acceptodds
Under review as a conference paper at ICLR 2027

Not the Scalar, but Its Evolution: Robust Source Attribution for Generated Images

Abstract

The widespread deployment of latent generative models (LGMs) makes generated image source attribution essential for copyright auditing and data governance, yet practical attribution should remain reliable under common post-processing and platform-induced processing. Existing reconstruction-based attribution methods summarize an image's reconstruction response to a single global *scalar*, such as the reconstruction error or the ratio of two consecutive errors. These summaries discard channel-frequency structure, allowing image content and post-processing variation to obscure source evidence and making attribution after spreading difficult. We instead uncover a more stable source-discriminative cue under post-processing: the frequency-resolved *evolution* of residual energy across repeated applications of the same reconstruction operator. Motivated by this insight, we identify a phenomenon we term ***Spectral Contraction*** and use it to build ***SpCon***, a source-attribution framework that accesses only the target model's VAE. SpCon obtains adjacent residuals through two deterministic reconstructions, builds a spectral-contraction representation from their frequency-resolved relative energy change, and aggregates Student- response densities for each source-condition pair over a finite set of fitted reference components. The same representation supports three settings according to non-belonging reference availability: none (*Target-Only*, TO; belonging references only), external out-of-domain sources (*External-Reference*, ER), or known in-domain competitors (*Known-Reference*, KR). Across seven LGMs, TO, ER, and KR attain average AUROC/AP of 0.8786/0.6344, 0.9404/0.7886, and 0.9652/0.8626 over the 35 conditions, all outperforming the strongest baseline (0.8154/0.5277), and achieve 0.8581/0.5507, 0.9369/0.7549, and 0.9734/0.8773 on the three real-world social platforms.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Generated Image Attribution, Latent Generative Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZXx1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SpCon for generated-image source verification under post-processing. It based on AEDR and applies a target VAE twice, represents adjacent residuals through log-energy ratios across RGB channels and radial frequency bands, and scores the resulting features using Student-t mixture densities. Three settings TO, ER, and KR reflect different levels of negative-reference availability. Evaluation covers seven generators, 35 synthetic conditions, and three social platforms.

Critical strengths
  1. The method is simple and well motivated. Preserving channel–frequency structure can retain source information discarded by a global reconstruction-error ratio, while requiring only two VAE reconstructions.
  2. The controlled experiments are substantial. AEDR-Mix matches reference access and density modeling, while additional controls examine single residuals, normalized residuals, and concatenated residual spectra. The gains therefore cannot be attributed solely to additional reference data or a more elaborate scoring backend.
  3. Reference-assisted variants achieve clear empirical improvements. ER/KR outperform the evaluated baselines under many single-operation conditions and the tested platform pipelines. The appendices also disclose informative failure cases involving shared VAEs, processing chains, and adversarial attacks.
Critical weaknesses
  1. Table 11 provides an important control: with matched references, fitting conditions, and density modeling, SpCon-ER achieves 0.9387/0.7825 AUROC/AP over all post-processing conditions, compared with 0.7444/0.4496 for AEDR-Mix-ER. The band-count ablation in Figure 5 further supports the value of spectral resolution. However, attribution metrics do not distinguish between smaller post-processing-induced feature changes and richer features that remain discriminative despite substantial changes. Figure 3 provides qualitative evidence through average response patterns, but a paired stability analysis across models and conditions is missing. Moreover, the common-factor cancellation identity in Appendix A is valid under its assumptions, but the extent to which actual transformations satisfy those assumptions is not quantified. The novelty should be framed accordingly. AEDR already uses double reconstruction and relative errors; the principal extension is spectral structure preservation and reference-density modeling.
  2. Appendix J already evaluates processing chains, but these findings deserve greater prominence. In Table 32, KR decreases from 0.9838/0.9208 AUROC/AP under individual medium-severity operations to 0.8633/0.5647 for two-operation chains and 0.6427/0.2286 for three-operation chains. Table 31 reports KR AUROC of 0.7677 for Resize→JPEG and 0.5365 for Rotation→Resize→JPEG. These transformations are directly relevant to the motivating dissemination scenario. Strong results on three platform pipelines do not establish generalization to repeated editing, resampling, and re-encoding. Representative chain results should appear in the main paper, with claims distinguishing individual operations, the tested platform pipelines, and sequential processing. Broader claims would benefit from evaluating composition-aware fitting on unseen operation combinations and orders.
  3. Section 3.1 defines membership with respect to a specific generator, whereas Section 5 and Appendix H primarily support VAE-family-level attribution. Table 29 reports approximately 0.50–0.55 AUROC when distinguishing SD3.5-Medium from Large using their shared VAE. The authors acknowledge this limitation, but the abstract, formal task definition, and application motivation should consistently reflect it. The seven-way evaluation establishes discrimination within the selected candidate set rather than general model-level identification. The paper should explicitly define its attribution classes and explain how generators sharing a VAE should be handled in the intended application.
Questions for authors

