acceptodds
Under review as a conference paper at ICLR 2027

Redistributive Unlearning for One-Step Generators via Unbalanced Optimal Transport

Abstract

Recent advances in one-step generative frameworks, such as flow map models, have significantly improved the efficiency of image generation by learning direct noise-to-data mappings in a single forward pass. However, machine unlearning for these powerful generators remains entirely unexplored, as existing diffusion unlearning methods rely on multi-step denoising and are inherently incompatible with one-step models. Moreover, prior methods focus only on suppressing the target concept, often leaving the unlearned probability mass to collapse into noise-like samples. In this work, we propose *UOT-Unlearn*, a redistributive class unlearning framework for one-step generative models based on the Unbalanced Optimal Transport (UOT). Our method formulates unlearning as a principled trade-off between a forget cost, which suppresses the target class, and an -divergence penalty, which confines the generated distribution to the data support via relaxed marginal constraints. As a result, the probability mass of the unlearned class is smoothly redistributed to the remaining classes rather than discarded, preserving overall generation fidelity. Experiments on CIFAR-10 and ImageNet-256 show that UOT-Unlearn achieves superior unlearning performance while maintaining generation quality, and it further extends to one-step text-to-image generation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, one-step generative models, flow map models, unbalanced optimal transport

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vPwpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

I am not an expert in this specific research area. I have made an effort to understand the method and evaluate the paper, but I still find some aspects difficult to assess confidently. I hope the authors and AC will also carefully consider the opinions of reviewers with more direct expertise in generative model unlearning and optimal transport.

This paper proposes UOT-Unlearn, which applies unbalanced optimal transport (UOT) to concept unlearning in pretrained one-step generative models. Compared with existing unlearning baselines, including GA, VDU, SA, and SalUn, the method achieves good unlearning performance measured by PUL while maintaining reasonable generation quality measured by u-FID. However, the paper is difficult to read, and the generation quality still degrades subs

Critical strengths
  1. The paper addresses the useful problem of unlearning concepts from one-step generative models.

  2. Compared with GA, VDU, SA, and SalUn, the proposed method achieves good unlearning performance while maintaining reasonable generation quality.

Critical weaknesses
  1. Generation quality degradation: Although the method achieves good PUL scores, generation quality still degrades substantially compared with the pretrained model. For example, on ImageNet-256, the u-FID of MeanFlow increases from 5.52 to 8.54 after unlearning. This degradation may limit the practical usefulness of the method.

  2. Unclear contribution of UOT: The paper needs to demonstrate whether the improvements genuinely come from the unbalanced optimal transport formulation, rather than mainly from the manually designed forgetting cost and class redirection strategy. The current results do not sufficiently isolate these contributions.

  3. Presentation and clarity: The paper presents a relatively intuitive goal through complex optimal transport formulations, making it difficult to understand the core idea and its specific contribution.

Questions for authors

Could the authors provide the following controlled comparisons, with other settings kept as consistent as possible?

  1. Full method: The proposed forgetting cost combined with UOT distribution constraints.

  2. Simpler regularization: The same forgetting cost combined with a simpler distribution-preserving regularization objective instead of UOT.

  3. Balanced OT: The same forgetting cost using standard OT with strict marginal constraints instead of UOT's relaxed marginal constraints. These comparisons would help establish whether the unbalanced nature of UOT is necessary for the reported improvements.

5 comments

[W1] Generation quality degradation: Although the method achieves good PUL scores, generation quality still degrades substantially compared with the pretrained model. For example, on ImageNet-256, the u-FID of MeanFlow increases from 5.52 to 8.54 after unlearning. This degradation may limit the practical usefulness of the method.

A. The increase from 5.52 to 8.54 is obtained at 98.75% PUL, i.e., near-complete erasure of the forget class, and it is the smallest degradation among all compared methods by a wide margin: the next best baseline on ImageNet-256 reaches u-FID 12.70 at only 61.29% PUL (SA), and the others exceed 49. Some loss of fidelity is inherent to unlearning, since the generator must leave the pretrained distribution on the forget region, and a trade-off between PUL and u-FID is observed for every method in Fig. 3. What distinguishes our method is the shape of this trade-off: the baselines' u-FID rises sharply once PUL becomes meaningful, whereas ours stays nearly flat, so a few FID points buy almost complete erasure. In practice, the practitioner can choose the operating point on this trade-off. A practitioner who prefers higher fidelity can simply stop earlier along the trajectory in Fig. 3. Even at such an earlier point, our method still achieves higher PUL than the baselines at the same u-FID.


[W2,Q] Could the authors provide the following controlled comparisons, with other settings kept as consistent as possible (Full method, Simpler regularization, Balanced OT)? These comparisons would help establish whether the unbalanced nature of UOT is necessary for the reported improvements.

