acceptodds
Under review as a conference paper at ICLR 2027

What CLIP Sees When Structure Breaks: Patch-Shuffle Response as a Generalizable Cue for AI-Generated Image Detection

Abstract

Existing CLIP-based detectors identify AI-generated images from the static representation of a single image, which is inevitably entangled with image semantics and thus generalizes poorly to unseen generators. In this work, we shift the focus from asking what an image is to how its representation reacts when its global structure is deliberately broken. Specifically, we compare an image's representation with that of its patch-shuffled counterpart, where global spatial arrangement is destroyed while local patches remain intact. We observe that real images exhibit significantly stronger response shifts than generated ones within a frozen CLIP feature space, suggesting that real representations rely heavily on globally coherent structures, whereas generated ones are dominated by local patterns. To harness this observation, we propose SPARE Structural Perturbation-Aware REpresentation), a framework that converts structural responses into a reliable discriminative signal, including a Patch-Shuffle structural response module constructs the perturbed view and computes layer-wise responses; a layer-wise response gating module adaptively weights each layer's contribution based on its reliability; and a dual-branch response fusion detector integrates the gated responses with the original representations as a content reference. Notably, the CLIP encoder remains entirely frozen throughout. Extensive experiments across multiple benchmarks demonstrate that SPARE consistently surpasses state-of-the-art detectors in cross-generator generalization.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Image Detection, CLIP, Structural response

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_DcnYOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPARE, a CLIP based AI-generated image detector that exploits representation changes induced by patch shuffling. The authors argue that real images rely more heavily on global spatial coherence than generated images, resulting in stronger feature responses when spatial structure is disrupted. To exploit this observation, SPARE computes layer-wise feature differences between original and path shuffled image, applies adaptive gating, and combines these responses with original representations for classification.

Critical strengths
  • The method is build upon a frozen CLIP backbone, avoiding expensive backbone fine-tuning.
  • The paper evaluates generalization across multiple GAN and diffusion based generators.
  • The authors include component ablations, perturbation comparisons, and robustness experiments.
Critical weaknesses
  • SFLD (gye et al., 205) already introduced patch shuffling with frozen CLIP representations for AI-generated image detection, explicitly aiming to reduce semantic bias while preserving local texture cues. SPARE builds upon highly similar underlying motivation and methodological foundation, yet does not adequately establish its novelty relative to this work.
  • The authors attribute stronger structural responses in real images to their greater dependence on physical and spatial coherence. However, the presented evidence does not establish this explanation or rule out alternative causes.
  • Table 5, gaussian noise achieves an average accuracy of 92.82 compared with 93.81 for patch shuffle. This relatively small difference raises questions about whether the reported improvements specifically arise from disrupting global structure.
  • The paper does not convincingly establish that this combination constitutes a sufficiently substantial methodological contribution.
Questions for authors

See weaknesses

7 comments
AuthorsOpenReview ↗

Response to Reviewer DcnY (Part 1/4)

We thank the reviewer for the careful review, and in particular for pointing us to SFLD, which we should have discussed. We also appreciate the recognition of SPARE's efficiency with a frozen backbone, its cross-generator evaluation, and its ablation and robustness experiments. We address each point below.

W1. Relation to SFLD.

We agree that SFLD (Gye et al., WACV 2025) is closely related and should have been discussed. Both methods apply patch shuffling with a frozen CLIP ViT-L/14, and both aim to reduce content bias. We also correct an inaccurate sentence in our introduction stating that patch shuffling had previously been used only as a training-time augmentation; SFLD uses shuffled views at inference as well. The revision discusses SFLD in the Introduction, Related Work and Experiments.

The two methods differ in which signal is classified:

  • SFLD trains an independent linear classifier on the final-layer CLIP feature of a single view: the original image, or a shuffled image at patch size 56 or 28. At test time it averages 10 shuffled views per patch size and averages the logits of the three classifiers. Each classifier separates real from fake from one view alone; views are combined only by averaging scores.

  • SPARE classifies the change between the two views, δℓ=Fclsℓ−F~clsℓ\delta^{\ell} = F^{\ell}_{cls} - \tilde{F}^{\ell}_{cls}, over a window of intermediate layers, jointly with the original representation. Its input is how the representation reacts to the shuffle, not what either view looks like.

