- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
In this paper, the authors propose sequential unlearning at the data-point level for Diffusion Models. The authors argue that existing unlearning methods, when applied to sequential unlearning, suffer from sequential reappearance. Basically, previously unlearned targets reappear when other targets are sequentially unlearned, which is a critical vulnerability. So here they propose LASTING, which extends SISS with PGD over bounded image/latent perturbations to identify low-denoising-loss neighbors for forget-side updates, reducing target reappearance during sequential diffusion unlearning.
Critical strengths
- This is a promising direction, as in practical real-world deployment, we need scalable and reliable unlearning methods that can handle sequential requests.
- Concept reappearance is a critical vulnerability with unlearning algorithms, and in a sequential setup, this will increase further. In this work, LASTING offers good improvement over other baseline methods with respect to this.
Critical weaknesses
- This sequential unlearning setup has been explored at the concept level in both the Diffusion and LLM setups. So this is a meaningful extension at the data-point level, but the paper lacks in the evaluation/metrics section. Catastrophic forgetting (model degradation) is a common problem at the concept level. In the case of CelebA-HQ, since it has access to full retention data, this won’t be very significant. But when it comes to SD models, we don’t have access to retain data, and we mostly use a small synthetic dataset, so retention is affected. So, for the SD experiments, we need more metrics to measure this. Here, the authors have not reported FID comparisons for the SD experiments as well.
- LASTING doesn’t perform very well on scaling to SD models, and also when full retention data is not available. With SD models, we can see that the CLIP-IQA metric shows that retention is comparatively poor compared to other methods, so, as I mentioned in the 1st point, the model might be facing catastrophic forgetting.
- The sequential unlearning setup for CelebA-HQ assumes that it knows all the future deletion requests, which is not practical (Appendix C.1). In the real-world deployment, we only have the current unlearned model along with the current unlearning request and previous requests (if stored).
- Novelty/Contribution - Sequential reappearance is not a significant contribution, as it has already been shown in the literature that fine-tuning the unlearned model on even random data causes reappearance of the unlearned target. Unlearning is essentially fine-tuning, so the reappearance phenomenon is expected. Even sequential unlearning works at the concept level for LLMs and Diffusion Models face the problem of concept reappearance, but they mainly focus on Catastrophic forgetting, as it’s an even bigger problem.
- Limited to just the SDv1.5 model only.
Questions for authors
- This is a promising direction, but currently it requires substantial changes. See the points mentioned in Weaknesses.
- Better evaluation metrics for SD experiments. It is better to focus experiments on SD models, as in practice, you don’t have access to retained data, so we need methods that perform well even without it. Also, unlearned models should have good retention performance, which isn’t really measured here.
- [Future direction] I would like to see how the unlearning persists when earlier deleted targets or related variants reappear in later retain batches, measuring recovery speed, reappearance rates, and retained semantic utility.
7 comments
Response to Reviewer LuJ4 (2/3)
W4 + part of W1. Contribution relative to prior recovery and sequential-unlearning studies
Prior work on recovery after fine-tuning and concept-level sequential unlearning also examines whether forgetting persists through later model updates. Our contribution, however, is to link each target's reappearance risk to its Recovery Accessibility after deletion. We use this analysis to decide which representations to update during unlearning. We track deleted targets through ordinary subsequent deletions without target reuse or a separate recovery intervention. This analysis motivates selecting locally recoverable, low-denoising-loss representations as update targets, with benefits under both SISS and FU.
The results below examine the direction of neighbor selection and its applicability to another objective. Bold indicates the lowest reappearance rate within each base objective.
Table C3. Extended neighbor-selection ablation (K=200, %)
| Forget-side update target | SISS Final Success ↑ | SISS Final Reapp. ↓ | FU Final Success ↑ | FU Final Reapp. ↓ |
|---|---|---|---|---|
| Original target | 94.50 | 5.50 | 65.50 | 34.17 |
| Random Neighbor | 75.50 | 23.74 | 61.50 | 38.19 |
| High-Loss Neighbor | 54.00 | 37.20 | 45.50 | 51.85 |
| Low-Loss Neighbor (LASTING, Ours) | 97.50 | 2.50 | 95.50 | 4.50 |
Our methodological contribution is to turn the Recovery Accessibility finding into an explicit rule for choosing what to unlearn. LASTING targets the locally recoverable, low-loss representations implicated by our analysis of sequential reappearance. The ablation directly tests this selection rule. Low-loss selection improves persistence under both SISS and FU, while random perturbations worsen it. High-loss selection substantially worsens persistence, raising reappearance rates to 6.8× for SISS and 1.5× for FU relative to original-target unlearning. This connects an explanation of why deleted targets return to a concrete intervention that improves forgetting persistence across different objectives.