A. Following the reviewer's suggestion, we added the requested controlled comparisons with all other settings kept fixed: (2) a simpler regularization variant that combines the same forgetting cost with a WGAN-GP objective (WGAN + culc_{\mathrm{ul}}; we also report WGAN w/o culc_{\mathrm{ul}} as a reference without the forgetting cost), and (3) a balanced OT variant that uses the same forgetting cost with standard OT under strict marginal constraints (OT + culc_{\mathrm{ul}}), alongside (1) our full method. The results, reported in Table 4 of the revised manuscript, show that the UOT formulation is necessary for the reported improvements: WGAN + culc_{\mathrm{ul}} retains moderate PUL but more than doubles u-FID (20.09 vs. 8.32), while balanced OT + culc_{\mathrm{ul}} loses about 30 points of PUL at comparable u-FID (52.76 vs. 84.32).

MethodMetricC0C1C2C3C4C5C6C7C8C9Avg
VDUPUL16.8153.8646.1013.7360.5136.4251.2538.6548.4332.2939.81
u-FID4.8271.9840.774.6124.8746.7047.1153.8152.0788.2143.50
WGAN w/o culc_{\mathrm{ul}}PUL38.6928.8348.8136.9840.7417.2042.4215.0646.1127.5834.24
u-FID7.166.405.736.036.006.629.426.277.565.666.69
OT + culc_{\mathrm{ul}}PUL51.5930.2253.5765.4362.0957.7659.3148.8159.2833.5352.76
u-FID8.278.026.156.037.526.296.867.867.1210.227.83
WGAN + culc_{\mathrm{ul}}PUL76.1166.6080.0079.2182.1481.1283.8849.2973.6557.1472.23
u-FID16.1215.2626.8817.4218.4038.2124.5616.4315.6715.6820.09
OursPUL80.7680.3283.1087.0993.2583.5590.9881.5085.2377.3884.32
u-FID9.949.908.348.498.299.785.118.405.889.038.32

We discuss why these three outcomes follow from the structure of the UOT objective in our response to [W3] below.

[W3] Presentation and clarity: The paper presents a relatively intuitive goal through complex optimal transport formulations, making it difficult to understand the core idea and its specific contribution.

A. We agree that the optimal transport formulations are complex, but we believe that UOT formulation is proper to mathematically state our core idea: unlearning requires the generated distribution to deviate from the pretrained one on the forget region while keeping the other modes unchanged. As optimal transport is well-established mathematical fields, most of the existing formulation follows the recent works.

Theoretically, the forgetting cost culc_{\mathrm{ul}} (Eq. 16) penalizes any generated sample that still looks like the forget class, pushing it out of the forget region SfS_f, which is our key idea and contribution of UOT compared to OT. Moreover, because the penalty is an ff-divergence, the displaced mass must be redistributed to the retain classes rather than to noise-like samples.

Empirically, the controlled comparisons in Table 4 (see [W2, Q1]) make the contribution of each part visible. With balanced OT, the marginal error is forced to zero, so PUL stays low even with the unlearning cost (52.76). With a simpler WGAN regularizer, the forgetting cost can raise PUL (72.23), but the regularizer lacks the support-confinement property, so the expelled mass is not redistributed to the retain classes and u-FID rises to 20.09. Only the forgetting cost combined with UOT achieves both high PUL (84.32) and low u-FID (8.32).

We have added the rationale of UOT and unlearning in the Section 3.2.

Dear Reviewer vPwp,

Thank you again for your constructive review. As the author–reviewer discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns.

On the contribution of UOT, we added the requested controlled comparisons with the same forgetting cost (Table 4): balanced OT loses about 30 points of PUL (52.76 vs. 84.32), and the simpler WGAN-GP regularizer more than doubles u-FID (20.09 vs. 8.32). On the quality degradation, we clarified that the u-FID increase from 5.52 to 8.54 on ImageNet-256 is obtained at 98.75% PUL and is the smallest among all compared methods, where the next best baseline reaches 12.70 at only 61.29% PUL. On the clarity, we added the rationale for the UOT formulation to Section 3.2. We have also revised the paper accordingly. We would be happy to clarify any remaining questions.

Reviewer_JsyXOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes UOT-Unlearn, an unlearning method for one-step generative models based on unbalanced optimal transport. It combines a feature based forgetting cost with distribution regularization to redirect generations toward retained concepts while preserving image quality. The method uses generated samples during fine-tuning, and uses a subset of real samples from the forgetting class to compute feature centroid for the class-unlearning objective. Experiments on CIFAR-10, ImageNet, and Van Gogh style removal using SDXL-Turbo shows improvements over the evaluated baselines.

