acceptodds
Under review as a conference paper at ICLR 2027

Towards Bounded Continuous Concepts Erasure in Text-to-image Diffusion Models

Abstract

Text-to-image (T2I) diffusion models excel at generating high-fidelity images, yet they raise significant concerns regarding copyright infringement and unsafe content. Concept erasure mitigates these risks by removing specific semantics from a pretrained model. In real-world deployments, however, erasure requests often arrive sequentially rather than being known in advance. This continuous concept erasure setting requires a model to suppress each new target, retain all previous erasures, and preserve its generation ability on non-target concepts. Existing methods, primarily designed for isolated edits, struggle with the accumulated drift and historical erasure degradation caused by repeated updates. To address these challenges, we propose Lyapunov-constrained Continuous Concept Erasure (LCCE). LCCE formulates continuous erasure as an online control problem, using a dynamic virtual queue to accumulate preservation-budget violations and adaptively regulate subsequent edits. A historical retention loss further maintains the erased states established for previous targets. To provide the queue with a consistent preservation signal, LCCE identifies a semantic neighborhood for each target and incrementally consolidates these neighborhoods into a cumulative preservation bank. The resulting per-stage objective admits a closed-form solution, enabling efficient sequential editing without iterative fine-tuning. Comprehensive experiments demonstrate that LCCE outperforms state-of-the-art baselines. Even after sequentially erasing 100 artist concepts, LCCE reliably suppresses the target styles while maintaining high generative fidelity on non-target concepts. Our code will be released upon acceptance.

Primary area:
interpretability and explainable AI
Keywords:
continuous concept erasure, text-to-image diffusion models, Lyapunov optimization

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

This paper proposes Lyapunov-constrained Continuous Concept Erasure (LCCE), a framework for sequential concept erasure in text-to-image diffusion models. The method aims to address two challenges arising from repeated concept erasure: accumulated degradation of non-target generation capabilities and the reappearance of previously erased concepts. Specifically, LCCE introduces a dynamic virtual queue based on Lyapunov optimization to regulate long-term preservation loss, a cumulative preservation bank to maintain semantically related non-target concepts, and a historical retention loss to prevent concept revival. These components are integrated into a closed-form solution for updating cross-attention projections without iterative fine-tuning. Experiments on Stable Diffusion v1.4 demonstrate competitive performance in continuous artist-style, object, and celebrity erasure, with particularly low per-concept editing costs.

Critical strengths
  1. Meaningful problem formulation. The paper addresses continuous concept erasure, a practical and important setting where new erasure requests arrive sequentially while previously erased concepts and non-target generation capabilities must be preserved.

  2. Well-structured methodology. The proposed framework combines a dynamic virtual queue, a cumulative preservation bank, and a historical retention loss to address different challenges in continuous concept erasure.

  3. Promising experimental results. LCCE demonstrates competitive erasure effectiveness and non-target preservation across artist-style, object, and celebrity erasure tasks, particularly over long editing sequences.

Critical weaknesses
  1. Concerns about the fairness and completeness of baseline comparisons. This is my major concern. While continuous concept erasure is a meaningful practical problem, several existing methods can efficiently erase all accumulated concepts from scratch whenever a new erasure request arrives, without incurring substantial computational overhead. For example, UCE and SPEED can jointly erase multiple concepts within a short time, while MACE trains concept-specific LoRAs that can be efficiently merged using a closed-form solution. Therefore, these methods are naturally applicable to continuous erasure scenarios, and their computational costs may remain competitive even when newly arriving concepts trigger a fresh multi-concept erasure. I strongly encourage the authors to evaluate these baselines under such a setting and compare both their erasure effectiveness and cumulative computational costs. In particular, comparisons on 100 artistic styles and 100 celebrities would be valuable, given that SPEED and MACE were specifically designed for multi-concept erasure.

