- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates how quantization affects the reliability of concept erasure in diffusion models. The authors examine the relationship between parameter fidelity, activation-level functional fidelity, and behavioral erasure effectiveness. Experiments on Stable Diffusion 1.5 show that weight quantization can substantially distort parameter updates while largely preserving their functional effects. In contrast, low-bit activation quantization significantly weakens nudity erasure for UCE and TIME, while other concepts remain largely unaffected. The paper highlights the importance of evaluating concept erasure under deployment-time numerical configurations.
Critical strengths
-
The research problem is Meaningful. The paper investigates an important yet underexplored issue: whether concept erasure remains effective after model quantization, a common practice in deployment.
-
The multi-level analysis is convincing. The distinction between parameter fidelity, functional fidelity, and behavioral effectiveness provides useful insights into the limitations of weight-space metrics for evaluating erasure preservation.
-
The design of experiments is reasonable. The paper includes multiple quantization settings, matched-noise controls, and additional checkpoints, providing empirical support for its main observations.
Critical weaknesses
-
Lack of a unified explanation for the observed phenomena. The paper reports substantially different effects of activation quantization across concepts: nudity erasure degrades significantly, whereas Van Gogh and church erasure remain largely preserved or even improve. However, the paper does not provide a convincing explanation for this discrepancy or establish a unified understanding of when and why quantization compromises concept erasure. The connection between quantization errors and behavioral degradation also remains insufficiently explained. This may affect the evaluation of the paper’s theoretical contributions.
-
The generalizability is limited. The main findings are primarily established on SD 1.5 with UCE and TIME, which are closely related editing approaches. The results for ESD-x are less conclusive, and the behavior on other architectures and concept-erasure methods remains unclear. Therefore, it is difficult to tell whether the phenomenon in this work is generalizable.
Questions for authors
-
Could the authors further clarify why activation quantization significantly weakens nudity erasure while Van Gogh and church erasure remain largely unaffected or even improve?
-
Following Q1, is there a unified explanation connecting quantization errors to these different behavioral outcomes?
-
Could the authors further discuss whether the observed phenomena are specific to UCE and TIME on SD 1.5, or whether they can generalize to other concept-erasure methods and model architectures? Additional evidence would help strengthen the generality of the conclusions.
I'm open to raising my score if the authors can address these concerns.
3 comments
Thank you for your response and additional experiments. I appreciate the authors' efforts to address my concerns. The new causal interventions and concept-specific analyses provide stronger empirical evidence and help clarify the observed phenomena. However, the underlying mechanism explaining why different concepts respond differently to quantization remains insufficiently understood. While the rebuttal strengthens the paper, my major concern is not fully addressed so I will maintain my original rating of 2.
Clarification of the Mechanistic Finding and Its Remaining Limit
Thank you for reassessing the revision and acknowledging the stronger intervention evidence and clearer concept-specific analysis. We agree that the network-level explanation for why the same perturbation produces different target-score directions remains incomplete. The revision distinguishes two questions. First, which computation can cause erasure weakening? The fixed-weight rounding and complementary preservation interventions localize a sufficient failure pathway to the BOS value inputs within the tested SD-1.x setting. Second, why does perturbing that pathway increase some concept scores and decrease others? The titration and decomposition characterize those outcomes but do not fully answer this second question. We therefore do not present the decomposition as a complete mechanistic explanation. Its narrower contribution is to show that apparently preserved or improved normalized efficacy can accompany substantial behavioral changes, including attenuation of the base concept or generation degradation. We understand that this remaining explanatory gap continues to affect your assessment. We respectfully maintain that experimentally identifying and intervening on an erasure failure pathway, alongside demonstrating the limitations of weight-fidelity diagnostics, constitutes a useful contribution within the stated scope. We have made the unresolved concept-direction question explicit in the revision. Thank you again for your careful review and engagement.
Response to JFY1: Causal Interventions, Concept Dependence, and Generalization
Thank you for recognizing the value of separating parameter fidelity, functional effects and behavioral erasure. Your request for a connection between quantization error and behavior motivated the new interventions and concept analysis in Section 6.
Q1-Q2: What connects activation quantization to behavioral weakening, and why do concepts differ?
We now identify a specific failure pathway by intervention. With deployed weights fixed, rounding only the BOS token's cross-attention value inputs weakens nudity erasure for UCE, TIME and a published ESD checkpoint. NudeNet efficacy changes are -0.243 [-0.454, -0.030], -0.224 [-0.434, -0.036] and -0.167 [-0.255, -0.088], respectively. CLIP agrees, and edited-model nudity scores rise, so the finding is not solely a normalization effect.
The complementary intervention preserves those inputs under scoped W8A4. It improves normalized efficacy for all three erasers on both detectors and directly outperforms preserving EOS: for example, UCE +0.201 [+0.052, +0.352] on NudeNet. Preservation is not established as equivalent to W8A16, and NudeNet absolute-suppression gains remain unresolved. Figure 2 and Table 3 report the interventions.
The concept analysis distinguishes outcomes that the original normalized metric obscured:
Writing efficacy as η=1−E/B, with base and edited target scores B,E and changes δB,δE, its change is exactly
Δη=B(B+δB)EδB−BδE.
This separates changes in the base score from changes in the edited score. It explains the metric's direction, not why the network produces those changes.
The edit-strength titration provides an additional distinction: for violence, the difference between the fully edited and base models' responses is resolved, +0.142 [+0.041, +0.254]. However, the complete prespecified three-contrast account of edit counteraction is not established; the corresponding interaction for nudity is unresolved on NudeNet. We therefore provide a common intervention pathway and differentiated behavioral evidence, without claiming a universal predictor of concept direction. See Figure 3 and Tables 15-16.
Q3: How far do the results generalize?
The published ESD checkpoint adds a fine-tuned eraser beyond UCE/TIME, and violence adds a second safety target with UCE. Same-GPU violence tests show BOS-only weakening, -0.175 [-0.334, -0.047], and an advantage of BOS over EOS preservation, +0.119 [+0.020, +0.230], on Q16.
We also evaluated FLUX.1-dev with an EraseAnything LoRA under actual SVDQuant INT4 execution and MixDQ's W8A8 simulation on SDXL-turbo. Neither shows statistically resolved weakening in the tested configuration. These extend coverage while bounding the claim: they do not establish transfer of the SD-1.x mechanism or equivalence to full precision. The FLUX adapter executes in a 16-bit branch, and implementation differences are not isolated. Tables 17-18 report the results and limitations.
All intervals are 95% paired prompt-bootstrap intervals. The general response details our correction of the submitted ESD E1/E4 base-checkpoint mismatch.
We hope the intervention evidence and concept-specific analysis address the central concerns behind your assessment. We would welcome your reassessment and clarification of any remaining substantive concerns.