acceptodds
Under review as a conference paper at ICLR 2027

Can Heterogeneous Artifacts Coexist? RGSA: Role-Specialized Global–Selective Adaptation for Mitigating Artifact Feature-Space Collapse in Unified Fake Image Detection

Abstract

Unified Fake Image Detection (FID) aims to detect forgeries across four image forensic domains within a single model. However, joint multi-domain training substantially degrades the discriminative capability originally learned for individual domains, leading to a collapse of the artifact feature space. To uncover the underlying mechanism, we quantify heterogeneous conflicts during joint optimization and establish an evidence chain linking artifact heterogeneity to feature-space collapse. Our analysis reveals that cross-domain updates exhibit mixed and asymmetric interactions, indicating that different artifacts have distinct sharing scopes. We further find that organizing feature-space capacity according to these sharing scopes can effectively mitigate such collapse. Building on these findings, we propose Role-Specialized Global-Selective Adaptation (RGSA), which organizes the adaptation capacity into two pathways with distinct roles to accommodate different sharing scopes. The global pathway captures representations shared across all domains, whereas the selective pathway models partially shared and domain-specific artifacts. Under the four-domain unified protocol of OpenMMSec, RGSA effectively mitigates cross-domain conflicts and artifact feature-space collapse, and outperforms 23 state-of-the-art methods, validating both the proposed mechanism and the role-specialized design.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Fake Image Detection, Unified Image Forensics, Artifact Heterogeneity, Feature Space Collapse

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_PMowOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper makes three main contributions. First, it analyzes cross-domain conflicts in unified fake image detection using gradient similarity, relative cross-domain harm, and long-term transfer to characterize how heterogeneous forensic domains interfere with each other during joint training. Second, it argues that different forensic artifacts have different sharing scopes and proposes that feature-space collapse is mainly caused by a mismatch between these sharing scopes and the organization of model capacity. Third, it proposes RGSA, which combines a globally shared OFT pathway with routed LoRA experts to model globally shared, partially shared, and domain-specific artifacts, and evaluates the method on the OpenMMSec benchmark.

Critical strengths

This work provides a systematic empirical analysis of cross-domain interactions in unified fake image detection, rather than only reporting final performance. The combination of gradient compatibility, relative cross-domain harm, and long-term transfer offers a relatively comprehensive view of how different forensic domains affect each other during joint optimization.

Critical weaknesses
  1. The issue of conflicts among heterogeneous data addressed in this paper has been repeatedly studied in previous works, which the authors also acknowledge in the Introduction and Related Work. The authors claim that their novelty lies in revealing the mechanism behind such conflicts. However, in practice, the authors still investigate the gradient conflicts among heterogeneous data and the influence of parameter updates on the amount of loss reduction in the same way as CS-DFD. To truly investigate the causes of these conflicts, the authors should at least examine the differences among different types of data and explain why gradient conflicts and optimization conflicts occur. The authors overstate the contribution of the current work and do not provide any substantial new insights.

  2. The core argument of the paper is that “feature-space collapse does not simply arise from insufficient capacity, but primarily from a mismatch between capacity organization and the distinct sharing scopes.” The former is already a common understanding, while the latter is not convincingly demonstrated by the authors. Even if the use of routers can be regarded as a better form of capacity organization, the authors still fail to prove the relationship between the router and the sharing scope.

  3. The two modules in this work are actually very simple, but I believe there are serious problems with the description of the method. The global module is updated using all data and keeps the relevant parameters orthogonal, but maintaining parameter orthogonality cannot guarantee the learning of shared artifacts.

  4. The Selective Pathway is essentially a LoRA-MoE and does not involve any relevant constraints. The router designed by the authors cannot distinguish different domains at all and does not know the artifact types. Therefore, it cannot guarantee that conflicting data will be assigned to different routers after using the router.