Q3. Reintroduction of previously deleted targets or variants
We thank the reviewer for suggesting this promising research direction. Given the time constraints, we conducted a small exploratory experiment on CelebA-HQ, starting from each method's K=200 checkpoint. We randomly selected 10 targets that had SSCD<0.7 both immediately after deletion and at K=200 under both methods. During each of the next 10 deletion stages (201–210), all 10 selected targets were reintroduced, one target in each of 10 distinct optimizer updates. Each target was inserted once per stage, giving 10 re-exposures per target and 100 target insertions in total. For each method, a matched run without reinsertion served as the control.
The rates below refer to the 10 selected targets. Bold indicates the lowest reappearance count within each reinsertion condition, including ties.
Table C4.
| Method | Reinsertion | Final Reapp. (K=210) ↓ | Ever Reapp. (K=201–210) ↓ | Re-exposures until first reappearance |
|---|---|---|---|---|
| SISS | No | 0/10 | 0/10 | — |
| SISS | Yes | 4/10 | 4/10 | 1, 3, 4, 10 |
| LASTING | No | 0/10 | 0/10 | — |
| LASTING | Yes | 2/10 | 2/10 | 5, 7 |
All 10 targets receive 10 re-exposures each over stages 201–210. The last column reports the cumulative re-exposure count at the first stage-end detection (SSCD≥0.7) for each reappearing target. A dash indicates no reappearance.
Neither method showed reappearance without reinsertion. With reinsertion, LASTING had fewer reappearing targets (2/10 versus 4/10), and its earliest return occurred after five re-exposures, compared with one for SISS. SISS may retain nearby representations that allow deleted targets to return quickly. LASTING updates these representations during unlearning, which may make rapid recovery less likely. Developing methods that remain robust to repeated re-exposure is a promising direction for future work. We thank the reviewer for this suggestion.
Response to Reviewer LuJ4 (3/3)
W5. Additional evaluation on SD v2.1
Thank you for suggesting evaluation on an additional model. We selected SD2.1 to evaluate LASTING beyond SD1.4.
Evaluating instance-level unlearning requires first establishing that the model reproduces the target image before deletion. Otherwise, failure to reproduce it afterward does not demonstrate suppression of previously memorized content. In its SD v1.4 experiments, SISS uses prompts that induce reproduction of specific training images to evaluate instance-level unlearning (Alberti et al., 2025, §5.3). Following this setting, we evaluate whether reproduction of memorized images is suppressed after deletion and whether this suppression persists through subsequent deletion updates.
Because the same prompt can elicit different memorization behavior across models, we used candidate memorization prompts from prior work (Jeon et al., 2025, Appendix D) rather than directly reusing the SD1.4 prompt set. We generated images using these candidates and manually inspected the outputs to identify 30 prompt–target pairs exhibiting memorization. We then used these pairs for sequential unlearning on SD2.1.
We evaluate the final model after 30 deletion requests (K=30) using both the original and modified prompts for each target, yielding 60 prompt–target cases in total. For each case, we generate 16 images and consider the target detected if at least one image has an SSCD similarity to the target reference at or above the threshold.
Somepalli et al. identify candidate copies of training images using SSCD similarity above 0.5 in their Stable Diffusion experiments (Somepalli et al., 2023, §3–4), while Finding DoRI uses scores above 0.7 to detect reproduction of memorized images (Kowalczuk et al., §3.2). To examine sensitivity to the detection threshold, we evaluate τ∈{0.4, 0.5, 0.6, 0.7, 0.8}, covering these prior cutoffs and nearby values. We apply all five thresholds to the same generated images and SSCD scores.
Table C5. Forgetting performance on SD2.1 (K=30, %)
| Metric | SISS | FU | ReTrack | LASTING (Ours) |
|---|---|---|---|---|
| Final Reappearance (τ=0.4) ↓ | 22.22 | 0.00 | 38.89 | 5.77 |
| Final Reappearance (τ=0.5) ↓ | 20.69 | 0.00 | 22.22 | 5.17 |
| Final Reappearance (τ=0.6) ↓ | 18.33 | 0.00 | 10.00 | 1.67 |
| Final Reappearance (τ=0.7) ↓ | 13.33 | 0.00 | 1.67 | 0.00 |
| Final Reappearance (τ=0.8) ↓ | 6.67 | 0.00 | 0.00 | 0.00 |
| Final Success (τ=0.4) ↑ | 76.67 | 100.00 | 61.67 | 95.00 |
| Final Success (τ=0.5) ↑ | 78.33 | 100.00 | 76.67 | 95.00 |
| Final Success (τ=0.6) ↑ | 81.67 | 100.00 | 90.00 | 98.33 |
| Final Success (τ=0.7) ↑ | 86.67 | 100.00 | 98.33 | 100.00 |
| Final Success (τ=0.8) ↑ | 93.33 | 100.00 | 100.00 | 100.00 |
LASTING matches or outperforms SISS and ReTrack on both forgetting metrics across all evaluated thresholds. Although FU achieves 100% Final Success at every evaluated threshold, we observed severe quality degradation when inspecting its generated images.