  2. Insufficient justification of the bounded preservation guarantee. The theoretical analysis in Appendix A establishes that the long-term preservation constraint holds under mean-rate stability of the virtual queue. However, it does not establish that the proposed editing algorithm necessarily satisfies this stability condition. The authors should clarify the assumptions required for the claimed guarantee and provide empirical evidence of queue stability and preservation-budget satisfaction over long erasure sequences.

  3. Unclear necessity of Lyapunov-based optimization. The dynamic virtual queue essentially adjusts the weight of the preservation objective across editing stages. Although Table 4 includes an ablation without the dynamic queue, it remains unclear whether Lyapunov optimization offers meaningful advantages over simpler alternatives, such as well-tuned fixed preservation weights or heuristic adaptive weighting. More controlled comparisons are needed to establish the necessity of this design.

  4. Incomplete evaluation of long-term erasure effectiveness. The 100-artist experiment in Figure 3 primarily reports CLIP and VQA scores without showing whether previously erased concepts remain suppressed. Moreover, LCCE's VQA score appears to increase substantially as more concepts are erased, which deserves further explanation. The authors should report historical erasure rates throughout the 100-stage process and clarify why the VQA score improves despite successive modifications to the model.

  5. Limited experimental validation across model architectures. All experiments are conducted on Stable Diffusion v1.4. Since the proposed approach is intended for practical continuous concept erasure, evaluating it on more recent architectures, such as SDXL or DiT-based models, would help establish its generalizability.

  6. Poor figure quality. The figures appear to be embedded as PNG images and become noticeably blurry when enlarged, particularly the plots and their annotations. For academic presentation quality, the authors should use vector graphics for plots and diagrams whenever possible, ensuring that text, curves, and labels remain readable at different scales.

  7. Non-standard mathematical notation. The mathematical notation throughout the paper could be improved for clarity and consistency. In particular, vectors and matrices are not consistently distinguished typographically. I suggest adopting a unified convention, such as bold lowercase letters for vectors and bold uppercase letters for matrices, and applying it consistently across the equations.

Questions for authors

See the weakness part

3 comments
AuthorsOpenReview ↗

Response to Reviewer xmUb

Summary. (W1) Against joint re-erasure of all 100 artists, LCCE achieves higher erasure while editing online. (W2) Our guarantee is conditional, as stated; we now verify it empirically. (W3) With everything else fixed, the queue outperforms a tuned fixed weight and a one-step adaptive rule. (W4) Early erasures do not revive over 100 edits; the VQA increase reflects content recognizability, not returning styles.

[W1] Joint re-erasure baselines. LCCE removes more target styles than UCE-joint and SPEED-joint, although they see all 100 targets at once and edit the original weights in a single step (Appendix C.5).

Method (100 artists)RR-LLM ↑ [95% CI]Non-target ΔCLIP ↑
UCE-joint0.764 [0.715, 0.812]−1.106
SPEED-joint0.735 [0.684, 0.780]−0.912
LCCE (online)0.829 [0.793, 0.863]−2.249

LCCE never revisits the original weights and stores no record of past requests — the defining constraint of the continuous setting (Sec. 3). A deployed model receives a removal request and must be updated in place. Joint re-erasure requires retaining every past request and the original checkpoint, then re-editing and re-validating the entire model at each new request; the CLIP gap reflects this fundamental difference in access.

[W2] Scope of the preservation guarantee. Clarification: the paper does not claim that LCCE guarantees queue stability for every request sequence. As the reviewer notes, Appendix A proves that mean-rate stability implies the long-term budget. The finite-horizon bound (Theorem A.2) holds unconditionally: 1T∑t=1TPL(t)≤D+Z(T+1)−Z(1)aT.\frac{1}{T}\sum_{t=1}^{T}\mathrm{PL}(t) \le D + \frac{Z(T+1)-Z(1)}{aT}. Stability itself requires the standard Lyapunov conditions, now stated explicitly in Appendix A.7: bounded per-stage loss, a budget DD that some policy can meet with a nonzero margin, and a per-stage action that minimizes the drift bound up to bounded error.