This difference is testable. Keeping the original branch, architecture, layers and training data unchanged, we replaced the response δ\delta with the static feature of the shuffled image F~\tilde{F}, so both variants receive exactly the same two views:

Second-branch inputForenSynthsGenImageAverage
Static shuffled feature F~\tilde{F} (SFLD-style)94.7487.6291.18
Structural response δ\delta (SPARE)95.6691.9693.81

Given the same two views, modeling their difference explicitly improves mean ACC by 2.63 pp, and by 4.34 pp on the cross-dataset GenImage benchmark.

SPARE is also cheaper at inference: 2 CLIP forward passes per image, compared with 21 for SFLD (1 original view + 2 patch sizes × 10 shuffled views).

Reference:
Seoyeon Gye, Junwon Ko, Hyounguk Shon, Minchan Kwon, and Junmo Kim . Sfld: Reducing thecontent bias for ai-generated image detection. In WACV, 2025 .

AuthorsOpenReview ↗

Response to Reviewer DcnY (Part 2/4)

W2. Evidence for the physical-coherence explanation.

About physical and spatial coherence. We clarify that physical constraints provide one possible explanation for the different structural responses observed in Fig. 1. As discussed by Garces et al. (2022), image appearance is jointly influenced by illumination, material properties, and geometry. We hypothesize that these factors induce spatial dependencies among different regions in real images, making their global representations more sensitive to structural perturbations. Nevertheless, SPARE focuses on capturing and exploiting these representation differences rather than identifying their underlying causes.

About empirical evidence. To investigate potential alternative explanations, we conduct additional experiments to examine the effects of image complexity, texture, and compression history on structural responses. Specifically, we evaluate BPP and Sobel gradient magnitude as proxies for image complexity and texture, respectively.

  1. Token-aligned shuffling. Because the patches coincide with ViT tokens, every token keeps exactly the same pixels and only its position changes. The response therefore contains no seam artifacts and no change in local texture within any token.
  2. Complexity and texture point the other way. For each image we compute JPEG compressibility (bits per pixel after re-encoding at Q = 75) and texture strength (mean grayscale Sobel gradient magnitude). Generated images are slightly more complex and textured than real ones (ProGAN: bpp 1.691 vs. 1.663, Sobel 84.18 vs. 83.14; SD v1.4: bpp 1.808 vs. 1.722, Sobel 85.54 vs. 82.93), yet real images respond more strongly. If ∥δ∥\Vert\delta\Vert merely tracked complexity, the opposite would hold.
  3. Controls do not reduce the separation. We (i) stratify images into five bins by bpp or by Sobel and compute the sample-weighted within-bin AUROC, (ii) regress ∥δ∥\Vert\delta\Vert on complexity (quadratic in log-bpp and Sobel with interaction, pooled over both classes) and use the residual as the score, and (iii) re-encode all images with JPEG Q = 95 to align compression history. AUROC of ∥δ∥\Vert\delta\Vert, with real images as the positive class:
SettingRaw (95% CI)Stratified by bppStratified by SobelResidualized
ProGAN0.7975 (0.7802–0.8198)0.81870.82430.8243
ProGAN, JPEG Q950.7902 (0.7676–0.8125)0.81090.81660.8181
SD v1.40.6150 (0.5876–0.6373)0.60430.62060.6557
SD v1.4, JPEG Q950.6015 (0.5742–0.6237)0.61040.63780.6208

Overall, these experiments suggest that the observed structural response differences cannot be simply explained by image complexity, texture intensity, or compression history. Although they do not establish physical and spatial coherence as the sole underlying cause, they provide further empirical support for the discriminative value of structural responses. We have included these additional experiments in the revised manuscript.

Reference:
Garces, E., Rodriguez-Pardo, C.,·Casas, D.,·Lopez-Moreno, J. A Survey on Intrinsic ImagesDelving: Deep Into Lambert and Beyond. International Journal of Computer Vision, 130(3):836–868, 2022.

AuthorsOpenReview ↗

Response to Reviewer DcnY (Part 3/4)

W3. Gaussian noise vs. Patch Shuffle (Table 5).