  5. There are many unclear symbol descriptions in the formulas of the paper. W0W_0, WOW_O, QQare not defined.

Questions for authors

I would appreciate clarification on whether the authors can provide a more concrete explanation of why conflicts arise among different forensic domains, beyond measuring gradient disagreement and cross-domain harm. In particular, can the authors identify what specific differences in artifacts, representations, or optimization requirements lead to these conflicts, and provide evidence that the proposed sharing scopes are meaningfully aligned with such conflict structures? A convincing analysis along these lines would substantially strengthen the paper and could affect my overall evaluation.

5 comments
AuthorsOpenReview ↗

Clarification and Response to W1

We thank the reviewer for recognizing our systematic analysis of cross-domain interactions in Unified FID. Your question is closely related to W1–W4, we will address together below.

We believe the concerns raised in W1 and W2 may stem from misunderstandings of the Unified FID task and existing research on artifact heterogeneity. We therefore first clarify the current research landscape and key challenges of Unified FID.

Clarification: Existing Research Gaps in Unified FID and Our Contributions

As explicitly stated in our Introduction, Unified FID focuses on “Unified FID aims to perform unified real–fake discrimination across forensic domain, while the underlying forgery method is unknown under realistic forensic scenarios.”

In this context, both academia and industry have investigated artifact heterogeneity and cross-domain interference. SICA [1]^{[1]} characterizes artifact feature space collapse as compressing “domain-specific, high-dimensional artifact features into a unified or domain-indiscriminative, low-dimensional representation space.” UniShield [3]^{[3]} notes that “models trained on mixed forgery datasets often experience performance degradation caused by domain conflicts.” Meanwhile, tech report [2]^{[2]} finds that certain IMDL manipulations involving generative editing benefit from AIGC data, demonstrating positive transfer alongside cross-domain interference.

However, existing studies have not systematically investigated conflicts among heterogeneous artifacts or distinguished whether cross-domain interference stems from conflicting optimization requirements or insufficient representation capacity. Therefore, we would appreciate it if you could provide specific references establishing that “feature space collapse does not simply arise from insufficient capacity” is already a common understanding in the literature.

Our work quantifies relative cross-domain harm at the optimization level,based on the results in Tables 2 and 10,

ModelRAcc
LoRA (r = 16)0.2080.854
LoRA (r = 64)0.2090.851
LoRA (r = 144)0.2100.855
LoRA+8 routers (r = 16)0.1880.859

LoRA adapters with different capacities exhibit similar cross-domain harm and detection performance, whereas separating update pathways at comparable capacity reduces harm and improves performance. This reveals “feature space collapse does not simply arise from insufficient capacity, but primarily from a mismatch between capacity organization and the distinct sharing scopes of heterogeneous artifacts” firstly, as accurately reflected in Contributions 1 and 2.

The mechanism we identify is explicitly stated in Sec. 3.5: “Reveal the mechanism which heterogeneous artifacts progressively lead to feature space collapse during joint optimization in FID”.

Further responses to your remaining concerns are provided below.

[1] Du et al. Can We Build a Monolithic Model for Fake Image Detection? SICA: Semantic-Induced Constrained Adaptation for Unified-Yet-Discriminative Artifact Feature Space Reconstruction. ICML, 2026.

[2] Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage. arXiv, 2026.

[3] Huang, Q., Xu, Z., Zhang, X., Yu, X., & Zhang, J. (2026). Unishield: An adaptive multi-agent framework for unified forgery image detection and localization. CVPR, 2026.

W1 :Is cross-domain conflict analysis distinct from CS-DFD?

Yes. The novelty of our cross-domain conflict analysis is clarified below.

We acknowledge that both CS-DFD and our work use loss changes induced by parameter updates to analyze optimization interference. However, CS-DFD focuses on coordinating gradient updates between samples, whereas our work quantifies relative optimization interference across heterogeneous forensic domains by comparing loss changes induced by cross-domain and within-domain sample-pair updates, and further investigates its cumulative relationship with feature space collapse.

Sec. 3.3 shows that instantaneous cross-domain gradients in Unified FID are nearly orthogonal rather than ideally aligned, while gradient similarity alone cannot capture actual cross-domain harm or its cumulative effects.

AuthorsOpenReview ↗

Response to W2,W3,W4,W5

W2,W3,W4: Do RGSA's two pathways learn their intended representations?