Critical strengths
  1. Preserving generation quality while removing unwanted concepts is a relevant problem for the current VLMs or text to image models, and the paper provides an approach for addressing this for single step generator models.
  2. The paper evaluates multiple onestep generators, including CTM , MeanFlow, and SDXL-turbo (text-to-image model) and reports improvements in forgetting performance and u-FID.
Critical weaknesses

It would be helpful for readers if the following points were clarified or improved:

  1. The evaluation scope is limited. Although several classes are evaluated independently, each experiment removes a single target. The text-to-image evaluation considers only Van Gogh style removal on SDXL-Turbo. Experiments on simultaneous or a set of concept removal, compositional prompts, and another recent text-to-image backbone would help establish efficacy and applicability as done in earlier paper of unlearning like Ablating Concepts in Text-to-Image Diffusion Models or SPEED: Scalable, Precise, and Efficient Concept Erasure for Diffusion Models.

  2. Maybe the baseline adaptations need stronger validation, as some of the baseline approaches show poor fid scores >50 in table2.

  3. The data requirements may be presented in more precisely. “Once” in Table 1 still means that real forget images are required during preprocessing to compute the anchor features. Maybe it would be clearer to write “Required for preprocessing only”.

  4. It would be helpful if the paper clarifies why existing concept-erasure methods cannot be adapted to one-step generators. Methods such as Unified Concept Editing in Diffusion Models (UCE) or SPEED. Could the authors compare against suitable adaptations.

  5. It would also be an interesting discussion for the readers to analyze or compare against the alternative of applying an existing concept erasure method to multistep generator and then distilling this model into a one-step generator.

Questions for authors

Please refer to Weakness section for questions.

6 comments

[W1] The evaluation scope is limited. Although several classes are evaluated independently, each experiment removes a single target. The text-to-image evaluation considers only Van Gogh style removal on SDXL-Turbo. Experiments on simultaneous or a set of concept removal, compositional prompts, and another recent text-to-image backbone would help establish efficacy and applicability as done in earlier paper of unlearning like Ablating Concepts in Text-to-Image Diffusion Models or SPEED: Scalable, Precise, and Efficient Concept Erasure for Diffusion Models.

A. We thank the reviewer for this suggestion. In the rebuttal, in addition to style erasure (Van Gogh), we have added (1) instance erasure and (2) multi-concept erasure experiments, as well as (3) on another one-step text-to-image backbone (SD-Turbo). Following the reviewer's suggestion, We also include the two training-free editing methods, UCE and SPEED. Following SPEED, we now evaluate preservation on COCO-1K and additionally report FID against the pretrained model's generations. The new results on SDXL-Turbo are as follows:

MethodSnoopy CLIP$_f$↓Snoopy CLIP$_p$↑Snoopy FID↓Van Gogh CLIP$_f$↓Van Gogh CLIP$_p$↑Van Gogh FID↓Snoopy+Van Gogh CLIP$_f$↓Snoopy+Van Gogh CLIP$_p$↑Snoopy+Van Gogh FID↓
SDXL-Turbo28.5526.57–27.3926.57–27.9726.57–
GA27.7126.5624.7225.5326.5118.2128.8026.5324.00
SalUn20.9926.5419.6823.7726.537.8922.8626.6924.08
UCE24.7126.6011.7126.1126.609.6525.4726.5913.55
SPEED23.7626.565.0226.1426.573.0824.9526.545.92
Ours21.3726.6213.5520.7927.1947.6220.9126.6113.94

(1) Instance Erasure. Following the common setup in the papers mentioned by the reviewer, we consider instance erasure of Snoopy. As shown in snoopy columns, our method achieves second-lowest CLIP$_f$ (21.37), while attaining the highest CLIP$_p$ (26.62) without a substantial increase in FID. In Qualitative results, Our method replaces Snoopy with a realistic dog while preserving the scene. We refer the reviewer to Figure 4 in the main paper and Figure 12 in the Appendix for qualitative results.

(2) Multi-concept erasure. We erase Snoopy and Van Gogh simultaneously with a single model. Our objective extends directly: each forget prompt cfc_f of concept kk uses the forget anchor μf(k)\mu_f^{(k)} and margin mkm_k of its own concept, so the forget region becomes the union of the per-concept regions, Rf=⋃k{x:dcos⁡(k)(x)<mk}\mathcal{R}_f=\bigcup_{k}\{x : d_{\cos}^{(k)}(x)<m_k\}. On Snoopy$+VanGogh,ourmethodachievesthelowestCLIPVan Gogh, our method achieves the lowest CLIP_f$ (20.91) while maintaining CLIP$_p$ comparable to the pre-trained model, without a substantial increase in FID. In Qualitative results, our method erases both the instance and the style. We refer the reviewer to Figure 4 in the main paper and Figure 14 in the Appendix for qualitative results.