To further clarify the effectiveness of Patch Shuffle, we conduct additional ablation experiments to distinguish the benefits of structural response modeling from those of simply introducing perturbed images:

PerturbationForenSynthsGenImageAverage
Gaussian Noise (Static shuffled feature F~\tilde{F})89.1381.6385.38
Gaussian Noise (Structural Response δ\delta)95.5690.0892.82
Patch Mask (Static shuffled feature F~\tilde{F})64.4055.0559.73
Patch Mask (Structural Response δ\delta)88.2586.7787.51
Patch Shuffle (Static shuffled feature F~\tilde{F})94.7487.6291.18
Patch Shuffle (Structural Response δ\delta)95.6691.9693.81

(a) Does the response formulation matter?
Yes. For every perturbation, classifying the response δ\delta outperforms classifying the perturbed view, with the original branch and all other settings unchanged.

(b) Does the type of perturbation matter?
Yes. When the perturbed view itself must carry the signal, Patch Shuffle outperforms Gaussian noise by 5.80 pp and Patch Mask by 31.45 pp, so the perturbation type strongly affects how much discriminative information the perturbed view holds. Under the response formulation the gap narrows to 0.99 pp.

(c) Positioning.
We do not claim that only structural perturbations carry information. RIGID (He et al., 2024) showed that sensitivity to small noise in foundation-model features is informative, and our Gaussian-noise results are consistent with this. Our claim is that the representation response is an effective cue, and that Patch Shuffle is its most effective instance among the perturbations we tested. It is also the perturbation that best isolates dependence on spatial arrangement: with token-aligned patches the content of every token is unchanged and only positions move, whereas noise and masking alter the content itself.

Reference:
Z. He, P.-Y. Chen, T.-Y. Ho. RIGID: A Training-Free and Model-Agnostic Framework for Robust AI-Generated Image Detection. arXiv:2405.20112, 2024.

Reviewer_9KXbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses semantic bias and limited cross-generator generalization in existing CLIP-based AI-generated image detectors. It proposes SPARE, a framework that exploits differences in feature responses between real and generated images under patch-shuffle structural perturbations. SPARE combines layer-wise response gating with dual-branch fusion of original representations and structural responses to distinguish real from generated images. By keeping the CLIP encoder frozen and training only lightweight modules, the method improves cross-generator and cross-dataset detection accuracy across multiple benchmarks.

Critical strengths

1.Novel perspective. The paper goes beyond static image representations by examining how representations change when global spatial structure is disrupted, providing a fresh perspective on AI-generated image detection. 2.Well-motivated framework design. Layer-wise gating adaptively weights structural responses to suppress perturbation-induced noise, while dual-branch fusion incorporates original image representations as a content-conditioned reference. The ablation results support the contributions of these components. 3.Strong generalization and encouraging robustness. The method achieves strong average detection accuracy across multiple benchmarks covering both GAN-based and diffusion-based generators. It also outperforms the evaluated baselines under JPEG compression and Gaussian noise, and shows reasonable stability across different patch-shuffle permutations, particularly in AP.

Critical weaknesses

1.Insufficient evidence for the proposed explanation. The paper attributes the different perturbation responses of real and generated images to a greater reliance on global structure in CLIP representations of real images and on local patterns in those of generated images. However, Figure 1 primarily provides empirical observations, without controlled experiments to rule out confounding factors such as content complexity and texture. Moreover, the sample size, sampling procedure, and aggregation procedure for the layer-wise curves are not clearly specified, making it difficult to assess the representativeness and reproducibility of this finding. 2.Additional inference overhead. SPARE requires encoding both the original image and its patch-shuffled counterpart, introducing an additional CLIP forward pass. Table 10 reports an inference time of 15 ms per image, compared with 10 ms for LTD and 12 ms for ForgeLens. Thus, freezing the backbone reduces training requirements but does not eliminate the additional computational cost at inference.

Questions for authors

1.Could the authors provide more details about the analysis in Figure 1? In particular, please specify the number and sources of real and generated images, the sampling procedure, the patch size and shuffle ratio, and how the layer-wise cosine distances and response norms are aggregated. Were the results averaged over multiple random shuffles, and how consistent are the observed differences across samples and permutations? 2.Could the authors clarify the exclusion of the DR subset from DRCT-2M? Could the authors also report performance on the DR subset to assess whether the claimed generalization extends to these samples?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9KXb (Part 1/4)

We thank the reviewer for the constructive feedback and for recognizing SPARE's perspective on structural perturbation responses, the design of LRG and DRF, and its generalization and robustness. We address each point below.