Please see the weaknesses。

7 comments
AuthorsOpenReview ↗

Overall Response to Reviewer ZXx1

We sincerely thank you for recognizing the simplicity and theoretical motivation of our method (S1), the substantial controlled experiments (S2), and the clear empirical improvements (S3). We also appreciate your detailed and valuable feedback. In response, we have added paired stability analyses, quantitatively examined the common-factor assumption, and further clarified our novelty (W1); moved sequential-processing results into the main paper and added composition-aware fitting experiments (W2); and unified the attribution-class definition while clarifying how shared-VAE models are handled (W3). The manuscript has been revised accordingly, with changes highlighted in blue. Each response identifies the corresponding manuscript revisions. Our detailed responses follow.

AuthorsOpenReview ↗

Re: Paired Stability, Common-Factor Cancellation, and Novelty (W1)-1/2

We thank you for recognizing the value of Table 11 and Figure 5. We agree that attribution metrics alone cannot distinguish the contributions of feature stability and discriminative power, and that the theoretical assumptions in Appendix A require further empirical validation. To address these concerns, we have added paired stability analyses across models and post-processing conditions, quantitative tests of the theoretical assumptions, and controlled representation-level experiments. We have also clarified SpCon's novelty relative to AEDR and revised the abstract and contributions accordingly.


(1) Stability or discriminability: Primarily richer discriminative features. We measured paired feature changes before and after post-processing, normalized by each representation's own between-class separation on clean images, and separately evaluated discriminative performance using only clean images for reference fitting.

ScopeRepresentationNormalized Drift (Belonging / Non-belonging)Clean-Only Fitting (AUROC / AP)
All post-proc.First-residual spectrum1.4899 / 1.17960.8082 / 0.4764
All post-proc.Global ratio (AEDR-style)0.6286 / 0.24060.7668 / 0.4704
All post-proc.SpCon full spectrum1.1088 / 0.84770.8506 / 0.6270
PlatformsFirst-residual spectrum1.3679 / 0.96760.8162 / 0.4078
PlatformsGlobal ratio (AEDR-style)0.6543 / 0.10180.7800 / 0.4792
PlatformsSpCon full spectrum1.0414 / 0.72410.9217 / 0.6850

Results show SpCon's advantage does not arise solely from smaller feature drift. Although SpCon exhibits larger normalized drift than the global ratio, it achieves higher average AUROC/AP with clean-only reference fitting (higher AUROC on 5 of 7 targets and higher AP on 6 of 7), indicating the preserved channel–frequency structure provides richer discriminative information. Compared with the first-residual spectrum, SpCon drifts less for belonging images in 14 of 15 conditions, consistent with partial cancellation by the relative response.


(2) Quantitative validation of the assumptions in Appendix A. We further quantified the extent to which the common-factor assumption holds. Let Δs1\Delta \mathbf{s}_1 and Δs2\Delta \mathbf{s}_2 denote the changes in channel–frequency log energies of the two residuals. We define the cancellation coefficient as:

κ=1−E[∥Δs1−Δs2∥2]E[∥Δs1∥2].\kappa = 1 - \frac{\mathbb{E}[\lVert \Delta \mathbf{s}_1 - \Delta \mathbf{s}_2 \rVert^2]}{\mathbb{E}[\lVert \Delta \mathbf{s}_1 \rVert^2]}.

Here, κ=1\kappa=1 indicates perfect common-factor cancellation, while lower values indicate greater deviations from the assumption. The results show that the assumption holds approximately for sharpening, blurring, contrast adjustment, resizing, brightness adjustment, and noise (κ=0.818\kappa=0.818–0.9610.961), moderately for cropping, rotation, and Telegram (κ=0.58\kappa=0.58–0.750.75), and less closely for saturation, WebP, JPEG, WeChat, flipping, and Facebook (κ≤0.33\kappa\leq0.33). Importantly, the band-wise cancellation coefficient exceeds its global (B=1B=1) counterpart for 11 of 12 operation types, all three platforms, and six of seven targets, demonstrating the advantage of band-wise comparison over global aggregation. Where κ\kappa is low, robustness is not explained by cancellation alone: reference densities directly model the remaining differential components for fitted conditions (e.g., JPEG), while their mixture supports generalization to unfitted conditions. These findings demonstrate that SpCon's robustness benefits from both band-wise cancellation and reference-density modeling, rather than relying on the cancellation assumption alone. We have added these quantitative results to Appendix A and clarified the theoretical scope of the cancellation identity.

