acceptodds
Under review as a conference paper at ICLR 2027

Mitigating Memorization In Text-To-Image Diffusion Models With Representation-Space Supervision

Abstract

Diffusion models are prone to memorization, i.e., reproducing individual training examples, which raises concerns about data privacy and copyright. Existing mitigation approaches primarily intervene through data preprocessing or specialized sampling procedures, leaving the role of the training objective comparatively underexplored. In this work we study whether changing the space in which diffusion predictions are supervised can mitigate memorization. We posit that direct data-space supervision contributes to memorization by encouraging model predictions to recover the exact training samples. Motivated by this view, we introduce Representation-Space Supervision (RSS), a general training framework that instead of supervising the models directly in data space, matches the model predictions to their targets in the feature spaces of pretrained visual representation models, which retain high-level perceptual and semantic information while reducing specific instance-level details in the training data that are essential for memorization. We evaluate RSS in controlled text-to-image supervised fine-tuning experiments, showing that it consistently and substantially improves the trade-off between generative performance and training-example reproduction relative to standard diffusion training.

Primary area:
generative models
Keywords:
diffusion model, memorization, privacy

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_RrMZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses memorization problem of diffusion models with a quite straightforward idea: revising prediction errors to those in a projected representation space. The paper mainly considers a fine-tuning scenario where the dataset is composed of public and private images. Through supervising the training w.r.t private images in a representations space, it shows that the resemblance to private training images can be controlled.

Critical strengths

The simple but effective idea of representation-space supervision for memorization-free training highlights the main contribution of this paper. Experimental results successfully demonstrate the superior performance-memorization trade-off of models trained using the proposed approach.

Critical weaknesses
  1. Application scenario. This paper mainly considers a downstream fine-tuning scenario where a subset of private data should be protected. Although quite common, it seems that the current training strategy cannot handle fine-tuning on all-private dataset, which could also happen in practice. According to the current objective function, fine-tuning on private data uses only RSS, while in lines 252-253, the paper states "In preliminary experiments, however, training with RSS alone can produce unwanted visual artifacts, ......".
  2. Robustness. As concluded by previous studies, memorization in diffusion models is caused by multifaceted factors, including training data duplication, training duration, model capacity. Evaluation in this paper lacks a comparison under variations of these factors, making the conclusion not sufficiently supported.
  3. Novelty. Training generative models with representation-space supervision is widely adopted technique often referred to as perceptual loss. Although adopted in a new problem, it still limits the novelty of this work.
Questions for authors
  1. The authors should discuss how to privately train diffusion models when the whole dataset is expected to be protected.
  2. How robust is the proposed training strategy when training dataset includes duplicated samples, the model is trained for longer time and bigger models are trained.
  3. The authors may justify the novelty of the key idea.
1 comment

Thank you for the thoughtful and constructive review. We appreciate your recognition of the practical motivation of RSS and its generation-quality--memorization trade-off. Below, we address your concerns.

Q1 — How can RSS train when the whole dataset is protected?

We thank the reviewer for highlighting this important setting. The current method is designed for mixed public/protected fine-tuning: standard diffusion supervision on public examples anchors low-level generation quality, while RSS allows protected examples to contribute semantic information without directly enforcing pixel-level reconstruction. We agree that an all-protected dataset removes this public anchor. In preliminary RSS-only experiments, we observed unwanted visual artifacts. We therefore do not claim that the present method fully solves the strictly all-private, zero-public-data setting, and we will make this scope explicit in the revision.


Q2 — Robustness to duplicated samples, longer training, and larger models. How robust is the proposed training strategy when the training dataset includes duplicated samples, the model is trained for longer time, and bigger models are trained?

Duplicated samples

We conducted a stress test using deliberately repeated protected canaries. At 100K steps, the test contains 8, 14, and 20 unique private canaries for the 50/50, 25/75, and 10/90 splits, respectively. The observed canary copy rates were 0% for RSS, compared with 63%, 50%, and 55% for Full-Data SFT across the three splits.

Longer training

For the 50/50 split, we extended the RSS run from 100K to 200K steps:

RSS checkpointGenEval ↑DPG ↑Private copy rate ↓
100K69.5277.383.0%
200K68.9877.084.9%

Generation quality remains stable from 100K to 200K, while private copying increases modestly from 3.0% to 4.9%. This longer run therefore supports robustness over twice the original training horizon.