(3) Another one-step T2I backbone. We repeat all T2I experiments on SD-Turbo, which differs from SDXL-Turbo in both its base model (Stable Diffusion 2.1) and text encoder, using the same prompts, evaluation protocol. The results on SD-Turbo are as follows:

MethodSnoopy CLIP$_f$↓Snoopy CLIP$_p$↑Snoopy FID↓Van Gogh CLIP$_f$↓Van Gogh CLIP$_p$↑Van Gogh FID↓Snoopy+Van Gogh CLIP$_f$↓Snoopy+Van Gogh CLIP$_p$↑Snoopy+Van Gogh FID↓
SD-Turbo28.4626.10–29.3426.10–28.9026.10–
GA26.3726.5740.8829.6226.4233.3828.8626.4534.94
SalUn20.1726.1919.2823.6526.137.6521.9426.1520.09
UCE25.4226.0019.5926.8826.1417.8426.0426.0120.82
SPEED26.5026.115.8527.4426.087.8226.9026.0914.82
Ours21.0726.1014.0223.0626.0420.4524.1826.0413.49

Consistent with the results on SDXL-Turbo, our method achieves strong forgetting on SD-Turbo while maintaining CLIP$_p$ and lower FID. Please see Figure 12-14 in the Appendix for qualitative results on SD-Turbo.

[W2] Maybe the baseline adaptations need stronger validation, as some of the baseline approaches show poor fid scores >50 in table2.

A. We totally agree with the reviewer about weak baselines. First, we want to emphasize that none of previous works investigate the unlearning formulation for one-step generators, which performs poor when fine-tuning with existing multi-step diffusion unlearning objectives. Thus, the high u-FID values of baselines are not a result of our adaptations but of the trade-off they exhibit once they reach a decent level of unlearning. Fig. 3 shows that u-FID rises gradually as unlearning proceeds for every method, so each method traces a PUL–u-FID curve rather than a single point. The baselines exhibit a sharp deterioration in u-FID as PUL increases: at the PUL levels reported in Table 2, their u-FID has already exceeded 50. Reporting the baselines at lower PUL would reduce their u-FID, but only by giving up the unlearning they are meant to achieve.

We would also like to note that the reported numbers are the best trade-off obtained under a hyperparameter search, not default settings. For each baseline we searched α∈{0.05,0.1}\alpha \in \{0.05, 0.1\}, β∈{0.5,5.0}\beta \in \{0.5, 5.0\}, λSA\lambda_{\mathrm{SA}}, γ\gamma, and the saliency ratio, and report the best PUL/u-FID trade-off. To further validate the baselines, we added WGAN-GP and OT-based comparisons in Table 4, which share our unlearning cost and differ only in the regularizer.


[W3] The data requirements may be presented in more precisely. “Once” in Table 1 still means that real forget images are required during preprocessing to compute the anchor features. Maybe it would be clearer to write “Required for preprocessing only”.

A. We thank the reviewer for the careful comment. Following the suggestion, we revised the entry for our method in Table 1 to "Preprocess". The revised table and caption are highlighted in blue.


[W4] It would be helpful if the paper clarifies why existing concept-erasure methods cannot be adapted to one-step generators. Methods such as Unified Concept Editing in Diffusion Models (UCE) or SPEED. Could the authors compare against suitable adaptations.

A. Thank you for suggesting comparison methods that we did not cover. After reading those papers, we agree that training-free editing methods such as UCE and SPEED can be directly applied to one-step generators: they modify only the cross-attention projection weighs in closed form, without any sampling or gradient updates, so this methods are independent of the number of inference steps. Following the reviewer's suggestion, we added both methods as baselines on SDXL-Turbo and SD-Turbo, even though their approaches are not exactly same.

Thsese methods erase instances reasonably well, e.g., on SDXL-Tubo they lower the Snoopy Clip$_f$ from 28.55 to 24.71 (UCE) and 23.76 (SPEED) while preserving generation capability. However, they barely remove the artistic style of Van Gogh: Clip$_f$ decreases only from 27.39 to 26.11 (UCE) and 26.14 (SPEED), and the Van Gogh style remains clearly visible in the generated images. This agree with the orginal papaers, where style erasure also yields only small changes; for instance, SPEED reports a Van Gogh ClIP score of 26.29 against 28.75 for SD v1.4, and UCE 28.10. In contrast, our training-based method erases the Van Gogh Style effectively as well (Clip$_f$ 20.79). We attribute this difference to their method design: the training-free edit only remaps the embedding of the concept token, which suits concepts tied to a specific word, such as an instance, but is limitied for more complex concepts such as artistic styles.


[W5] It would also be an interesting discussion for the readers to analyze or compare against the alternative of applying an existing concept erasure method to multistep generator and then distilling this model into a one-step generator.

A. We thank the reviewer for this interesting suggestion. Applying an unlearning method to a multi-step generator and then distilling it into a one-step generator is indeed a natural alternative. However, this requires re-running the full distillation for every concept to be erased, which is highly time-consuming and computationally expensive. In contrast, our method directly applys on the one-step generator with a lightweight LoRA update.

Dear Reviewer JsyX,

Thank you again for your constructive review. As the author–reviewer discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns.