Empirically, a feasible budget is met by every edited projection (Appendix G.2, Table 18). Values are cumulative loss divided by the budget; below 1 means the budget is satisfied.

BudgetHorizonMean over projectionsWorst projection
α=2\alpha = 216 edits0.8900.983
α=1\alpha = 1 (default)16 edits1.3081.396
α=1\alpha = 1 (default)100 edits1.342–

The default sets DD to the loss of a single edit, which is tighter than a growing preservation bank can sustain. The queue therefore keeps raising preservation pressure, yet the excess stays flat from 16 to 100 edits rather than growing with TT. α\alpha selects the erasure–preservation operating point.

AuthorsOpenReview ↗

Response to Reviewer xmUb

[W3] Necessity of the virtual queue. With solver, bank, retention statistics, and artist order all fixed, the queue outperforms both a tuned fixed weight and a one-step adaptive rule (Appendix G.4). The fixed weight was selected on a separate development set by a rule fixed in advance (Table 20); all three controllers were evaluated once on a confirmation set with new prompts.

ControllerRR-LLM ↑Non-target ΔCLIP ↑
Virtual queue0.9031+0.6594
Fixed weight (c=10c = 10)0.7906+0.5607
One-step heuristic0.8738+0.3425

The queue is not just a varying weight: it accumulates preservation debt across all earlier stages. A fixed weight cannot react to accumulated drift, and the one-step rule sees only the last loss. The queue's advantage in erasure over the fixed weight, and in preservation over the one-step rule, comes from this history.

[W4] Long-term erasure and the VQA trend. Early-erased artists stay erased throughout 100 sequential edits (Appendix C.4, Fig. 5). After the first ten edits, RR-CLS and RR-LLM on these artists remain flat through the hundredth edit. UCE and MACE also keep erasure high, but collapse non-target CLIP and VQA (Fig. 3); SPEED preserves quality but leaves styles recognizable. LCCE is the only method that keeps both.

Clarification: the rising VQA score does not indicate returning styles. VQA measures whether the objects named in the prompt are recognizable, not whether a style is present. Abstract styles distort object structure; removing the style makes objects easier to recognize, raising VQA.

[W5] Other architectures. LCCE transfers to SDXL and FLUX.1-schnell without changing the method. Appendix C.8 applies the same closed-form update to both backbones over four sequential artist erasures:

BackboneRR-LLM ↑ original → LCCECLIP original → LCCE
SDXL0.350 → 0.87529.940 → 30.504
FLUX.1-schnell0.825 → 1.00028.698 → 28.449

LCCE edits the cross-attention key/value projections of SDXL and the two text-conditioning projections of FLUX. Erasure improves on both architectures; CLIP remains at the original level.

[W6, W7] Figures and notation. Fixed. All plots are now vector PDFs. Vectors use bold lowercase, matrices bold uppercase. The derivations in Appendix A are rewritten and cross-checked.

I remain unconvinced of the practical necessity of the proposed continuous editing setting, as existing editing-based methods can efficiently handle multi-concept erasure with limited additional overhead, while LCCE has not demonstrated a clear advantage over these alternatives (e.g., in the additional experiments, LCCE exhibits a substantially larger non-target CLIP degradation of −2.249 compared to −1.106 for UCE-joint and −0.912 for SPEED-joint); therefore, I maintain my original score.

Reviewer_DB2NOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Lyapunov-constrained Continuous Concept Erasure (LCCE) for sequential erasure in text-to-image diffusion models. A virtual queue adaptively regulates preservation strength, a cumulative preservation bank anchors non-target responses to the original model, and a historical retention loss maintains previous erasures. These components yield an equality-constrained closed-form update. Experiments on SD v1.4 show a favorable erasure–preservation trade-off and low editing cost, including preservation results over 100 artist-style erasures.

Critical strengths
  1. The virtual queue and cumulative preservation bank address accumulated preservation errors across sequential edits, providing history-dependent control rather than treating requests independently. The historical retention loss explicitly addresses interference with previous erasures, while fixed-size sufficient statistics and closed-form updates make repeated editing computationally attractive.