Q1. Details of the Fig. 1 analysis.

We apologize for omitting these details. The protocol is as follows:

  • Data and sampling: 1,000 real and 1,000 generated images per generator, sampled uniformly at random (seed 0) from the test sets of the ProGAN subset of ForenSynths and the SD v1.4 subset of GenImage.
  • Preprocessing: center-crop to 224×224, as at inference.
  • Encoder and layers: frozen CLIP ViT-L/14. Class tokens of all 24 layers are used for the cosine-distance curves, and those of the fixed layers 13–17 for ∥δ∥\Vert\delta\Vert. The learned ALS module is not used, so the analysis involves no trained component.
  • Patch Shuffle: 14×14 patches aligned with the ViT patch grid; ρ = 0.7.
  • Aggregation: for each layer, the L2 distance (or cosine distance) between the class tokens of the original and shuffled image; averaged over the five layers, then over 10 random permutations. The same 10 permutations are applied to all images, giving one response value per image. Layer-wise curves average over images.

Consistency across samples and permutations. For across-sample consistency, we independently draw 10 random subsets of 2,000 images from the full dataset for each class. For each subset, we first average the structural responses over 10 shuffle rounds for each image and then calculate the mean response across the sampled images. We subsequently compute the mean and standard deviation across the 10 sampling trials. For across-permutation statistics, we first average the structural responses over the 10 shuffle rounds for each image, and then calculate the mean and standard deviation, across the images of each class. The results are summarized as follow:

DatasetClassAcross-samplesAcross-permutations
ProGANReal3.7031 ± 0.18393.7031 ± 0.2939
ProGANFake3.0111 ± 0.17373.0111 ± 0.2258
SD v1.4Real4.3713 ± 0.12044.3713 ± 0.3373
SD v1.4Fake3.5531 ± 0.16903.5531 ± 0.3668

As shown, real images consistently exhibit larger structural response magnitudes than generated images across both datasets and under both statistical evaluations. Moreover, the relatively small standard deviations across random permutations indicate that the observed differences are not highly sensitive to the specific patch shuffling patterns.

AuthorsOpenReview ↗

Response to Reviewer 9KXb (Part 2/4)

Q2. DR subsets of DRCT-2M.

Following the evaluation setting adopted by existing work (LTD, Yang et al . (2026)), we initially excluded the DR subset to maintain consistency with previous works. According your suggestion, we further evaluate our SPARE on the DR subset, including SDv1-DR, SDv2-DR, and SDXL-DR. The DR subsets are diffusion reconstructions of real MSCOCO photographs and share their content and global layout, which makes them a demanding test for any detector.

MethodSDv1-DR ACCSDv1-DR APSDv2-DR ACCSDv2-DR APSDXL-DR ACCSDXL-DR AP
CNNSpot50.1556.5550.0252.2150.0560.66
UnivFD50.2555.9451.0564.9050.8567.38
NPR50.0060.5050.1063.1050.0255.00
FatFormer64.4478.2759.5771.8654.1762.26
RINE49.5041.6049.6545.2049.5045.60
D³50.0967.6650.0161.4849.9751.08
ForgeLens50.3758.8249.5051.7250.1849.89
LTD50.3268.6050.1465.9149.8955.52
SPARE (Ours)71.0083.6361.5075.4258.0074.98

All methods find these subsets difficult, and most baselines stay near 50% ACC. SPARE achieves the highest ACC and AP on all three subsets, but its accuracy on SDv2-DR and SDXL-DR remains limited. We view forgeries whose global layout is inherited from a real photograph as a boundary case for structure-based cues, and will add this to the Limitations.

AuthorsOpenReview ↗

Response to Reviewer 9KXb (Part 3/4)

W1. Evidence for the proposed explanation.

We agree that Fig. 1 alone cannot rule out confounders such as content complexity and texture. Using the same images and protocol as in Q1, we address this in three ways.