On the evaluation scope, we added instance erasure (Snoopy), multi-concept erasure (Snoopy + Van Gogh), and a second one-step backbone (SD-Turbo); on SDXL-Turbo, our method attains the lowest CLIP$_f$ on Van Gogh (20.79) and on the multi-concept task (20.91). On the baseline validation, we clarified that the high u-FID values in Table 2 reflect the PUL–u-FID trade-off of the baselines (Fig. 3) rather than our adaptations, and that the reported numbers are the best trade-off found by a hyperparameter search. We also added controlled WGAN-GP and OT baselines that share our unlearning cost (Table 4). We have also revised the paper accordingly. We would be happy to clarify any remaining questions.

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

The manuscript proposes UOT-unlearn which fine tunes a one step image generator to move the generator away from producing an unwanted class while replacing those generations with plausible images. This is achieved by redistributing the probability associated with the unwanted class to plausible alternatives.

The target is to make: $ \begin{equation*} Pr(G_{\theta}(z) \in S_f) \rightarrow 0 \quad \text{,} \quad Pr(G_{\theta}(z) \in S_r) \rightarrow 1 \end{equation*} $

where SfS_f denotes the semantic region containing the forget class, and SrS_r contains the remaining class.

The authors assume that the real retain class images are unavailable during fine tuning. They approach the problem by treating the unlearning process as distribution transport achieved using Unbalanced Optimal Transport (UOT). UOT is used to avoid exact marginal mapping, which can result in reproducing the original data distribution again.

Since original retain data is unavailable, they use the frozen original generator to supply reference samples.

To identify the unwanted concept, they first extract the features of the forget class images using a frozen pretrained feature extractor. Then, they compute the centroid of the forget class images. After that, the cosine differences are calculated between the current images and the computed centroid. If the cosine distance is below a threshold (mm), the current image is marked as forget-like.

Then, a cost that changes according to the current output is defined. Minimising this increases the cosine difference between the generated image and the forget class centroid when the current image is treated as forget-like, and it encourages the generated output remain closer to the original output when the generated image lies outside the forget class.

To discourage producing unrealistic images, the authors penalise the generator if it produced probabilities where the reference distribution has a low probability. This is achieved using KL divergence.

The training alternates between optimising a dual potential network through UOT semi dual objective and fine tuning the one step generator. The potential provides distributional guidance using the current outputs and the samples from the pretrained generator. The generator pushes forget-like outputs away from the forget centroid and limits displacement outside the forget region.

Validation includes 2D synthetic example, CIFAR-10 experiments with CTM and MeanFlow, class conditional ImageNet-256 experiments with MeanFlow, and Van Gogh style removal using SDXL-Turbo. Comparisons are presented against adapted unlearning baselines evaluate forgetting and generation quality through PUL and u-FID. Text-to-image experiments use CLIP-based metrics for style suppression and prompt alignement. Ablations are performed for forgetting strength, feature space margin, centroid sample size, and marginal relaxation.

Critical strengths
  1. A formulation suited to one-step generators under restricted data access.
  • The method combines feature-based forgetting derived using UOT objective operating directly on generated outputs. Assuming real retain samples are not available, the method uses synthetic reference samples or intermediate denoising trajectories. This addressed practically relevant setting preserving one step inference.
  1. Strong reported forgetting–fidelity trade-offs -On CIFAR-10 with CTM, the average PUL is 85.51%. The u-FID value is 6.96, compared with a baseline average of at least 42.02. Experiments with CTM and MeanFlow, including ImageNet-256, provide evidence beyond a single model or dataset.
Critical weaknesses
  1. The comparisons do not cleanly establish the benefit of the UOT formulation.
  • The unconditional SA and SalUn adaptations explicitly train forget-associated inputs to produce Gaussian noise (Eq. 32). Their poorer generation quality may therefore partly reflect this target choice.
  • The WGAN comparison in Appendix F.2 removes both the UOT formulation and the proposed unlearning cost. Consequently, its lower forgetting performance cannot establish which component produces the improvement. Table 6 isolates source-marginal relaxation, but does not resolve this broader attribution issue.
  • A comparison retaining the same unlearning cost while changing only the distributional regularizer would better substantiate the paper’s central methodological claim.
  1. The theoretical analysis does not establish the claimed redistribution guarantee for the implemented generator.
  • The main suppression bound concerns a potentially unnormalized UOT marginal. Appendix B provides a normalized guarantee under the additional condition (C_0+\delta<1), whose applicability to the experiments is not established.
  • The practical method also replaces the true data distribution with the pretrained distribution, and the paper does not establish that the trained generator’s output distribution realizes the normalized optimal UOT marginal.
  • These gaps limit how strongly the ideal transport result supports the claim that the implemented method redistributes forget probability into valid retain classes. This matters because principled redistribution is a central contribution of the paper.
Questions for authors

Clarification—connection between theory and implementation: How does the generator distribution GθμZG_{\theta}\mu_Z relate to the normalized optimal UOT marginal analyzed in Appendix B? Is the condition C0+δ<1C_0+\delta<1 established for the evaluated settings, and which guarantees remain applicable when pdatap_{\mathrm{data}} is replaced by pprep_{\mathrm{pre}}?