  2. Artist-style and object-erasure results demonstrate favorable target suppression and non-target preservation relative to the evaluated baselines, together with low per-concept editing cost.

Critical weaknesses
  1. Despite motivating the method through unsafe-content removal, the experiments cover only artist styles, objects, and celebrity identities. Broad semantic concepts such as nudity and violence involve diverse visual patterns and indirect descriptions, leaving this important application unvalidated.

  2. All experiments use U-Net-based SD v1.4. Although Appendix H discusses diffusion transformers, the absence of results on DiT-based models such as PixArt or FLUX leaves the transferability of the editing formulation, preservation mechanism, and efficiency unclear.

Questions for authors
  1. Figure 3 reports CLIP and VQA, which primarily assess alignment and content preservation rather than target-style removal. Could the authors report RR-CLS and RR-LLM across all 100 erased artists after the final edit, alongside preservation metrics, and track early-erased artists to assess concept revival?

  2. ETC(Erasing Thousands of Concepts: Towards Scalable and Practical Concept Erasure for Text-to-Image Diffusion Models) and OCE(Orthogonal Concept Erasure for Diffusion Models) report large-scale multi-concept erasure, while LACU(Locality-Aware Continual Unlearning for Diffusion Models) directly addresses continual unlearning. Can you clarify the practical advantages of LCCE over these methods, particularly its comparative performance in large-scale concept erasure?

  3. The reported sequences treat concept domains separately, whereas deployment may involve interleaved requests concerning celebrities, objects, characters, and artistic styles. Could the authors evaluate a heterogeneous erasing sequence and report erasure efficacy, historical erasure retention, and non-target preservation for each domain?

  4. Could the authors provide results on at least one DiT-based text-to-image model, such as PixArt or FLUX? It would be particularly useful to clarify which projections would be edited, how the invariant constraints would be constructed, and whether the erasure–preservation trade-off remains favorable as sequential edits accumulate.

3 comments
AuthorsOpenReview ↗

Response to Reviewer DB2N

Summary. (W1) On I2P nudity prompts, LCCE eliminates NudeNet detections with the best CLIP. (W2/Q4) LCCE runs unchanged on FLUX.1-schnell and SDXL. (Q1) Erasure holds over all 100 artists and early targets do not revive. (Q2) ETC/OCE address a different setting; against LACU, LCCE erases far more at equal preservation and orders of magnitude lower cost. (Q3) On a mixed-domain sequence, LCCE improves average removal and content preservation.

[W1] Unsafe and indirectly specified concepts. LCCE eliminates all NudeNet detections on I2P while keeping the highest CLIP among edited models (Appendix C.7, Table 14). We sequentially erase "nudity," "naked," "erotic," and "sexual," then evaluate on sampled I2P prompts, most of which do not contain these words.

StageLCCE NudeNet ↓ / CLIP ↑UCESPEED
0 (original)14.375% / 26.1714.375% / 26.1714.375% / 26.17
12.500% / 26.416.875% / 25.719.0% / 23.61
40% / 25.891.875% / 25.200% / 23.60

A single request removes most detections. SPEED also reaches zero by the last stage, but at notably lower CLIP.

[W2, Q4] DiT-based architectures. The editing formulation is architecture-agnostic: it applies wherever text enters through linear projections. For FLUX.1-schnell, LCCE edits the two text-conditioning projections and keeps all other weights fixed; the invariant subspace is built in FLUX's own conditioning space, including its pooled CLIP features.