We also repeated the 25/75 RSS experiment at 100K with a second independent training seed:

MetricSeed 0Seed 1Absolute difference
GenEval ↑66.9666.810.15
DPG ↑74.8676.711.85
Private SSCD ↓23.9824.300.32
Public SSCD ↓32.0632.060.00

The central quality--memorization behavior is consistent across these two runs.

Larger models

The present controlled experiments use an approximately 1B-parameter DiT, and we do not currently have a controlled larger-model run. Nevertheless, memorization is already substantial at this scale, and our results show that RSS significantly mitigates it under our evaluation protocol. Whether this benefit persists at larger model scales remains an important direction for future work.


Q3 — What is novel relative to perceptual loss?

To the best of our knowledge, RSS is the first training-time method to mitigate memorization in diffusion models by selectively replacing direct reconstruction supervision on protected examples with representation-space supervision. We do not claim that representation-based losses themselves are new, our novelty lies in using them selectively according to the protection status of each training example.

Reviewer_8qsFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Representation-Space Supervision (RSS) to reduce memorization during text-to-image diffusion fine-tuning. Protected examples are supervised by matching predicted images and training targets in a frozen visual encoder’s feature space, while public examples retain the standard diffusion objective. A simplified linear analysis illustrates how transformed supervision relaxes the requirement to reconstruct every component of a training sample. The main experiments fine-tune an approximately 1B-parameter model on 120K synthetic image–text pairs, comparing three public/private splits and multiple training checkpoints. RSS improves the measured trade-off between generation performance and average training-image similarity relative to full-data and public-only fine-tuning.

Critical strengths
  1. The paper identifies a practical use of representation-based supervision for memorization mitigation. The important design choice is to remove direct reconstruction supervision from protected examples while retaining their contribution through visual features. This differs from adding representation alignment to an otherwise unchanged diffusion objective, as in REPA [1]. Standard supervision on public data provides an anchor for generation quality. The resulting method addresses learning from protected examples without requiring a specialized sampling procedure.

  2. The main experiments provide meaningful evidence of a better measured trade-off. Figure 1 compares methods from the same pretrained checkpoint across three data partitions and multiple training stages. Section 5.2 also considers each method’s best-GenEval checkpoint, addressing the concern that the improvement merely reflects comparison against an overtrained baseline. Under the 50/50 split, RSS achieves GenEval 69.8 with mean SSCD 25.9, compared with 68.3 and 44.5 for full-data fine-tuning, using the paper’s reported scales. The improvement over public-only training additionally supports that RSS extracts useful information from the protected subset.

[1] Yu, S., Kwak, S., Jang, H., Jeong, J., Huang, J., Shin, J., & Xie, S. (2025). Representation alignment for generation: Training diffusion transformers is easier than you think. ICLR 2025

Critical weaknesses
  1. The main memorization metric does not sufficiently establish a reduction in actual copying. Section 5.1 averages similarity over four generations per training prompt and then over 2,000 prompt–image pairs. A lower mean can coexist with a substantial population of near-exact copies, and comparing each generation only with its corresponding training image can miss reproduction of other training examples. This distinction is consequential because the paper’s central claim concerns training-example reproduction, rather than average resemblance. Somepalli et al. [2], for example, evaluate nearest-training-image similarity and emphasize the high-similarity tail. Figure 2 provides useful qualitative evidence, but thresholded copying rates, upper-tail statistics, and nearest-neighbor checks over the protected set are needed to substantiate the claimed mitigation.

  2. The improvement over closely related training-time mitigation methods remains unclear. The main comparison in Figure 1 establishes an advantage over conventional full-data and public-only fine-tuning. However, Shah et al. [3] already modify diffusion training through ambient supervision to improve the quality–memorization trade-off, and Somepalli et al. [2] propose training-time caption diversification and randomization. Both are relevant alternatives for the fine-tuning setting studied here. Section 5.3 reports additional comparisons with masked-pixel supervision and DP-SGD, which are useful but do not establish the advantage over these existing approaches. Since representation-based objectives already have substantial precedent, stronger comparisons are important for determining the added value of selectively replacing direct supervision on protected examples.

[2] Somepalli, G., Singla, V., Goldblum, M., Geiping, J., & Goldstein, T. (2023). Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36, 47783-47803.