9 comments

[W1-a] The unconditional SA and SalUn adaptations explicitly train forget-associated inputs to produce Gaussian noise (Eq. 38). Their poorer generation quality may therefore partly reflect this target choice.

A. We agree that the target choice affects generation quality, and that it explains part of the quality gap on CIFAR-10. We would like to clarify two points.

The noise target follows the original SA design rather than our adaptation. For class forgetting, SA sets the surrogates q(x∣cf)q(x\mid c_f) to a uniform distribution over pixel values, and Eq.38 keeps this surrogate with Gaussian noise in its place. SalUn instead relabels the forget concept to a misaligned concept c′≠cc'\neq c, but relabeling presupposes a conditioning input, which an unconditional generator does not have. In the unconditional setting, we therefore adopted the SA surrogate as a common target for both baselines.

On class-conditional ImageNet-256, the target design differs accordingly: SA maps cfc_f to a retain class, while SalUn uses its original random relabeling. Consistent with the reviewer's observation, SA achieves a substantially lower u-FID of 12.70 in this setting, compared with 51.61 on unconditional CIFAR-10. Although these results come from separately trained models on different datasets, they suggest that the choice of surrogate target may influence generation quality. But the ranking does not change: SA and SalUn reach PUL 61.29 and 57.62 versus 98.75 for ours, and our u-FID remains the lowest (8.54 vs. 12.70 and 49.87). Per class, our PUL is higher than SA’s on all five forget classes, and our u-FID is lower on four of the five (Table 2).


[W1-b,c] A comparison retaining the same unlearning cost while changing only the distributional regularizer would better substantiate the paper’s central methodological claim.

A. Following the reviewer's suggestion, we added three baselines that isolate the effect of the distributional regularizer: WGAN-GP with and without the unlearning cost (WGAN + culc_{\mathrm{ul}} and WGAN w/o culc_{\mathrm{ul}}), and an OT formulation that retains the same unlearning cost but replaces UOT with a linear-potential OT objective (OT + culc_{\mathrm{ul}}). Specifically, experimental settings are as follows. Each baseline differs from ours in exactly one component: WGAN w/o culc_{\mathrm{ul}} replaces the UOT objective with WGAN-GP and uses classifier-selected pseudo-retain samples without the unlearning cost; WGAN + culc_{\mathrm{ul}} adds the same unlearning cost to the WGAN-GP objective; and OT + culc_{\mathrm{ul}} keeps our setup unchanged except for the linear potential Ψi∗(x)=x\Psi_i^*(x) = x, i.e., exact marginal matching. Full implementation details are given in Appendix G.2.

MethodMetricC0C1C2C3C4C5C6C7C8C9Avg
VDUPUL16.8153.8646.1013.7360.5136.4251.2538.6548.4332.2939.81
u-FID4.8271.9840.774.6124.8746.7047.1153.8152.0788.2143.50
WGAN w/o culc_{\mathrm{ul}}PUL38.6928.8348.8136.9840.7417.2042.4215.0646.1127.5834.24
u-FID7.166.405.736.036.006.629.426.277.565.666.69
OT + culc_{\mathrm{ul}}PUL51.5930.2253.5765.4362.0957.7659.3148.8159.2833.5352.76
u-FID8.278.026.156.037.526.296.867.867.1210.227.83
WGAN + culc_{\mathrm{ul}}PUL76.1166.6080.0079.2182.1481.1283.8849.2973.6557.1472.23
u-FID16.1215.2626.8817.4218.4038.2124.5616.4315.6715.6820.09
OursPUL80.7680.3283.1087.0993.2583.5590.9881.5085.2377.3884.32
u-FID9.949.908.348.498.299.785.118.405.889.038.32

The results, reported in Table 4 of the revised manuscript, show that neither alternative regularizer matches UOT under the same unlearning cost: OT + culc_{\mathrm{ul}} loses about 30 points of PUL at comparable u-FID (52.76 vs. 84.32), while WGAN + culc_{\mathrm{ul}} retains moderate PUL but more than doubles u-FID (20.09 vs. 8.32). The gap between OT + culc_{\mathrm{ul}} and ours shows that the relaxed marginal constraint is what allows the forget mass to be redistributed, and the gap between WGAN + culc_{\mathrm{ul}} and ours shows that the UOT regularizer preserves retain-class fidelity where a generic adversarial regularizer does not.

[Q1, W2] How does the generator distribution (Gθ)♯μZ(G_\theta)_\sharp\mu_Z relate to the normalized optimal UOT marginal analyzed in Appendix B?

A. We appreciate the reviewer's insightful comment. The generator distribution (Gθ)♯μZ(G_\theta)_\sharp\mu_Z and the normalized optimal UOT marginal πˉ1λ\bar\pi_1^\lambda are pushforwards of two nearby source measures under the same optimal map TT. The two differ in general, but their W2W_2 distance is bounded by the source divergence DΨ1(π0∣μ)≤C0+δD_{\Psi_1}(\pi_0|\mu)\le C_0+\delta, and both guarantees of Prop. 2 transfer to the generator distribution.