General generation quality. To assess whether improved forgetting persistence preserves general generation ability, we evaluate the same checkpoints on non-deleted COCO prompts and compare them with Base. FID-COCO measures distributional differences from real COCO images, while FID-base measures deviation from Base's generation distribution. CLIPScore measures image–text alignment, and CLIP-IQA and MUSIQ assess perceptual quality without a reference image.
Table C6. General generation quality on COCO for SD2.1 (K=30)
| Metric | Base | SISS | FU | ReTrack | LASTING (Ours) |
|---|---|---|---|---|---|
| FID-COCO ↓ | 24.41 | 30.07 | 98.03 | 41.61 | 26.98 |
| FID-base ↓ | 0.00 | 14.83 | 85.90 | 32.07 | 9.47 |
| CLIPScore ↑ | 0.8185 | 0.8182 | 0.6849 | 0.7877 | 0.8179 |
| CLIP-IQA ↑ | 0.9371 | 0.8686 | 0.5645 | 0.8604 | 0.8857 |
| MUSIQ ↑ | 71.15 | 67.21 | 59.72 | 69.34 | 67.75 |
LASTING largely preserves Base's text alignment and perceptual quality, with nearly unchanged CLIPScore and modest reductions in CLIP-IQA and MUSIQ. Its FID-COCO remains close to Base, and it achieves the lowest FID-COCO and FID-base among the updated methods. FU's strong forgetting performance comes with substantial generation-quality degradation. Its FID-COCO is approximately 4.0× that of Base.
Overall, LASTING achieves stronger forgetting persistence than SISS and ReTrack while maintaining competitive generation quality and text alignment on non-deleted prompts.
We will release the images and the full list of prompts used in this evaluation on GitHub.
Response to Reviewer LuJ4 (1/3)
We thank the reviewer for the constructive feedback. We agree that forgetting persistence should be evaluated alongside generation performance on non-deleted content. Below, we present additional SD generation evaluations, examine sequential unlearning without advance knowledge of future deletion requests, and clarify our contribution relative to prior work.
W1·W2 + Q2. Generation quality and retention in SD
CLIP-IQA in Table 4 measures perceptual image quality without a reference image and was evaluated on outputs from deletion-target prompts. Since these results alone cannot establish general generation ability or text alignment, we additionally compare the Base model with the models at K=10, 20, and 30 using fixed COCO prompts.
We select one caption for each of 10,000 distinct COCO val2014 images using selection seed 2026 and generate 10,000 images per model state. All methods use the same captions, 512×512 resolution, PNDM with 50 steps, and a guidance scale of 7.5.
We evaluate SD v1.4 using FID-COCO, FID-base, CLIPScore, CLIP-IQA, and MUSIQ. FID-COCO measures distributional differences from real COCO images, while FID-base measures deviation from the original model's generation distribution. CLIPScore measures image–text alignment (Hessel et al., 2021), whereas CLIP-IQA (Wang et al., 2023) and MUSIQ (Ke et al., 2021) assess perceptual image quality without a reference image. The quality metrics use fixed COCO prompts and seeds, while Final Reappearance uses the deletion-target evaluation from Table 4. Base is the model before unlearning. Bold indicates the best value among updated methods, excluding Base and including ties.
Below, we show K=30, the final checkpoint of the 30-request sequence. Please refer to Table 20 in the appendix for the full COCO generation-quality results at K=10, 20, and 30.
Table C1. SD v1.4 (K=30)
LASTING largely preserves perceptual quality and text alignment relative to Base, with higher CLIPScore, nearly unchanged CLIP-IQA, and only a 0.50-point decrease in MUSIQ. It achieves the highest CLIPScore and MUSIQ among the updated methods and lower FID-COCO than SISS and ReTrack, alongside the lowest reappearance rate. These results show that LASTING maintains competitive generation quality on non-deleted content while improving forgetting persistence.
W3. Sequential unlearning without advance knowledge of future deletion requests
The original protocol in Appendix C.1 uses a fixed retain pool that excludes all 300 deletion targets in advance. To test whether LASTING's persistence gains hold without advance knowledge of future deletion requests, we ran sequential unlearning experiments that exclude only previously deleted targets and the current deletion target from the retain pool. Future targets remain eligible for retention until their deletion requests arrive. We compare LASTING with SISS, ReTrack, and FU over 300 deletion requests. FID follows the original CelebA-HQ evaluation protocol.
Below, we show K=300, the final checkpoint. Please refer to Table 22 in the appendix for the K=100 and K=200 results.
Table C2. Sequential unlearning (K=300, seed=42)
LASTING achieves the lowest Final Reappearance and highest Final Success among the compared methods. These results show that LASTING's persistence gains hold without advance knowledge of future deletion requests.