Yes. As explained in our Clarification, RGSA is “simple yet effective” because it identifies and directly addresses the key cause of artifact feature space collapse in Unified FID: conflicting update requirements of heterogeneous artifacts.

Below, we explain how RGSA learns representations with different cross-domain sharing scopes and address your concerns.

  • I believe there are serious problems with the description of the method.

If the “serious problems” is that training on all domains inevitably introduces conflicts, we clarify that “global-shared” in Sec. 4.2 refers to artifact features compatible with all domains, rather than universally beneficial features. Partially shared artifacts are handled through selective expert activation. As explicitly stated in Contribution 3 and Sec. 4, conflict isolation is achieved through “complementary update pathways according to the cross-domain sharing scopes of artifacts”, rather than merely relying on adapter properties.

Using the trained checkpoints under the configuration in Appendix G.1, we sample 10,000 test images per forensic domain from OpenMMSec (fake : real = 3:1). We compute the relative cross-domain harm R when updating only the Global pathway and compare it with models without pathway separation.

ROnly OFTOFT+Router LoRARGSA
Global-Path0.2370.1280.112

To validate these features capture forgery artifacts and the pathway complementarity, using identical trained weights, we compare the performance of (i) features with only Selective activated and (ii) Full − Global features, obtained by subtracting Global-only features (Selective disabled) from Full features (both pathways activated) (Original Table 12, revision Table 16), on the OpenMMSec test set:

Accuracy(%)OFT+Router LoRAAnchor-LoRA+Router LoRAAnchor-OFT+Router LoRA
Global disabled81.1779.5381.48
Global active (Full − Global)85.7385.9086.33

The results indicate that the Global pathway has learned more cross-domain-compatible shared artifacts.

  • The Selective Pathway is essentially a LoRA-MoE and does not involve any relevant constraints.

We introduce a load-balancing loss based on pre-truncation routing probabilities to prevent expert concentration and ensure routing gradients reach all experts, thereby encouraging differentiated learning across experts. Beyond the significant cross-domain differences in expert activation patterns reported in Table 7, we further examine the correlation between expert activation patterns and specific artifact characteristics to verify whether expert differentiation reflects conflicts among heterogeneous artifacts.

We use RGSA trained on OpenMMSec to analyze expert activation similarity across manipulation types using six test subsets: Coverage (IML_cm1) and DEFACTO CopyMove Subset (IML_cm2), RTM CopyMove Subset (Doc_cm), DocTamper Splicing Subset (Doc_sp), generative inpainting datasets CocoGlide(IML_gp), and DiffusionForensic (AIGC). Only fake images from the test sets are included in this analysis.

DomainManipulation TypeDatasetExperts Similarity
SameSameIML_cm1 vs IML_cm20.834
SameCrossIML_cm1 vs IML_gp0.589
CrossSameAIGC vs IML_gp0.830
CrossSameIML_cm1 vs Doc_cm0.569
CrossCrossIML_cm1 vs Doc_sp0.277

The results show that expert activation patterns reflect specific manipulation types: similar manipulations exhibit high expert similarity within the same domain and can share experts across domains, while conflicting artifacts activate different experts for specialized representations.

Specifically, copymove focuses on anomalous similarities between authentic regions [5]^{[5]}, whereas generative inpainting and AIGC focus on synthetic–authentic discrepancies [6,7]^{[6,7]}, leading to distinct update requirements.

[5] Busternet: Detecting copy-move image forgery with source/target localization. ECCV, 2018.

[6] Noise Doesn't Lie: Towards Universal Detection of Deep Inpainting. IJCAI, 2021.

[7] Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection. CVPR, 2024.

Weakness 5:Clarification of Previously Undefined Symbols

Thank you for pointing this out. We acknowledge that these symbols were not explicitly defined at their first appearance.

Specifically, W0W_0 denotes the frozen pretrained linear projection weights, WOW_O denotes the effective weights after OFT adaptation, and QQ denotes the trainable orthogonal transformation. We will clarify their definitions, dimensions, and orthogonality constraints in the revised manuscript.