First, the gap W22(T♯μ,πˉ1λ)W_2^2(T_\sharp\mu, \bar\pi_1^\lambda) is controlled by the source divergence that the UOT objective already bounds. Here, let μ=ppre\mu = p_{\mathrm{pre}} denote the source of the UOT problem (Eq. 8), so that (Gθ)♯μZ=T♯μ(G_\theta)_\sharp\mu_Z = T_\sharp\mu by the identification Gθ=T∘GpreG_\theta = T\circ G_{\mathrm{pre}} in Sec 3.2. Combining Pinsker's inequality and the data-processing inequality with the trivial bound W22≤DW1≤D2TVW_2^2 \le D W_1 \le D^2 \mathrm{TV} on a domain of diameter DD, we have

W22(T♯μ,πˉ1λ)≤D212KL(πˉ1λ∣T♯μ)≤D212KL(πˉ0λ∣μ)≤D2C0+δ2Mλ,W_2^2(T_\sharp\mu, \bar\pi_1^\lambda) \le D^2\sqrt{\frac{1}{2}\mathrm{KL}(\bar\pi_1^\lambda | T_\sharp\mu)} \le D^2\sqrt{\frac{1}{2}\mathrm{KL}(\bar\pi_0^\lambda | \mu)} \le D^2\sqrt{\frac{C_0+\delta}{2M_\lambda}},

where r0=dπ0/dμr_0 = d\pi_0/d\mu and Mλ=π0(X)M_\lambda = \pi_0(\mathcal{X}) denote the density and total mass of the source marginal, so that πˉ0λ=r0μ/Mλ\bar\pi_0^\lambda = r_0\mu/M_\lambda and πˉ1λ=T♯πˉ0λ\bar\pi_1^\lambda = T_\sharp\bar\pi_0^\lambda. Therefore, the two distributions are close whenever the source marginal π0\pi_0 stays close to μ\mu, which is exactly what the source penalty DΨ1(π0∣μ)D_{\Psi_1}(\pi_0|\mu) in the UOT objective enforces.

Second, regardless of how close π0\pi_0 is to μ\mu, the support confinement (Prop. 2(ii)), which is the key argument for unlearning, holds for the generator distribution. Since we use the KL divergence for the source penalty DΨ1D_{\Psi_1}, the source density r0r_0 is strictly positive μ\mu-a.e., so π0\pi_0 and μ\mu are mutually absolutely continuous. Hence, for any Borel AA,

ν(A)=0⇒π1(A)=π0(T−1A)=0⇒μ(T−1A)=0⇒T♯μ(A)=0,\nu(A) = 0 \Rightarrow \pi_1(A) = \pi_0(T^{-1}A) = 0 \Rightarrow \mu(T^{-1}A) = 0 \Rightarrow T_\sharp\mu(A) = 0,

i.e., T♯μ≪νT_\sharp\mu \ll \nu and T♯μT_\sharp\mu is concentrated on Sf∪SrS_f\cup S_r. In other words, the generator distribution never leaves the support of the data distribution ν\nu.

Moreover, we extend the forget suppression bound (Prop. 2(iii)) to the generator distribution under a bounded source density, r0≥rmin⁡>0r_0\ge r_{\min}>0 μ\mu-a.e. (Prop. 3). Under this assumption, since T♯μ(Sf)=μ(T−1Sf)≤π1(Sf)/rmin⁡T_\sharp\mu(S_f)=\mu(T^{-1}S_f)\le\pi_1(S_f)/r_{\min}, Prop. 2(iii) gives

T♯μ(Sf)≤C0+δrmin⁡ε⋅1λ,T♯μ(Sr)≥1−C0+δrmin⁡ε⋅1λ.T_\sharp\mu(S_f)\le\frac{C_0+\delta}{r_{\min}\varepsilon}\cdot\frac{1}{\lambda}, \qquad T_\sharp\mu(S_r)\ge 1-\frac{C_0+\delta}{r_{\min}\varepsilon}\cdot\frac{1}{\lambda}.

Note that T♯μT_\sharp\mu is already a probability measure, so the normalization step of Prop. 2(iv)-(v) and the condition C0+δ<1C_0+\delta<1 are not needed here. We have incorporated these results into Appendix B.3 (highlighted in blue), and we thank the reviewer for the comment that led us to extend our analysis to the generator distribution. The formal statement of Proposition 3 is as follows.