AuthorsOpenReview ↗

Re: Paired Stability, Common-Factor Cancellation, and Novelty (W1)-2/2

(3) Novelty clarification. We agree with your characterization of SpCon's contribution. SpCon builds upon AEDR's double-reconstruction and relative-error framework, while related double-reconstruction or second-order-difference designs also appear in [1–4]. We do not claim this underlying strategy as our contribution. Contribution 2 (Section 1) and Section 2 now explicitly state two extensions: (i) preserving the 48-dimensional channel–frequency structure of relative changes between consecutive reconstruction residuals, rather than collapsing them into a global scalar; and (ii) reference-density modeling. We have also replaced "identify a phenomenon" with "characterize." These extensions are not merely a domain transformation. By Parseval's theorem, transforming to the frequency domain and then aggregating global energy is mathematically equivalent to the spatial-domain global ratio (Appendix A). Thus, the improvement cannot be attributed to the domain transformation alone, but to the preserved structure. To further demonstrate the distinction from AEDR, we added controlled representation-level experiments using identical references, fitting conditions, and density-modeling backends, varying only the input representation.

[1] SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution, CVPR 2026.

[2] Data Provenance for Image Auto-Regressive Generation, ICLR 2026.

[3] MGI: Member vs Generated Inference, ECCV 2026.

[4] A Difference-in-Difference Approach to Detecting AI-Generated Images, CVPR 2026.

Controlled representation-level results (AUROC / AP):

RepresentationDimensionTOERKR
Global Ratio10.6999 / 0.32390.7447 / 0.43720.7498 / 0.4502
AEDR-Mix10.7495 / 0.42360.7444 / 0.44960.7579 / 0.4724
Mean of 48 Band Responses10.5370 / 0.19500.7315 / 0.42300.7527 / 0.4408
RGB Channel Means30.6161 / 0.25730.7850 / 0.47270.7995 / 0.4893
SpCon480.8756 / 0.62510.9387 / 0.78250.9642 / 0.8586

Global Ratio uses B=1B=1 with summed RGB energies and is equivalent to −log⁡(MSE2/MSE1)-\log(\mathrm{MSE}_2/\mathrm{MSE}_1) by Parseval's theorem, without the GLCM term. Mean of 48 Band Responses averages the 48-dimensional response vector z\mathbf{z}, while RGB Channel Means averages the 16 band responses within each RGB channel. Under identical backends, Global Ratio, AEDR-Mix, Mean of 48 Band Responses, and RGB Channel Means achieve only 0.73–0.80 AUROC under ER/KR, whereas the full 48-dimensional SpCon representation reaches 0.9387/0.9642. This demonstrates that the gains arise from preserving channel–frequency structure rather than the domain transformation itself. Moreover, references alone do not close the gap: with identical references and backend, no low-dimensional summary exceeds 0.80 AUROC under KR, whereas the 48-dimensional response reaches 0.9642.


Manuscript revisions: Appendix A now includes Empirical Check of the Common-Factor Assumption, with the definition of κ\kappa and per-condition results (Table R2), and Paired Stability versus Discriminability, reporting paired drift and clean-only discriminative performance across conditions and targets (Tables R2–R3). Appendix I adds the controlled representation comparisons (Table R8). We have reframed the novelty as spectral structure preservation and reference-density modeling in Section 1 (including the replacement of "identify" with "characterize" and the revision of Contribution 2) and Section 2. The abstract, introduction, and opening of Appendix A no longer claim inherently greater feature stability, instead emphasizing richer and more robust discriminative cues. Appendix A also reports the clean-only results for the first-residual spectrum.

Reviewer_ZPimOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

SpCon tests whether an image comes from a target latent generative source. It passes the image through the source's frozen VAE twice, then measures the log-energy change between the residuals in each RGB frequency band. Source-condition Student-t densities score this response. Target-Only, External-Reference, and Known-Reference scoring differ in their reference sets. Experiments cover seven target generators and synthetic processing conditions, followed by three platform paths. Further studies examine shared VAEs, processing chains, and adaptive attacks.

Critical strengths

Equations 9-14 give the representation a concrete motivation: global ratios can shift with input-dependent energy weights, while band-wise comparison can cancel changes shared within a coordinate. The limits of this argument are stated.

Critical weaknesses
  1. The idea is quite similar to this paper "AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction", except for transforming the pixel domain to the frequency domain.