  1. Token-aligned shuffling. The 14×14 patches coincide with ViT tokens, so every token keeps exactly the same pixels and only its position changes. The response contains no seam artifacts and no change in local texture within any token; it reflects how much the representation depends on the arrangement of token contents.
  2. Complexity and texture point the other way. For each image we compute JPEG compressibility (bits per pixel after re-encoding at Q = 75) and texture strength (mean grayscale Sobel gradient magnitude). Generated images are slightly more complex and textured than real ones (ProGAN: bpp 1.691 vs. 1.663, Sobel 84.18 vs. 83.14; SD v1.4: bpp 1.808 vs. 1.722, Sobel 85.54 vs. 82.93), yet real images respond more strongly. If ∥δ∥\Vert\delta\Vert merely tracked complexity, the opposite would hold.
  3. Controls do not reduce the separation. We (i) stratify images into five bins by bpp or by Sobel and compute the sample-weighted within-bin AUROC, (ii) regress ∥δ∥\Vert\delta\Vert on complexity (quadratic in log-bpp and Sobel with interaction, pooled over both classes) and use the residual as the score, and (iii) re-encode all images with JPEG Q = 95 to align compression history. AUROC of ∥δ∥\Vert\delta\Vert, with real images as the positive class:
SettingRaw (95% CI)Stratified by bppStratified by SobelResidualized
ProGAN0.7975 (0.7802–0.8198)0.81870.82430.8243
ProGAN, JPEG Q950.7902 (0.7676–0.8125)0.81090.81660.8181
SD v1.40.6150 (0.5876–0.6373)0.60430.62060.6557
SD v1.4, JPEG Q950.6015 (0.5742–0.6237)0.61040.63780.6208

Across all controls, the AUROC stays essentially unchanged or slightly higher.

We also note the limits of this evidence. The complexity differences between classes are small (at most about 5%), so these results show that the response is not a proxy for complexity, texture strength or compression history, but they do not test robustness under large shifts in these statistics. The training-free separation is also weaker on SD v1.4, which is why SPARE learns layer-wise gating and conditions on the original representation rather than thresholding ∥δ∥\Vert\delta\Vert directly.

Reviewer_bA2cOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SPARE (Structural Perturbation Aware REpresentation), a novel method for AI-generated image detection that exploits differences in structural responses between original images and their patch-shuffled counterparts. Visualizations of these structural responses reveal a clear separation between real and synthetic images. Extensive experiments across multiple datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  • This work introduces a relatively novel perspective on AI-generated image detection by exploiting differences in representation responses between original images and their structurally perturbed counterparts. Specifically, it investigates whether responses to disruptions in global spatial structure can serve as discriminative cues for distinguishing real from generated images.

  • The paper presents a clear motivation and a well-structured framework that is easy to follow. The key components are well motivated, and the overall methodology is clearly described. Furthermore, the ablation studies provide empirical evidence supporting the effectiveness of the proposed modules.

Critical weaknesses
  • Insufficient differentiation from closely related work. The authors acknowledge that SemAnti also employs Patch Shuffle, but the paper does not sufficiently clarify the methodological differences or conceptual novelty of the proposed approach. Given the similarity in the use of structural perturbations, a more detailed comparison is necessary to establish the distinct contribution of SPARE. Without such clarification, it is difficult to assess the novelty of the proposed method relative to existing work.

  • Insufficient justification of the core design choice. The proposed PSR module relies on feature differences between original and patch-shuffled images to capture structural responses. However, the paper does not adequately justify why this difference-based formulation is preferable to alternative feature fusion strategies, such as concatenation or learnable fusion. Since PSR is a central component of the proposed framework, the lack of controlled comparisons makes it difficult to determine whether the reported improvements stem from the proposed structural response formulation or simply from incorporating features of perturbed images. This weakens the empirical support for the paper’s main methodological contribution.

Questions for authors

please refer to the weaknesses above

4 comments
AuthorsOpenReview ↗

Response to Reviewer bA2c (Part 1/2)

We sincerely thank Reviewer for the careful reading and constructive feedback, as well as for recognizing the novelty of our perspective on AI-generated image detection through structural perturbation responses (Strength 1), the clear motivation and well-structured methodology of SPARE (Strength 2), and the empirical evidence supporting the effectiveness of our proposed modules through ablation studies (Strength 3). We also appreciate the reviewer's valuable suggestions and address the raised concerns point by point below.