Proposition 3 (Guarantees for the generator distribution). Let Assumptions 1-3 hold, let δ≥0\delta \ge 0, and let πλ=(id,T)♯π0λ\pi^\lambda = (\mathrm{id}, T)_\sharp\pi_0^\lambda be a δ\delta-approximate minimizer of FλF_\lambda induced by a map TT, with source density r0=dπ0λ/dμr_0 = d\pi_0^\lambda/d\mu and total mass Mλ=π0λ(X)M_\lambda = \pi_0^\lambda(\mathcal{X}). Write πˉ1λ=π1λ/Mλ\bar\pi_1^\lambda = \pi_1^\lambda/M_\lambda and suppose diam(X)≤D\mathrm{diam}(\mathcal{X})\le D. Then: (i) W22(T♯μ,πˉ1λ)≤D2C0+δ2MλW_2^2(T_\sharp\mu, \bar\pi_1^\lambda) \le D^2\sqrt{\frac{C_0+\delta}{2M_\lambda}}. (ii) If r0>0r_0 > 0 μ\mu-a.e. (in particular, if DΨ1D_{\Psi_1} is the KL divergence), then T♯μ≪νT_\sharp\mu \ll \nu; in particular, T♯μT_\sharp\mu is concentrated on Sf∪SrS_f \cup S_r. (iii) If moreover r0≥rmin⁡>0r_0 \ge r_{\min} > 0 μ\mu-a.e., then T♯μ(Sf)≤C0+δrmin⁡ε⋅1λT_\sharp\mu(S_f)\le\frac{C_0+\delta}{r_{\min}\varepsilon}\cdot\frac{1}{\lambda} and T♯μ(Sr)≥1−C0+δrmin⁡ε⋅1λT_\sharp\mu(S_r)\ge 1-\frac{C_0+\delta}{r_{\min}\varepsilon}\cdot\frac{1}{\lambda}.

[Q2, W2] Is the condition C0+δ<1C_{0} + \delta < 1 established for the evaluated settings?

A. Yes. On the 2D synthetic experiment of Fig. 1, we obtain C0=0.404C_0 = 0.404 and C0+δ^≤0.420C_0 + \widehat{\delta} \le 0.420 across ten independent runs, so the condition holds with a wide margin. Here, C0C_{0} is computable from the experimental setting and hyperparameter τ\tau, and δ\delta is the primal-dual gap of the trained network pair (Gθ,vϕ)(G_{\theta}, v_{\phi}). Since evaluating δ\delta requires the exact UOT plan, we measured both terms on the 2D synthetic dataset: δ^\widehat{\delta} is estimated as the gap between the empirical UOT optimum and the training dual objective in Eq. (13) over ten independent sample sets. In all runs, C0+δ^C_0 + \widehat{\delta} remained well below 11 (mean: 0.4120.412, max: 0.4200.420), demonstrating that the sufficient condition is satisfied in this experiment.

The condition can also be satisfied on the image benchmarks. Recall that C0=τW22(ppre,νr)+log⁡11−pfC_0 = \tau W_2^2(p_{\mathrm{pre}}, \nu_r) + \log\frac{1}{1-p_f} (Eq. 22). The forget ratio pfp_f is small by the nature of class unlearning (one class out of 10 or 1,000), so the second term is close to zero. τ\tau is the weight of the retain cost (Eq. 16) and is a hyperparameter under our control, so C0C_0 can always be kept well below 1. Regarding δ\delta, a pair (Gθ,vϕ)(G_\theta, v_\phi), that is δ\delta-optimal for the semi-dual objective (Eq. 13), induces a plan that is δ\delta-optimal for FλF_\lambda by strong duality of UOT. It is therefore controlled by the approximation quality of the dual potential and decreases as vϕv_\phi is optimized.


[Q3, W2] Which guarantees remain applicable when pdatap_{data} is replaced by pprep_{pre}?

A. All of the guarantees remain applicable. The only difference is that the reference distribution changes from the ground-truth data distribution pdatap_{data} to the pretrained generator distribution pprep_{pre}. Appendix B is stated for a generic target ν\nu, so all steps of the proof remain valid when ν=ppre\nu = p_{\mathrm{pre}}. Specifically, the support-confinement statement, Prop. 2(ii), becomes π1λ≪ppre\pi_1^\lambda \ll p_{pre}: the fine-tuned distribution is confined to the support of pprep_{pre}. The forget-suppression bounds, Prop. 2(iii) and (v), are unaffected, since they are stated in terms of the forget support SfS_f, not of pdatap_{data}.

Whether ν\nu is pdatap_{\mathrm{data}} or pprep_{\mathrm{pre}} depends on how the unlearning task is defined. In this work, we consider a more challenging setting, where no real retain data are accessible (Table 1). Since the data distribution is unavailable in this setting, the natural unlearning goal is to redistribute the forget mass onto the support of pprep_{\mathrm{pre}} rather than onto that of pdatap_{\mathrm{data}}. This is exactly what our guarantees deliver. We also note that our framework extends directly to the case where real data are available: it suffices to replace the pretrained mini-batch B2\mathcal{B}_2 in Line 2 of Algorithm 1 with a mini-batch of real data, in which case the same analysis applies with ν=pdata\nu = p_{\mathrm{data}}.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.