[3] Shah, K., Kalavasis, A., Klivans, A. R., & Daras, G. (2025). Does generation require memorization? creative diffusion models using ambient diffusion. ICML 2025

Questions for authors
  1. Does RSS also reduce thresholded copying rates and the upper tail of SSCD similarity when generated images are compared against their nearest neighbors across the protected training set? Reporting these quantities from the existing evaluation generations would help distinguish reduced average resemblance from reduced reproduction.

  2. How does RSS compare with ambient-diffusion training [3] and training-time caption diversification or randomization [2] under the same initialization, data partitions, and evaluation protocol? Comparisons at similar generation performance, with training cost reported, would materially affect my assessment of the contribution.

5 comments

Thank you for the thoughtful and constructive review. We appreciate your recognition of the practical motivation of RSS and the measured generation-quality--memorization trade-off. Below, we address your questions with additional nearest-neighbor, baseline, performance-matched, and computational-cost results.

Q1 — Does RSS reduce thresholded copying and the upper tail of nearest-neighbor SSCD? Does RSS also reduce thresholded copying rates and the upper tail of SSCD similarity when generated images are compared against their nearest neighbors across the protected training set?

We agree that mean similarity to the paired training target may conceal a high-similarity tail and may miss reproduction of a different protected example. We therefore added an exhaustive protected-set nearest-neighbor evaluation using the same SSCD model as our other copying evaluations.

For the 50/50 split, we use 1,000 private probe prompts and generate four images per prompt. For each prompt, we take the maximum SSCD cosine similarity over both its four generations and all 61,706 protected training images; denote this value by sis_i. The copy rate is the fraction of prompts with si>0.5s_i>0.5. The NN mean is the average of sis_i over all prompts. P95 is the value below which 95% of prompts fall, so only the highest-scoring 5% of prompts exceed it; P99 is defined analogously for the highest-scoring 1%. These statistics therefore describe the overall level and the high-similarity tail, respectively. SSCD statistics are multiplied by 100.

Method, 100KNN copy rate ↓NN mean ↓P95 ↓P99 ↓
Full-Data SFT84.90%58.98%72.50%76.63%
Public-Only SFT0.20%26.15%32.35%43.16%
RSS4.20%35.94%49.17%57.28%

The full thresholded trajectory gives NN copy rates of 48.0%, 84.7%, and 84.9% for Full-Data SFT at 20K, 50K, and 100K, compared with 9.0%, 5.2%, and 4.2% for RSS. Thus, RSS substantially reduces both thresholded nearest-neighbor copying and the upper tail of protected-set similarity.

Q2 — Comparison with Ambient training and caption diversification/randomization. How does RSS compare with ambient-diffusion training [3] and training-time caption diversification or randomization [2] under the same initialization, data partitions, and evaluation protocol?

We agree that Ambient supervision and caption randomization are important training-time comparisons. We evaluate two Ambient noise levels and two caption interventions using the same 50/50 initialization, partition, training horizon, checkpoint schedule, and evaluation protocol as RSS.

For caption randomization, each caption undergoes two perturbation trials. Each trial is activated with probability 0.1 and either replaces a randomly selected word/token or inserts a random word.

Details of our implementation is given in appendix E of our revised draft.

50/50 results at 100K steps

MethodGenEval ↑DPG ↑Private copy rate ↓Private mean SSCD ↓Public mean SSCD ↓
Full-Data SFT59.8670.9480.8%53.4653.27
Public-Only SFT22.8444.670.2%9.8644.56
Caption randomization + text noise60.671.310.3%33.055.3
Caption randomization only64.973.344.8%43.054.8
Ambient-inspired, σn=1\sigma_n=1, seed 055.066.81.7%20.9144.03
Ambient-inspired, σn=3\sigma_n=3, seed 063.977.114.2%26.939.4
Ambient-inspired, σn=3\sigma_n=3, seed 162.976.814.3%26.939.4
RSS69.5277.383.0%27.5741.41

The σn=3\sigma_n=3 Ambient-inspired configuration is the closest competitor: across two seeds, its DPG and private mean SSCD are close to RSS, but its GenEval is 5.6–6.6 points lower and its private copy rate is slightly higher. The σn=1\sigma_n=1 configuration reaches lower mean similarity and copying, but at a substantial cost in GenEval and DPG. Both caption interventions obtain lower generation quality and higher private copying than RSS.