I thank the authors for their response. After considering it, I have decided to maintain my original score.

Reviewer_LEJ4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies Unified Fake Image Detection (FID) across Deepfake, AIGC, IMDL and Doc. It traces the joint-training degradation ("artifact feature-space collapse") with gradient cosine similarity C, a relative cross-domain harm metric R, and the long-term transfer effect ΔM (Eq. 6). It argues that artifacts have different sharing scopes and that capacity organization matters more than capacity size. The proposed RGSA combines a shared OFT pathway, anchored to a frozen teacher via patch-relation MSE, with 8 Top-2 routed LoRA experts. On OpenMMSec (26 seen / 72 unseen forgery types) it reports 86.4 ACC vs 85.4 for SICA, and lowers mean R from 0.208 to 0.130.

Critical strengths
  1. Attempts to measure which cross-domain directions hurt and by how much, rather than only reporting that joint training hurts.
  2. The in-domain-normalized harm metric R is sensible, and the LoRA-rank sweep (Table 10, r = 8–144) cleanly shows that harm plateaus with rank.
  3. Broad experiments: 23 baselines, four metrics, four backbones, pathway ablations and routing analysis.
Critical weaknesses
  1. The case for unification is not established. The paper justifies a single model because "detectors cannot know the domain of a test image in advance," yet separating faces, documents and natural scenes is near-trivial, and the routing error used to dismiss ensembles is never quantified. The paper's own Table 11 shows LoRA outperform RGSA; AIGC and IMDL drops under joint training. No specialist ensemble (oracle or learned routing) is compared, and no experiment tests whether joint training improves within-domain generalization to unseen forgeries, which is the practically important target.
  2. The collapse-mitigation claim uses a weak reference. Table 5 and table 11 compare unified RGSA against RGSA's own single-domain ensemble (0.8615), which is weaker than LoRA's (0.8687). Calling this row "Best single-domain" is misleading, and mean SIL still drops (0.203 to 0.188).
  3. The evidence chain is overstated. C lies within ±0.05 for every pair, so gradient conflict is marginal. R is positive in 10/12 directions, so the 10/12 sign agreement with ΔM has a high chance baseline, and magnitudes disagree (Doc-AIGC: R = 0.556, ΔM = −0.10 pp; DF-Doc: R = 0.136, ΔM = −0.51 pp). There are no seeds, correlations or intervals, yet Sec. 3.5 makes causal claims.
  4. Gains are small and unverified. The improvement over SICA is about 1.0 ACC with no variance; all Table 7 ablations sit within 0.9 pp. RGSA is not best on Doc AUC/AP or IMDL AUC, and on DeepfakeBench cross-domain (Table 16) it trails SICA (0.8772 vs 0.8945) and Effort (0.8803) without discussion.
  5. Limited novelty and missing baselines. The method composes OFT, MoE-LoRA and relational distillation. Gradient-surgery baselines are absent despite the conflict framing, and shared/specific expert designs are not discussed. The global pathway's role is post hoc: alone, the anchored version (0.7430) is worse than unanchored OFT (0.7626), and the claimed "absorption" of shared artifacts is never tested directly.
  6. Protocol concerns. All baselines run at 224×224 with ForensicHub defaults, which likely handicaps IMDL/Doc methods designed for high-resolution traces; localization, central to IMDL/Doc, is not evaluated. Reporting is also inconsistent: ACC appears as percentages in Tables 3, 6 and 17 but as fractions in Tables 2, 4, 5, 7, 11 and 12, with precision varying between one and four decimals; four-decimal values imply a precision that single-run results without variance cannot support.
Questions for authors
  1. Why is domain separation of face, document and natural scene images considered difficult, and can you quantify the routing error used to dismiss ensembles? Why omit specialist ensemble baselines and experiments on within-domain generalization to unseen forgeries?
  2. Why label RGSA’s weaker single-domain ensemble as “Best single-domain”, and how do you explain the SIL drop despite your collapse-mitigation claim?
  3. With marginal gradient conflict (C within ±0.05), chance-level sign agreement and mismatched magnitudes between R and ΔM, plus missing seeds/intervals, what supports the causal claims in Sec 3.5?
  4. Can you provide variance estimates to validate the ~1.0 ACC gain over SICA? How do you interpret the narrow range of Table7 ablations and RGSA’s cross-domain underperformance on DeepfakeBench?
  5. What core novelty exists beyond combining OFT, MoE-LoRA and relational distillation? Why omit gradient-surgery baselines, and what direct evidence supports the global pathway’s artifact absorption claim?
  6. Does the 224×224 ForensicHub setup unfairly penalize high-resolution IMDL/document baselines? Why skip localization evaluation and use inconsistent ACC formatting/precision without repeated-run variance?
4 comments
AuthorsOpenReview ↗