  2. State the attribution granularity and processing limits earlier. In Appendix H, Medium-versus-Large discrimination with a shared VAE is near chance: Table 29 gives AUROC 0.5006 for ER and 0.5518 for KR. The other shared-VAE checkpoints are excluded from KR fitting, so this comparison does not include reference samples from those exact competitors. Processing chains pose a separate limit. KR falls from 0.9838 on the constituent single operations to 0.6427 on three-operation chains in Appendix J (Table 32). Section 5 already reports both findings. The abstract and initial claims should carry the same VAE-family and processing qualifications. Within that scope, the measured improvements remain useful.

  3. Threshold selection uses additional negative data. Appendix D.3 keeps calibration and evaluation separate, but all scoring functions receive the same held-out labeled negatives for threshold selection (p.23, lines 1215-1217). Table 15 is therefore a controlled comparison under common calibration access. It leaves open how TO would be deployed without negative references, and whether ER can use only external negatives for calibration. Please label that access next to the results. Calibrating ER on its permitted external pool, if feasible, would answer the practical question with the existing method.

  4. I could not trace the configuration-selection split or the variability from reference fitting. Figure 5 motivates the choice of 16 frequency bands. Section 4.7 uses a smaller ablation test split, but its relation to the final test set is unspecified. Please give the split indices and say when the configuration was frozen. The bootstrap unit also needs to be stated. Since the current intervals condition on the fitted references, a small reference-resampling check would help establish stability. The description leaves these procedures unclear; it does not establish leakage. Optimizer-seed experiments would add little for deterministic inference.

Questions for authors
  1. Can the abstract and task definition state the VAE-family scope and distinguish the tested platform paths from arbitrary processing chains? That qualification would address the scope concern without changing the method.
  2. What happens when ER thresholds are calibrated only on its allowed external negatives? This deployment result should be reported separately from the threshold-free comparison.
  3. How are band-selection images separated from final-test images? How stable are the main gains under a different 100-image reference draw? A split manifest and a bounded sensitivity check would increase confidence in the result.
  4. Carefully clarify the fundamental difference from the existing similar paper.
6 comments
AuthorsOpenReview ↗

Overall Response to Reviewer ZPim

We sincerely thank you for recognizing the theoretical motivation and clearly stated limits of our approach (S1), as well as your valuable suggestions regarding novelty, application scope, and experimental protocols. In response, we have clarified the fundamental difference from AEDR and added controlled representation-level experiments (W1 & Q4); specified the attribution granularity and processing limitations (W2 & Q1); clarified threshold-calibration access and added deployment experiments (W3 & Q2); and documented the configuration-selection split, freezing point, and reference-resampling stability (W4 & Q3). The manuscript has been revised accordingly, with changes highlighted in blue. Each response identifies the corresponding manuscript revisions. Our detailed responses follow.

AuthorsOpenReview ↗

Re: Fundamental Difference from AEDR (W1 & Q4)

SpCon's frequency-domain extension is not merely a domain transformation, but is specifically designed to address information loss during real-world image dissemination. We acknowledge that SpCon follows AEDR's general double-reconstruction strategy, a design also adopted by SWIFT [1], Data Provenance [2], MGI [3], and DID [4]. The novelty lies in adapting this strategy to more realistic scenarios where information loss during image propagation challenges robust attribution. Specifically, SpCon preserves the 48-dimensional channel–frequency structure of energy changes between consecutive reconstruction residuals, which cannot be recovered from a global scalar, and characterizes its distribution through reference-density modeling to provide richer discriminative information for robust attribution. Sharing the same general strategy does not imply representational equivalence: simply transforming to the frequency domain and aggregating global energy yields a result mathematically equivalent to the spatial-domain global ratio. To further demonstrate the fundamental difference from AEDR, we added controlled representation-level experiments using identical references, fitting conditions, and density-modeling backends, varying only the input representation to directly assess its discriminative value and robustness.

[1] SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution, CVPR 2026.

[2] Data Provenance for Image Auto-Regressive Generation, ICLR 2026.

[3] MGI: Member vs Generated Inference, ECCV 2026.

[4] A Difference-in-Difference Approach to Detecting AI-Generated Images, CVPR 2026.