W1. Insufficient differentiation from closely related work.

We acknowledge that both SemAnti and our proposed SPARE employs Patch Shuffle. We would like to point out that the distinction between these two methods has already been discussed in the Related Works section (Section 2, "Local Forensic Cues Beyond Semantics"). Nevertheless, we further clarify their differences in terms of motivation and methodological design below.

SemAnti employs Patch Shuffle as data augmentation to alleviate the semantic bias in the pre-trained CLIP models by disrupting the spatial arrangement of image patches. Meanwhile, it fine-tunes artifacts’ sensitive shallow and middle layers, thereby guiding the model to learn more generalizable artifact cues.

In contrast, SPARE is motivated by a novel empirical observation: the real and generated images exhibit different representation responses to global structural perturbations for a frozen CLIP model. Specifically, we apply Patch Shuffle to images from the ProGAN and SD v1.4 datasets and measure the feature discrepancies between the original and perturbed views across different Transformer layers. As illustrated in Figure 1, the distributions of structural response magnitudes exhibit clear differences between real and generated images, demonstrating clear separation between real and generated images. This observation suggests that representation changes induced by structural perturbations can serve as effective discriminative cues for distinguishing real images from AI-generated ones.

Motivated by this finding, SPARE employs Patch Shuffle not for data augmentation or CLIP fine-tuning, but as structural perturbations to construct perturbed views. It extracts structural responses by measuring the representation discrepancies between the original and perturbed views in the frozen CLIP feature space. To further exploit these responses, SPARE introduces three modules: the Patch-Shuffle Structural Response (PSR) module extracts structural responses from different Transformer layers; the Layer-wise Response Gating (LRG) module adaptively weights these responses; and the Dual-branch Response Fusion (DRF) module integrates the structural responses with the original image representations through a dual-branch architecture, thereby improving detection reliability. We will add the difference discussion of these two methods in the related works after the Reviewer’s discussions.

AuthorsOpenReview ↗

Response to Reviewer bA2c (Part 2/2)

W2. Insufficient justification of the core design choice.

We first clarify that our feature difference design to capture structural response is not arbitrary, but is motivated by our empirical observations.

Specifically, we observe that the real and generated images exhibit distinct representation responses to global structural perturbations in the frozen CLIP feature space. As shown in Figure 1, we extract features from different layers of the original and patch-shuffled images, and calculate the cosine distances between two views and the L2 norms of their feature differences. We observe clear differences between real and generated images across different layers, indicating that representation differences induced by structural perturbations can provide discriminative cues for distinguishing real images from generated images. Therefore, we adopt the representation differences between the original and perturbed views to measure structural responses, rather than simply fusing features of the two views.

Furthermore, we conduct additional ablation studies to provide empirical justification for this design choice. We compare our representation difference (denoted as Difference) with two alternative baselines: feature concatenation (denoted as Concatenation) and learnable feature fusion (denoted as Learnable Fusion). In these experiments, we only change the feature interaction strategy between the original and perturbed views while keeping all other experimental settings unchanged. The results are reported in the following Table 1. As shown, our proposed representation difference strategy achieves better performance than both feature concatenation and learnable feature fusion. These results further verify the effectiveness of the structural response in PSR module and show that the performance gains of SPARE are not simply due to the additional features from perturbed images.

Table 1:Ablation study of feature interaction strategies for structural response on ForenSynths and GenImage. Results are reported in Mean ACC (%).

Feature Interaction StrategyForenSynthsGenImageAverage
Concatenation93.0789.5291.30
Learnable Fusion94.0690.8992.48
Difference (Ours)95.6691.9693.81

Thank you for the detailed response. The clarification on SemAnti addresses my first concern. SPARE uses Patch Shuffle to measure structural responses in a frozen CLIP, while SemAnti uses it as augmentation for fine-tuning. The new ablation supports the difference-based design in PSR. It also shows that the gains do not come only from adding perturbed-view features. Please include the SemAnti comparison in the revised version. Please also include this ablation. My concerns have been addressed, and I will maintain my positive rating.

Dear Reviewer bA2c:

Thank you for reviewing our response and for recommending acceptance. We will incorporate the clarification of SPARE's differences from closely related work and the newly added experimental results into the revised manuscript.

Thank you again for your valuable comments and suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.