Clarification and Response to W1

We sincerely thank the reviewer for recognizing our systematic analysis of cross-domain optimization harm and extensive experiments. We would also like to clarify that the observation in Strength 2—“the LoRA-rank sweep (Table 10, r = 8–144) cleanly shows that harm plateaus with rank”—is not entirely accurate. The harm actually plateaus at ranks 16, 64, and 144. We will further clarify this finding in our response to W5 and Q5, in connection with the core architectural innovation of RGSA.

Your Questions Q1–Q6 correspond to Weaknesses W1–W6, respectively, and are addressed below.

We believe your concerns in W1, W2, Q1, and Q2 largely stem from misunderstandings of the Unified FID task and artifact feature space collapse. We therefore first clarify the motivation and research significance of Unified FID.

Clarification: The Research Significance of Unified FID and Artifact Feature Space Collapse

The definition and significance of Unified FID have already been stated in our Introduction: “Unified FID aims to perform unified real–fake discrimination across forensic domain, while the underlying forgery method is unknown under realistic forensic scenarios.” Here, we further clarify its significance : “Forensic detectors cannot know the domain of a test image in advance”[1]^{[1]}. Its practical importance stems from the fact that “In real-world scenarios, it is often impossible to predetermine the type of manipulation (Deepfake, IMDL, AIGC, and Document) present in an image, making unified detection necessary.[2]^{[2]}” For instance, a forged image containing a face could result from any of Deepfake, AIGC, or IMDL.

The importance of Unified FID is recognized by both academia and industry. Academically, SICA [1]^{[1]} and UniShield [5]^{[5]} identify Unified FID as an important research task, focusing on artifact heterogeneity. Practically, existing tech reports [3],[4]^{[3],[4]} highlight its real-world necessity, driven by increasingly diverse forgery techniques and blurred forensic-domain boundaries, and propose corresponding solutions.

SICA explicitly defines artifact feature space collapse as the compression of “domain-specific, high-dimensional artifact features into a unified or domain-indiscriminative, low-dimensional representation space.” It further identifies its direct evidence as “involving extra training data from other subdomains will degrade model performance on the current subdomain.” Accordingly, in Sec. 5.3 (Feature Space Reconstruction Analysis), SICA evaluates feature space reconstruction by examining whether the performance degradation from single-domain to unified training is mitigated.

[1] Du et al. Can We Build a Monolithic Model for Fake Image Detection? SICA: Semantic-Induced Constrained Adaptation for Unified-Yet-Discriminative Artifact Feature Space Reconstruction. ICML, 2026.

[2] Du et al. ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization. NeurIPS, 2025.

[3] GuangJian Team. Venus-DeFakerOne: Unified Fake Image Detection & Localization. arXiv, 2026.

[4] Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage. arXiv, 2026.

[5] Huang, Q., Xu, Z., Zhang, X., Yu, X., & Zhang, J. (2026). Unishield: An adaptive multi-agent framework for unified forgery image detection and localization. CVPR, 2026.

Please find our responses to the remaining concerns below.

W1: Can Different Forensic Domains Be Directly Separated?

No. As clarified above, faces, documents, and natural images may all involve manipulations from different forensic domains. We provide the following evidence to support this point.