Controlled representation-level results (AUROC / AP):

RepresentationDimensionTOERKR
Global Ratio10.6999 / 0.32390.7447 / 0.43720.7498 / 0.4502
AEDR-Mix10.7495 / 0.42360.7444 / 0.44960.7579 / 0.4724
Mean of 48 Band Responses10.5370 / 0.19500.7315 / 0.42300.7527 / 0.4408
RGB Channel Means30.6161 / 0.25730.7850 / 0.47270.7995 / 0.4893
SpCon480.8756 / 0.62510.9387 / 0.78250.9642 / 0.8586

Global Ratio uses B=1B=1 with RGB energies summed and, by Parseval's theorem, is equivalent to −log⁡(MSE2/MSE1)-\log(\mathrm{MSE}_2/\mathrm{MSE}_1), without the GLCM term. Mean of 48 Band Responses averages the 48-dimensional response vector z\mathbf{z}, while RGB Channel Means averages the 16 band responses within each channel. Under identical backends, Global Ratio, AEDR-Mix, Mean of 48 Band Responses, and RGB Channel Means achieve only 0.73–0.80 AUROC under ER/KR, whereas the full 48-dimensional SpCon representation reaches 0.9387/0.9642. This demonstrates that the gains arise from preserving channel–frequency structure rather than the domain transformation itself. Moreover, references alone do not close the gap: with identical references and backend, no low-dimensional summary exceeds 0.80 AUROC under KR, whereas the 48-dimensional response reaches 0.9642.


Manuscript revisions: Contribution 2 in Section 1 now explicitly describes two extensions to AEDR's double-reconstruction framework (channel–frequency structure preservation and reference-density modeling). Section 2 (Related Work) acknowledges the shared double-reconstruction strategy with AEDR. Section 4.7 summarizes the low-dimensional representation comparisons. Appendix I adds Low-Dimensional Summaries Under the Final Protocol and Table R8, comparing Global Ratio, AEDR-Mix, Mean of 48 Band Responses, and RGB Channel Means against SpCon.

AuthorsOpenReview ↗

Re: Attribution Granularity and Processing Limits (W2 & Q1)-1/2

We have clarified two scope limitations in the abstract, introduction, Section 3.1, and Section 5: SpCon primarily provides VAE-family attribution, and robustness is reported separately for individual post-processing operations, tested social-platform pipelines, and sequential processing chains. To address your two concerns, we conducted two additional investigations.

(1) VAE-family attribution. We explored including samples from competing shared-VAE models in KR reference fitting. We found that discrimination improves substantially when the non-belonging references include images generated by other models within the same VAE family.


Specifically, we evaluated three models sharing the same VAE in the SD3.5 family: SD3.5-Medium (M), SD3.5-Large (L), and SD3.5-Large-Turbo (LT). We considered all six configurations with two or three shared-VAE models in the candidate set. In each comparison, the left side of "vs" denotes the target model and the right side denotes the non-target model(s). SpCon-KR+Comp extends KR by including samples from competing shared-VAE models in its known non-belonging references.

Shared-VAE model discrimination (AUROC / AP):

SD3.5 FamilySpCon-ER (Original)SpCon-KR (Original)SpCon-KR+CompAUROC gain [95% CI]AP gain [95% CI]
M vs L0.5006 / 0.50010.5518 / 0.55970.7420 / 0.71110.1902 [0.1395, 0.2393]0.1514 [0.0994, 0.2040]
M vs LT0.7037 / 0.74150.6797 / 0.72020.9353 / 0.93880.2555 [0.2227, 0.2870]0.2186 [0.1891, 0.2489]
L vs LT0.7718 / 0.81860.7289 / 0.77970.9531 / 0.95880.2242 [0.1930, 0.2559]0.1791 [0.1507, 0.2047]
M vs L+LT0.6022 / 0.42050.6158 / 0.46480.8043 / 0.64680.1885 [0.1539, 0.2243]0.1820 [0.1320, 0.2326]
L vs M+LT0.6891 / 0.54140.6514 / 0.50170.8307 / 0.70960.1794 [0.1523, 0.2052]0.2079 [0.1655, 0.2425]
LT vs M+L0.6385 / 0.51730.6966 / 0.56500.9431 / 0.89040.2465 [0.2203, 0.2738]0.3254 [0.2833, 0.3655]

Both AUROC and AP improve substantially across all six configurations. AUROC exceeds 0.93 for M vs LT, L vs LT, and LT vs M+L, while M vs L remains the most challenging case (0.7420). This extension does not modify the method, but incorporates competing shared-VAE models into KR's known non-belonging references, directly addressing your observation that the original comparison excluded reference samples from those exact competitors. This requires access to samples from individual models within the same VAE family. Since shared-VAE models are often released and maintained by the same organization, this requirement can be satisfied in internal auditing scenarios.

Reviewer_zJ3cOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SpCon for source verification of generated images. It represents the energy evolution between two consecutive VAE reconstruction residuals across RGB channels and frequency bands, then models these responses using multivariate Student-t distributions. Even with multiple clues, the traceability rate of post-processed images remains outstanding.