Backbone (4 sequential edits)RR-LLM ↑ original → LCCECLIP original → LCCE
SDXL0.350 → 0.87529.940 → 30.504
FLUX.1-schnell0.825 → 1.00028.698 → 28.449
AuthorsOpenReview ↗

Response to Reviewer DB2N

[Q1] All 100 artists and revival of early targets. After 100 edits, LCCE erases more of the 100 artists than joint baselines that erase all of them at once (Table 12):

Method (100 artists)RR-LLM ↑ [95% CI]Non-target ΔCLIP ↑
UCE-joint0.764 [0.715, 0.812]−1.106
SPEED-joint0.735 [0.684, 0.780]−0.912
LCCE (online)0.829 [0.793, 0.863]−2.249

Early targets do not revive. Fig. 5 tracks the earliest-erased artists at every tenth stage; their RR-CLS and RR-LLM stay flat from the tenth edit to the hundredth.

[Q2] ETC, OCE, and LACU. Clarification: ETC and OCE solve a different problem. They erase a target set known in advance in one batch. LCCE handles requests that arrive one at a time and must preserve earlier erasures without access to past requests. Running ETC/OCE in our setting reduces to joint re-erasure after every new request; the joint UCE/SPEED comparison above is exactly that experiment at 100 concepts, and LCCE erases more.

Against LACU, the only direct continual-unlearning competitor, LCCE erases far more at equal preservation and runs orders of magnitude faster (Table 2):

MethodAcceAcc_e ↓AccpAcc_p ↑H0H_0 ↑Time / concept
LACU0.3440.8600.744≈40 min
LCCE0.0020.8600.924≈8 s

[Q3] Heterogeneous sequence. In a single sequence mixing style, object, identity, and character, LCCE improves both average removal and content preservation over UCE (Appendix C.6, Table 13). All four targets are evaluated after the last edit, so earlier targets directly measure retention across later cross-domain edits.

Concept (order)RR: UCE / LCCEContent-CLIP: UCE / LCCE
Van Gogh (1)0.6 / 0.728.405 / 29.826
Golf ball (2)0.1 / 0.920.062 / 20.583
Brad Pitt (3)1.0 / 0.822.130 / 22.450
Pikachu (4)1.0 / 1.027.069 / 26.638
Average0.675 / 0.85024.417 / 24.874

Thank you for the response and the additional experiments. The new results address several of my concerns and help clarify the method’s effectiveness. Although some concerns remain, I will raise my rating to 2 based on the additional evidence.

Reviewer_LKNZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies continuous concept erasure, where erasure requests arrive sequentially and the model must suppress each new target while retaining all previous erasures and preserving generation on non-target concepts. The authors identify two failure modes of mainstream single-step editing methods: cumulative parameter drift progressively degrades non-target representations and generation quality, and later edits weaken or reverse earlier erasures, causing removed concepts to re-emerge. They propose LCCE, which formulates continuous erasure as a Lyapunov-constrained online control problem: a dynamic virtual queue accumulates preservation-budget violations and adaptively regulates subsequent edits, a historical retention loss maintains previously erased states, and a cumulative preservation bank incrementally merges target semantic neighborhoods to provide the queue with a consistent preservation signal. Each stage admits a closed-form solution, avoiding iterative fine-tuning. Experiments on Stable Diffusion v1.4 cover continuous artist-style and object-category erasure, a 100-concept stress test, runtime, and ablations.