Q3 — Performance-matched comparison and training cost. Comparisons at similar generation performance, with training cost reported, would materially affect my assessment of the contribution.

Training cost

We present the training cost as follows. Every listed run uses 8 GPUs and effective batch size 512. The table jointly reports the performance, memorization, and training cost of the corresponding 100K run:

Method, 50/50, 100KSeedGenEval ↑DPG ↑Private copy ↓Private SSCD ↓Wall hoursGPU-hours
Full-Data SFT059.8670.9480.8%53.4638.97311.7
RSS069.5277.383.0%27.5742.42339.3
Ambient-inspired, σn=1\sigma_n=1055.066.81.7%20.9126.88215.1
Ambient-inspired, σn=3\sigma_n=3063.977.114.2%26.926.60212.8
Ambient-inspired, σn=3\sigma_n=3162.976.814.3%26.927.10216.8
Caption randomization064.973.344.8%43.027.18217.5
Caption randomization + text noise060.671.310.3%33.026.82214.5

Relative to Full-Data SFT, RSS uses 27.6 additional GPU-hours (8.9%) but improves GenEval by 9.66 points and DPG by 6.44 points, while reducing private copying from 80.8% to 3.0% and private mean SSCD from 53.46 to 27.57. The σn=3\sigma_n=3 Ambient-inspired and caption baselines have lower recorded GPU-hours, but they also obtain lower generation quality; the Ambient-inspired runs are 5.6–6.6 GenEval points below RSS, while caption randomization has substantially higher private copying. The σn=1\sigma_n=1 configuration reaches lower copying and mean SSCD than RSS, but its GenEval and DPG are lower by 14.52 and 10.58 points.

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

This paper proposed a framework, called Representation-Space Supervision (RSS), to align the model prediction with their target model in the feature spaces, and mainly focus on the test-to-image doain in the supervised fine-tuning experiements.

Critical strengths
  • the method is simple, just add another loss about feature spaces to the model, which can be used in any pretrained models. It seems that we don't need much effort to implement the proposed framework.

  • the author conduct rich experimental results. All methods share some config like the initialization, optimizer and fixed partitions etc. From the experimental results, it seems that the propsoed RSS substantially improves the trade-off between generative capability and memorization in text-to-image diffusion models.

  • the paper gives useful baselines and ablations. Masked-pixel supervision, DP-SGD, pretraining-data mixing, RSS added on top of features-space loss, several several representation spaces and a pixel-space model give a broad first picture.

Critical weaknesses
  • The source code and data are not released, the reprodubility is unclear. For example, The SFT data is entirely synthetic, it is not possible to know if the experiments is solid and reliable. Also, It is unclear if the paper reports single runs with no seeds or confidence intervals.

  • the gain is modest when Against a properly early-stopped baseline, and in one split it reverses. The headline comparisons (Fig. 1 and the 100K-step numbers in §5.2) are made where the baselines have badly overfit. RSS has higher private similarity, how to explain this metric? for example: Pixel similarity barely changes across methods (about 73–76 on private data in Table 1), so it is not very clear.

  • The paper contrasts itself with data-side and sampling-side defenses (Somepalli et al. 2023b; Wen et al. 2024; Chen et al. 2024; 2025a) but never compares against them. The only baselines are masked-pixel loss and DP-SGD.

  • A key basline is missing: Public-Only SFT plus a comparable regularizer, such as RSS applied to the public images. That would show whether the private data contributes anything.

Questions for authors
  • The source code and data are not released, the reprodubility is unclear.

  • the gain is mild when compating to early-stopped baseline. also how I read the resutls for private similrity?

  • The paper contrasts itself with data-side and sampling-side defenses (Somepalli et al. 2023b; Wen et al. 2024; Chen et al. 2024; 2025a) but never compares against them. The only baselines are masked-pixel loss and DP-SGD.

  • A key basline is missing: Public-Only SFT plus a comparable regularizer, such as RSS applied to the public images. That would show whether the private data contributes anything.

5 comments

Thank you for the thoughtful and constructive review. We appreciate your recognition of the simplicity of RSS and the breadth of our experiments. Below, we address each concern and describe the additional analyses and experiments conducted in response.

Q1 — The source code and data are not released, and reproducibility is unclear. For example, the SFT data are entirely synthetic, making it difficult to assess whether the experiments are reliable.