Critical strengths
  1. A well-motivated methodological contribution. SpCon characterizes how reconstruction residual energy evolves across RGB channels and frequency bands, preserving information discarded by scalar reconstruction statistics. The paper clearly explains this design, and the residual-spectrum ablations and backend-matched AEDR comparison support the representation's contribution. Requiring only two reconstructions through a frozen target VAE makes the approach computationally practical. This combination of methodological originality and supporting evidence is a principal reason for my positive assessment.

  2. Substantial improvements under the evaluated post-processing conditions. Across the 28 held-out conditions, SpCon-ER and SpCon-KR achieve average AUROC/AP of 0.9342/0.7734 and 0.9602/0.8459, respectively, compared with 0.8222/0.5415 for AEDR. Evaluation on three social platforms further supports the practical relevance of these gains. Although AEDR already performs strongly on clean images, SpCon addresses an important weakness of existing reconstruction-based methods under post-processing. These results are the main empirical basis for my Weak Accept recommendation.

Critical weaknesses
  1. Concerns regarding application scenarios: For unpost-processed images, SpCon shows no obvious advantage over AEDR, which already achieves a tracing rate of nearly 100%. For post-processed images, SpCon delivers significantly better performance compared with AEDR. However, post-processing workflows for many images today are not limited to traditional operations such as cropping and resizing. Instead, AI assistance is more widely adopted for resolution enhancement, watermark removal, or detail adjustment. Therefore, from a practical perspective, such AI-post-processed images should be included in the research scope.

  2. Application scope and deployment: The abstract and introduction describe model-source attribution without stating the shared-VAE limitation discussed in Section 5 and Appendix H. Please clarify in these sections that SpCon primarily provides VAE-family attribution and cannot reliably distinguish individual models sharing the same VAE.

Questions for authors
  1. Under what conditions should a positive SpCon decision be interpreted as identifying an individual generator, rather than statistical compatibility with its VAE family?

  2. For the fixed-threshold experiment, how sensitive are the realized TPR and FPR to the calibration split, particularly across individual targets and unseen conditions?

7 comments
AuthorsOpenReview ↗

Overall Response to Reviewer zJ3c

We sincerely thank you for recognizing the methodological originality and computational practicality (S1), as well as the post-processing robustness and practical relevance (S2) of our work. We appreciate your valuable feedback. In response, we have added AI-assisted post-processing experiments (W1), clarified the scope of VAE-family-level attribution (W2), specified the conditions for individual-generator identification (Q1), and examined the sensitivity of TPR and FPR to calibration splits (Q2). The manuscript has been revised accordingly, with changes highlighted in blue. Each response also identifies the corresponding revisions in the manuscript. Our detailed responses follow.

AuthorsOpenReview ↗

Re: AI-Assisted Post-Processing (W1)

We have added 10 AI-assisted post-processing operations covering the three scenarios you highlighted: resolution enhancement, watermark removal, and detail adjustment. None of these operations was used for reference fitting. SpCon-ER/KR outperform all baselines in average AUROC and AP, with KR outperforming the strongest baseline on every operation. We also note that although AEDR already achieves nearly 100% traceability on clean images, SpCon offers an advantage at low false-positive rates. At 1% FPR, the false-negative rate decreases from 5.4% (AEDR) to 1.1% (ER) and 0.14% (KR).


The added AI-assisted operations are as follows:

  1. Resolution enhancement: Real-ESRGAN ×2 after 0.5× downsampling; SwinIR ×2 after 0.5× downsampling; and Real-ESRGAN ×2 at the original resolution. Real-ESRGAN (ICCVW 2021, 37K+ stars) and SwinIR (ICCVW 2021, 5K+ stars) are widely adopted super-resolution methods.
  2. Watermark removal: LaMa inpainting (WACV 2022, 10K+ stars) to remove synthetic visible watermarks at known locations; and re-encoding with a learned compression model (CVPR 2020), a common regeneration step in watermark removal attacks.
  3. Detail adjustment: SCUNet denoising (MIR 2023, 800+ stars); FBCNN de-artifacting after JPEG-50 compression (ICCV 2021); and Img2Img regeneration using Qwen-Image (strengths 0.1 and 0.3) and SDXL (strength 0.2). Img2Img regeneration is commonly used for both detail adjustment and watermark removal.