Critical strengths
  1. The problem formulation is novel: casting continuous erasure as online control with a long-term preservation budget, and feeding the virtual queue's preservation debt into the closed-form update, distinguishes it from fixed per-stage penalties.
  2. The two failure modes (cumulative drift and weakening of historical erasures) are clearly characterized and matched by the virtual queue plus preservation bank, and the historical retention loss, respectively.
  3. The cumulative preservation bank uses incremental semantic neighborhoods to keep the queue's signal comparable across stages, and the historical retention loss summarizes history with two fixed-size sufficient statistics, so storage does not grow with the sequence—an elegant engineering design.
  4. Each stage has a closed-form solution, taking about 8 seconds per concept versus 20 minutes for CUIG, two orders of magnitude faster, with clear deployment value.
  5. Experiments cover style and object tasks, a 100-step stress test, runtime, and a three-component ablation, with reasonably comprehensive metrics.
Critical weaknesses
  1. The Lyapunov guarantee relies on standard boundedness and feasibility assumptions, but the budget D is calibrated from the preservation loss after the first edit; there is no empirical verification of D's sensitivity or whether the long-term budget is actually satisfied.
  2. The preservation bank's semantic neighborhoods are LLM-generated and top-K selected by CLIP cosine similarity; neighborhood quality and the choice of K strongly affect the preservation signal, yet the main text lacks analysis of neighborhood construction quality and K sensitivity.
  3. The style task uses only 16 artists, and the 100-step stress test reports only CLIP/VQA curves without systematic quantification of erasure metrics (RR-CLS/RR-LLM) at 100 steps; the object task uses only 10 classes, a limited scale.
  4. Experiments are essentially limited to SD v1.4 without validating other backbones; against CUIG on the object task, the H0 advantage is marginal (0.924 vs. 0.888).
  5. The ablation shows that removing the dynamic queue yields the strongest erasure but worst fidelity, while removing the historical retention loss yields the best fidelity but weakest erasure, indicating the method still moves along the erase–fidelity trade-off rather than strictly dominating it; a quantitative characterization of this trade-off is missing.
Questions for authors
  1. How is the budget D calibrated? Under different D, is the long-term average preservation loss actually within the constraint? Can a sensitivity curve for D be provided?
  2. How much do the LLM neighborhoods and top-K choice in the preservation bank affect results? Can erroneous neighborhood construction contaminate the queue signal?
  3. For the 100-step stress test, can erasure metrics be reported as a function of step count, rather than only CLIP/VQA?
  4. Does the method remain effective on SD v2.1, SDXL, or FLUX? How stable is the closed-form solution when updating both key and value projections?
  5. Since H0 on the object task is close to CUIG, does LCCE's advantage come mainly from closed-form efficiency or from the control mechanism? Can a fair comparison under matched budgets be provided?
2 comments
AuthorsOpenReview ↗

Response to Reviewer LKNZ

Summary. (W1) The budget is met whenever it is feasible; α\alpha controls the operating point. (W2) Semantic neighborhoods matter; random ones hurt preservation, not erasure. (W5) Clarification: against every baseline, LCCE leads on all four metrics; there is no trade-off. (Q5) The gain comes from the control mechanism, isolated with the solver fixed.

[W1, Q1] Budget calibration and satisfaction. Clarification: sensitivity to the budget was already reported in the submission (Table 17, Fig. 9). The revision adds what was missing: whether the budget is actually met. The budget is D(ℓ)=αDbase(ℓ)D^{(\ell)} = \alpha D^{(\ell)}_{\mathrm{base}}, where Dbase(ℓ)D^{(\ell)}_{\mathrm{base}} is the loss of projection ℓ\ell after the first edit (Table 18):

α\alphaHorizonCum. loss / budget (mean)Worst projectionRR-LLM ↑
0.5161.6421.7770.825
1161.3081.3960.838
11001.342––
2160.8900.9830.925

With a feasible budget, every edited projection meets it. Smaller α\alpha sets the budget below what a growing preservation bank can sustain, so the feasibility condition (Appendix A.7) fails; the queue then keeps raising preservation pressure, yet the excess stays flat from 16 to 100 edits.

[W2, Q2] Neighborhood quality and KK. Semantic Top-50 gives the best erasure and preservation; a random neighborhood keeps erasure but damages preservation (Appendix G.3, Table 19).

NeighborhoodRR-LLM ↑Non-target ΔCLIP ↑Bank size
Top-250.8375−0.0422231
Top-500.8875+0.4485365
Top-1000.8250+0.3360568
Random-500.8700−0.9312650

Random selection is the worst-case neighborhood error, and it degrades preservation despite a larger bank. Errors affect the preservation signal, not erasure; more anchors do not compensate for wrong ones.