A.1: Thank you for raising the reproducibility concern. We will release the complete training and evaluation code, exact data indices, preprocessing pipeline, public/private split construction, training configurations, checkpoint-selection rule, and memorization-evaluation scripts upon publication. We will also add these implementation details to the revised paper.

Second, our SFT dataset is not a proprietary collection generated specifically for this paper. It is a common practice to perform SFT with synthetic data [1] because they provide high-quality images and detailed captions. Moreover, our training data is constructed with publicly available image-text pairs from BLIP3o-60K [2], the LAION DALL-E 3 Discord dataset [3], and ShareGPT-4o-Image [4].

Consequently, our experiments do not depend on inaccessible proprietary data. The complete dataset composition can be independently reconstructed from the released sources, and we will provide the exact sample indices and processing scripts used in our experiments. Together with the released code, configurations, and fixed data partitions, this will make the complete experimental pipeline reproducible.

[1] Wang, Xianbang, et al. "Minit2i: A minimalist baseline for text-to-image generation, 2026." URL https://peppaking8. github. io/#/post/minit2i.

[2] Chen, Jiuhai, et al. "Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset." arXiv preprint arXiv:2505.09568 (2025).

[3] [OpenDatasets 2023] OpenDatasets. 2023. "DALL·E 3 Dataset (LAION Discord set)." https://huggingface.co/datasets/OpenDatasets/dalle-3-dataset.

[4] Chen, Junying, et al. "Sharegpt-4o-image: Aligning multimodal models with gpt-4o-level image generation." arXiv preprint arXiv:2506.18095 (2025).

Q2 — It is unclear if the paper reports single runs with no seeds or confidence intervals.

We thank the reviewer for raising this concern. We agree that the original submission did not clearly state the number of independent training runs. To evaluate run-to-run variation, we repeated the key RSS experiment under the 25/75 public/private split for 100K steps using a second independent training seed. The two runs produce the following results:

MetricTraining seed 0Training seed 1Absolute difference
GenEval ↑\uparrow66.9666.810.15
DPG ↑\uparrow74.8676.711.85
Private SSCD ↓\downarrow23.9824.300.32

The central performance--memorization behavior is consistent across the two independent training runs. In particular, GenEval differs by only 0.15 points and private-set SSCD by 0.32 points. DPG exhibits somewhat larger variation of 1.85 points, which we will report explicitly rather than hiding through aggregation.

Q3 — The gain is modest when Against a properly early-stopped baseline, and in one split it reverses:

A3: First, relative to the early-stopped Public-Only baselines, RSS consistently improves the best achievable GenEval score across all three public/private splits:

SplitFull-Data SFT: GenEval / private SSCDPublic-Only SFT: GenEval / private SSCDRSS: GenEval / private SSCD
50/5068.3 / 44.566.9 / 28.369.8 / 25.9
25/7568.3 / 44.163.7 / 26.368.7 / 23.8
10/9068.3 / 43.959.6 / 20.764.7 / 22.1

Compared with Public-Only SFT, RSS improves the best GenEval score by 2.9, 5.0, and 5.1 points under the 50/50, 25/75, and 10/90 splits, respectively. This demonstrates that the private data provide useful learning signals through RSS beyond simply discarding them.

Compared with early-stopped Full-Data SFT, RSS matches or exceeds its GenEval performance under the 50/50 and 25/75 splits, while reducing private-set SSCD by 18.6 and 20.3 points, respectively. Thus, RSS achieves comparable generative performance while substantially mitigating private-data memorization.

We acknowledge the reviewer's observation regarding the 10/90 split: early-stopped Full-Data SFT achieves a higher GenEval score than RSS (68.3 versus 64.7). This result reflects the intended information–memorization trade-off: when 90% of the data receive only representation-space supervision, RSS intentionally removes instance-specific information that can improve fitting but also causes memorization. Nevertheless, RSS reduces private-set SSCD from 43.9 to 22.1 relative to Full-Data SFT. It also improves GenEval by 5.1 points over Public-Only SFT, while its private-set SSCD remains close to the no-access reference (22.1 versus 20.7). RSS therefore occupies a substantially more favorable performance–memorization trade-off point.

The improvement in DPG is less pronounced. This behavior is partly due to the properties of our fine-tuning dataset and setup: even Full-Data SFT produces only limited improvement in DPG. Thus, this particular SFT dataset provides a stronger learning signal for the capabilities measured by GenEval than for those measured by DPG.