AI-assisted post-processing results (AUROC/AP):

OperationAEDROCC-CLIPSpCon-TOSpCon-ERSpCon-KR
Real-ESRGAN ×2 (after ×0.5)0.6703/0.18870.6580/0.21100.7378/0.25700.8060/0.40370.7661/0.3374
SwinIR ×2 (after ×0.5)0.8468/0.41030.6712/0.22260.8190/0.41210.9214/0.68360.9483/0.7333
Real-ESRGAN ×2 (full size)0.4832/0.14310.6444/0.20400.7480/0.34230.8342/0.55690.8711/0.6092
LaMa inpainting0.9918/0.93450.6664/0.21650.8975/0.77080.9893/0.95610.9974/0.9889
Learned compression0.7223/0.22020.6494/0.20850.5725/0.21090.7395/0.28010.8054/0.3563
SCUNet denoising0.9058/0.52180.6488/0.20450.7428/0.34140.9076/0.63810.9566/0.8010
FBCNN de-artifacting0.8222/0.37160.6528/0.21100.7796/0.35880.8573/0.53540.8883/0.5788
Qwen redraw 0.10.5474/0.16370.6591/0.21130.5468/0.18030.7305/0.25810.7923/0.3001
Qwen redraw 0.30.5138/0.15330.6539/0.21040.5206/0.15990.6776/0.21530.7432/0.2559
SDXL redraw 0.20.4725/0.15390.6476/0.20890.6804/0.21310.7045/0.25310.7502/0.2694
Average0.6976/0.32610.6552/0.21090.7045/0.32470.8168/0.47800.8519/0.5230

The results show that SpCon-ER/KR achieve average AUROC/AP of 0.8168/0.4780 and 0.8519/0.5230, respectively, outperforming all baselines (the strongest baseline, AEDR: 0.6976/0.3261). Notably, KR outperforms the strongest baseline across all 10 operations. Since LatentTracer requires computationally expensive multi-step gradient-based reconstruction, we prioritized evaluation on SD1.5 and SD2.1, where it performs best, to provide a timely response. Across the 10 operations, LatentTracer achieves average AUROC/AP of 0.6375/0.2519, compared with 0.8688/0.5925 for SpCon-ER and 0.8724/0.5886 for SpCon-KR on the same subset.


Manuscript revisions: Added Appendix L, AI-Assisted Post-Processing, detailing the protocols and per-operation results for all 10 AI-assisted operations (Table R10). A summary of these findings has also been added to Section 4.7, Additional Analyses.

AuthorsOpenReview ↗

Re: Application Scope and Deployment (W2)

We have clarified in the abstract that SpCon attributes at the level of the target VAE family, and stated in the introduction and the task definition (Section 3.1) that generators sharing the same VAE cannot be reliably distinguished from VAE-level evidence alone. To further investigate this limitation, we explored including samples from other shared-VAE models in KR reference fitting. We found that discrimination improves substantially when the non-belonging references include images generated by competing models within the same VAE family.


Specifically, we evaluated three models sharing the same VAE in the SD3.5 family: SD3.5-Medium (M), SD3.5-Large (L), and SD3.5-Large-Turbo (LT). We considered all six configurations with two or three shared-VAE models in the candidate set. In each comparison, the left side of "vs" denotes the target model and the right side denotes the non-target model(s). SpCon-KR+Comp extends KR by including samples from competing shared-VAE models in its known non-belonging references.

Shared-VAE model discrimination (AUROC / AP):

SD3.5 FamilySpCon-ER (Original)SpCon-KR (Original)SpCon-KR+CompAUROC gain [95% CI]AP gain [95% CI]
M vs L0.5006 / 0.50010.5518 / 0.55970.7420 / 0.71110.1902 [0.1395, 0.2393]0.1514 [0.0994, 0.2040]
M vs LT0.7037 / 0.74150.6797 / 0.72020.9353 / 0.93880.2555 [0.2227, 0.2870]0.2186 [0.1891, 0.2489]
L vs LT0.7718 / 0.81860.7289 / 0.77970.9531 / 0.95880.2242 [0.1930, 0.2559]0.1791 [0.1507, 0.2047]
M vs L+LT0.6022 / 0.42050.6158 / 0.46480.8043 / 0.64680.1885 [0.1539, 0.2243]0.1820 [0.1320, 0.2326]
L vs M+LT0.6891 / 0.54140.6514 / 0.50170.8307 / 0.70960.1794 [0.1523, 0.2052]0.2079 [0.1655, 0.2425]
LT vs M+L0.6385 / 0.51730.6966 / 0.56500.9431 / 0.89040.2465 [0.2203, 0.2738]0.3254 [0.2833, 0.3655]

The results show substantial improvements in both AUROC and AP across all six configurations. AUROC exceeds 0.93 for M vs LT, L vs LT, and LT vs M+L, while M vs L remains the most challenging case (0.7420). This extension does not change the method, but incorporates competing shared-VAE models into KR's known non-belonging references. It requires access to samples from individual models within the same VAE family. Since shared-VAE models are often released and maintained by the same organization, this requirement can be satisfied in internal auditing scenarios.