AuthorsOpenReview ↗

Response to Reviewer LKNZ

[W3, Q3] Erasure over 100 steps and scale. Early erasures persist through 100 edits (Fig. 5): RR-CLS and RR-LLM remain flat from the tenth edit to the hundredth. Clarification: our 100-artist sequence already exceeds all continual-unlearning baselines (CUIG and LACU use 10 concepts), extending the horizon by an order of magnitude.

[W4, Q4] Other backbones and closed-form stability. LCCE transfers to SDXL and FLUX.1-schnell with CLIP at the original level (Table 15; reproduced in our response to Reviewer xmUb, W5). The closed form is stable by construction: each projection is solved independently, and B(t)=λI+(PSD Gram terms)≻0\mathbf{B}(t) = \lambda\mathbf{I} + (\text{PSD Gram terms}) \succ 0 for λ>0\lambda > 0, so the solution exists and is unique at every stage.

[W4, Q5] Advantage over CUIG and its source. LCCE beats CUIG on every metric of both tasks (Tables 1 and 2):

TaskMethodMetrics
10 objectsCUIG GradProjAcceAcc_e 0.005 / AccpAcc_p 0.802 / H0H_0 0.888
10 objectsLCCEAcceAcc_e 0.002 / AccpAcc_p 0.860 / H0H_0 0.924
16 artistsCUIG GradProjRR-CLS 0.692 / RR-LLM 0.932 / CLIP 24.89 / VQA 39.16
16 artistsLCCERR-CLS 0.704 / RR-LLM 0.982 / CLIP 27.43 / VQA 46.05

The gain comes from the control mechanism, not only the closed form: Appendix G.4 fixes the solver and varies only the controller; the queue outperforms a tuned fixed weight and a one-step rule (Table 21). The closed form contributes speed (Table 4). A matched-compute comparison is ill-defined: CUIG's optimization cannot complete within one LCCE update, and running LCCE for CUIG's duration would change the method.

[W5] Erasure–fidelity trade-off. Clarification: against every baseline, LCCE does not trade erasure for fidelity; it leads on all four metrics at 16 artists (Table 1):

Method (16 artists)RR-CLS ↑RR-LLM ↑CLIP ↑VQA ↑
UCE0.6850.96512.205.40
ESD0.6010.96021.1531.20
CCRT0.6250.92019.5029.10
MACE0.4500.69025.9037.10
SPEED0.5150.69525.8040.20
CUIG GradProj0.6920.93224.8939.16
LCCE0.7040.98227.4346.05

The trade-off the reviewer observes is between ablated variants of LCCE (Table 3), which confirms that each component contributes: removing the queue maximizes erasure at the cost of fidelity; removing historical retention does the opposite. The full method combines both, and α\alpha controls the operating point explicitly (Table 18).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading. We highlight three points for the AC, followed by the additions (all marked in blue in the revision).

1. At the end of the 16-artist sequence, LCCE is best on all four metrics simultaneously (Table 1; reproduced in our response to Reviewer LKNZ, W5). Against the baselines, LCCE does not trade erasure for fidelity; the trade-off discussed by Reviewer LKNZ concerns ablated variants of LCCE itself.

2. LCCE removes more concepts than offline joint re-erasure, which sees all 100 targets at once, while processing requests one at a time (Table 12).

3. The theoretical claim is, and was, conditional. The paper states that mean-rate stability of the queue implies the long-term preservation budget; it never claims that stability holds for every request sequence. The revision makes the assumptions explicit (Appendix A.7) and measures the budget directly (Appendix G.2).

Additions. Erasure trajectories over 100 edits (C.4, Fig. 5); joint baselines (C.5, Table 12); LACU (Table 2); fixed-weight and one-step controllers (G.4, Tables 20–21); budget satisfaction and sensitivity (G.2, Table 18); neighborhood size and random selection (G.3, Table 19); mixed-domain sequence (C.6, Table 13); I2P nudity (C.7, Table 14); SDXL and FLUX.1-schnell (C.8, Table 15); vector figures and unified notation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.