719.48%−41.96 [−45.32, −38.54]
AIN 0.90.23800.2667−0.350815.58%−45.86 [−49.14, −42.56]
AIN 1.10.28870.2758−0.250122.28%−39.16 [−42.34, −36.00]

The IF experiment demonstrates transfer of suppression to T5 conditioning and pixel-space generation. The SD-selected configuration at σ=4\sigma=4 and α=0.625\alpha=0.625 reduces template matches from 56/144 to 29/144 without IF-specific tuning. ImageReward decreases from 0.3988 to 0.1058 at this transferred operating point, reflecting the suppression–quality trade-off. Appendix H evaluates these matches using spatially masked reference templates, showing a reduction of 18.75 percentage points.

AuthorsOpenReview ↗

Response [2/2]

W4. Combining our method with memorization detection

Like AIN [2] and the method of Jain et al. [3], our method provides an inference-time intervention to mitigate memorization. Appendix I illustrates how our method can be combined with the existing detector of Wen et al. [4] to selectively intervene while limiting unnecessary changes to ordinary generations. At the selected 5% false-positive operating point, the detector’s recall of approximately 0.55 limits the coverage of this gated configuration. Our main benchmarks apply our method directly and evaluate its suppression–quality trade-off independently of gating.

Q1. Figure 2 operating points and sensitivity to prompt resampling

Our objective in Figure 2 was to characterize the suppression–quality trade-off across a range of intervention strengths. The benchmark does not provide a separate validation split, and we used the same benchmark prompts for parameter exploration and evaluation. Starting from σ=1\sigma=1, λp=0.1\lambda_p=0.1, and λw=2\lambda_w=2, we increased σ\sigma and selected the reported operating points based on the observed trade-offs.

For prompt-resampling analysis, we kept the configurations in the new SD1.4 comparison fixed. Our σ=1\sigma=1 and σ=4\sigma=4 configurations remain nondominated in 100% and 99.96% of prompt-bootstrap resamples, respectively, on both the SSCD–CLIP and SSCD–ImageReward frontiers. Their frontier membership is therefore stable to prompt resampling within this comparison.

References

[1] Ren et al. “Unveiling and Mitigating Memorization in Text-to-Image Diffusion Models through Cross Attention.” ECCV, 2024.

[2] Han et al. “Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models.” NeurIPS, 2025.

[3] Jain et al. “Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization.” CVPR, 2025.

[4] Wen et al. “Detecting, Explaining, and Mitigating Memorization in Diffusion Models.” ICLR, 2024.

Thank you for the detailed response. The controls with the deployed operator clearly show that suppression stems from guidance direction rather than magnitude attenuation alone, and the thresholded event rates with confidence intervals show stronger suppression at comparable or better quality. The local conditions cover only part of the attention rows, and residual copying risk and the IF quality cost remain limitations, but my main concerns are addressed. I will raise my score from 2 to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.