Manuscript revisions: The abstract, end of Section 1 (Introduction), Section 3.1 (revised Belonging definition and Attribution Goal to reflect VAE-family attribution, with a new Attribution Class paragraph), Section 5 (Limitations), and Section 6 (Conclusion) now explicitly clarify the VAE-family attribution scope. Appendix H, Table 29 has been extended with SpCon-KR+Comp results and confidence intervals for the performance gains, together with a new Known Shared-VAE References paragraph.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Rebuttal Summary (Submission 6648)

Dear Program Chairs, Senior Area Chairs, Area Chair, and Reviewers,

We sincerely thank all reviewers for their careful assessments and valuable suggestions. To facilitate the final evaluation, we briefly summarize the strengths recognized by the reviewers, our key responses during the rebuttal, and the reviewers' feedback. All revisions are highlighted in blue in the revised manuscript.

1. Strengths Recognized by Reviewers

  1. Clear methodological motivation and sound theoretical rationale (zJ3c, ZPim, ZXx1).
  2. Comprehensive experiments and substantial performance improvements under post-processing (zJ3c, ZXx1).
  3. Controlled comparisons and ablations supporting the independent contribution of the feature representation (zJ3c, ZXx1).
  4. Methodological simplicity and low computational overhead (zJ3c, ZXx1).

2. Key Responses During Rebuttal

We carefully addressed each reviewer concern and conducted targeted additional experiments and analyses:

  1. Novelty Beyond AEDR (ZPim, ZXx1): We clarified that SpCon's core contribution is not merely a frequency-domain transformation, but rather preserving relative residual changes across channel–frequency structures and modeling their reference distributions. Under strictly matched reference data, fitting conditions, and modeling backends, SpCon-ER achieves an AUROC of 0.9387, substantially outperforming AEDR-Mix-ER (0.7444), demonstrating the independent contribution of the structured feature representation (Sec. 4.7, App. I).

  2. Robustness Under Complex Post-Processing (zJ3c, ZXx1): We added 10 AI-assisted post-processing operations and 15 sequential-processing evaluation settings. SpCon-KR outperforms the strongest baseline across all 10 AI-assisted operations. Under matched chain-reference access, SpCon-KR achieves higher AUROC than both native AEDR and reference-matched AEDR-Mix-KR across all 15 settings. Composition-aware fitting also substantially improves performance on certain unfitted three-operation chains (e.g., 0.6645→0.9171) (Sec. 4.5, Apps. J, L).

  3. Attribution Scope and Applicability (zJ3c, ZPim, ZXx1): We clarified the scope of VAE-family-level attribution and the necessary conditions for model-level discrimination. Additional shared-VAE experiments demonstrate AUROC improvements across all six configurations when reference samples from competing models within the same VAE family are incorporated. We further clarified that model-level attribution requires a defined candidate set and reliability validation on held-out data (Sec. 3.1, App. H).

  4. Experimental Protocol and Calibration Reliability (ZPim, zJ3c): We added documentation of data splits and configuration freezing, reference-resampling stability experiments, threshold calibration under deployment-specific reference-access constraints, and sensitivity analyses across 50 random calibration splits. We also explicitly distinguished calibration-split stability from FPR control under distribution shifts (Sec. 4.1, Apps. B, D, E).

3. Reviewer Feedback

  1. zJ3c: Explicitly acknowledged that our responses and additional experiments had substantially addressed their concerns and maintained the positive Weak Accept assessment. Both follow-up suggestions have also been incorporated into the revised manuscript.

  2. ZXx1: Acknowledged that most concerns had been addressed, while retaining reservations regarding novelty and robustness under complex post-processing. We provided further clarifications supported by controlled experimental evidence in our subsequent response (see Points 1 and 2 in Section 2).

  3. ZPim: As of this writing, we have not received further feedback on our rebuttal. Nevertheless, we have addressed all four major concerns regarding novelty, attribution scope, calibration access, and experimental stability through detailed responses, additional experiments, and corresponding manuscript revisions.

We believe that our rebuttal has substantially addressed the reviewers' main concerns and further strengthened the paper's novelty justification, empirical evidence, and rigor. We respectfully hope that the PCs, SACs, and AC will give due consideration to the additional evidence and substantive manuscript revisions during the final evaluation.

We sincerely thank the PCs, SACs, AC, and all reviewers once again for their valuable time, constructive feedback, and thoughtful consideration.

All Authors of Submission 6648

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.