Finally, the 100K-step and early-stopping comparisons answer complementary questions. The 100K-step results provide an equal-compute comparison and measure training stability. The best-checkpoint results are an optimistic oracle analysis in which each method is retrospectively selected using its highest GenEval score. Even under this baseline-favorable comparison, RSS matches Full-Data SFT in two splits while substantially reducing private-data memorization, and consistently outperforms Public-Only SFT in GenEval.

Q4 — RSS has higher private similarity, how to explain this metric? How to read private similarity?

A.4: Private-set similarity measures training-example reproduction on the private dataset. Specifically, we sample 2,000 text-image pairs from the private training set. For each training prompt, we generate four images using independent initial noise and compare them with the corresponding private training image in pixel space and SSCD feature space. We then average over the four generations and the complete probe set. A lower value indicates less reproduction of the corresponding private training images.

This metric should be interpreted jointly with generation quality and relative to the information available to each method. First, the central controlled comparison is between RSS and Full-Data SFT, because both methods use the private data during training. Full-Data SFT directly supervises the model using the private images in data space, whereas RSS uses them only through representation-space supervision. Across all three splits, RSS substantially reduces private-set SSCD relative to Full-Data SFT. At the best-GenEval checkpoints, the values are:

SplitFull-Data SFTPublic-Only SFTRSS
50/5044.528.325.9
25/7544.126.323.8
10/9043.920.722.1

Thus, relative to direct supervision on the same private data, RSS reduces private-set SSCD by 18.6, 20.3, and 21.8 points, respectively.

Second, Public-Only SFT never accesses the private images and therefore serves as a no-access reference. At the best-GenEval checkpoints, RSS obtains even lower private similarity than Public-Only SFT under the 50/50 and 25/75 splits. Under the 10/90 split, RSS is only 1.4 points higher (22.1 versus 20.7), while improving GenEval by 5.1 points (64.7 versus 59.6). This shows that RSS extracts useful information from the private data while keeping reproduction close to the level of a model that never accesses those images.

The very low Public-Only similarity values at 100K steps should not be interpreted as stronger memorization mitigation. At that checkpoint, Public-Only SFT has severely degraded: its GenEval scores fall to 22.8, 6.9, and 9.9 across the three splits. A collapsed model naturally has low similarity because it no longer generates meaningful images corresponding to the prompts.


Q5 — Why pixel similarity changes little across methods

A.5: The private-set pixel similarity does change substantially across methods. The range of approximately 73–76 highlighted by the reviewer applies mainly to the base model, Public-Only SFT, and RSS. It does not include the Full-Data SFT baseline, which directly supervises the model on the private images and exhibits substantially higher pixel similarity.

At 100K steps, the private-set pixel similarities of Full-Data SFT are 93.4, 93.5, and 93.3 under the 50/50, 25/75, and 10/90 splits, respectively, whereas the corresponding RSS values are 76.4, 76.4, and 76.4. RSS therefore reduces private-set pixel similarity by approximately 17 points across all three splits.

The same conclusion holds at each method's best-GenEval checkpoint:

SplitFull-Data SFTPublic-Only SFTRSS
50/5083.274.575.2
25/7583.774.775.7
10/9083.276.972.9

Thus, even after early stopping, RSS reduces private-set pixel similarity by 8.0, 8.0, and 10.3 points relative to Full-Data SFT. These results support the conclusion that RSS substantially reduces private training-image reproduction relative to direct data-space supervision.

The repeated value of 76.4 for RSS at 100K steps is a rounding coincidence, not an evaluation error. The unrounded values from the three independently constructed private probe sets are 76.44, 76.39, and 76.45 respectively. We will report these values to two decimal places in the revision.

Pixel similarity is less discriminative among models that do not directly reproduce the training image: in our experiments, these models remain close to the pretrained model's pixel-similarity level of approximately 74. However, it clearly detects the strong increase caused by Full-Data SFT. SSCD, which is specifically designed for image-copy detection, provides finer discrimination within this non-copying regime. RSS's private-set SSCD values—27.6, 24.0, and 20.6—vary across splits while remaining substantially below those of Full-Data SFT. We therefore report both metrics and interpret them together with generation quality and qualitative reproduction examples.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.