SICA (Fig. 1) uses CLIP to construct a semantic manifold and t-SNE to visualize the semantic distributions of forensic subdomains, showing that all domain pairs exhibit partial semantic overlap except Doc–Deepfake. We further quantify the error by training a forensic domain 4 classifier using the same OpenMMSec train/test splits and training configuration as our main experiments, with a DINOv3 backbone and rank-16 LoRA. The classifier achieves only 87% test accuracy, with error rates of 17% for AIGC and 21% for IMDL. Both domains span overlapping semantic categories, including faces and documents, and exhibit substantial mutual confusion.

Label/PredictionDeepfakeAIGCIMDLDoc
Deepfake63,1273,0799230
AIGC1,03362,88411,517370
IMDL1812,26445,981182
Doc077739733,439

(Response to W1 continues in the next post.)

AuthorsOpenReview ↗

Response to W1,W2,W3

(W1 Continued)

  • Why omit specialist ensemble baselines and experiments on within-domain generalization to unseen forgeries?

Regarding specialist ensembles, UniShield$^{[5]}$ is, to our knowledge, the only existing Unified FID ensemble baseline, but its code and weights are unavailable. Building a comparable ensemble also requires addressing the “barrel effect” noted by SICA, complicating fair comparisons. We will clarify these reasons in the revision.

Our experiments already evaluate generalization to unseen manipulation types under the unified detection setting. As detailed in SICA (Sec. 3), OpenMMSec contains 98 forgery types, of which 72 are unseen during training and reserved for testing, explicitly evaluating generalization to unseen forgeries.

W2: Is the reported evidence sufficient to support RGSA's mitigation of artifact feature space collapse?

Yse, as clarified above, the evidence is whether the performance degradation from single-domain to unified training is mitigated(SICA$^{[1]}$).

As shown in the table below(Table.11), RGSA exhibits less degradation than LoRA in both ACC and SIL across all four domains and on average, indicating better preservation of domain-specific discriminative information and more effective mitigation of artifact feature space collapse. We will move Table 11 to Sec. 5.4 to more clearly present the evidence for mitigating artifact feature-space collapse.

Acc-DegenerationDeepfakeAIGCIMDLDocMean
LoRA Degeneration-0.38%-4.25%-1.82%+0.55%-1.48%
RGSA Degeneration+3.43%-2.49%-1.76%+1.73%+0.22%
SIL DegenerationDeepfakeAIGCIMDLDocMean
LoRA Degeneration-0.022-0.144-0.056-0.010-0.058
RGSA Degeneration+0.061-0.076-0.046-0.001-0.015

The original label “Best Single Domain” of four independently trained RGSA models represents single domain feature spaces unaffected by joint-training-induced collapse. To avoid ambiguity, we will rename it “RGSA Single Domain”.

W3: Is the Evidence Chain for Cross-Domain Harm Analysis Reliable?

Yes. In Sec. 3, we provide substantial evidence supporting the distinct meanings of the three metrics and their interconnections within a coherent evidence chain.

To address your concern, we further clarify the reliability. First, our conventional gradient similarity analysis reveals nearly orthogonal instantaneous gradients rather than ideally aligned updates, while it cannot capture actual optimization interference or cumulative effects. We therefore introduce R to quantify asymmetric cross-domain optimization interference by comparing the loss changes induced by cross-domain and in-domain parameter updates. Finally, the consistency between R and ΔM supports our conclusion that cross-domain optimization interference leads to artifact feature space collapse during unified training.

We would like to clarify that R is positive in 9 of the 12 directions, rather than 10 as stated by the reviewer. More importantly, its correspondence with ΔM\Delta M captures both positive and negative transfer, rather than merely reflecting the majority of positive R values. å As discussed in Sec. 3.4, their magnitude differences arise because R measures instantaneous interference from individual updates, whereas ΔM\Delta M reflects cumulative effects throughout joint training, further influenced by imbalanced domain sample sizes (Deepfake:AIGC:IMDL = 25:13:31:13).

We conducted experiments with two additional random seeds. Averaged over three runs, R and ΔM\Delta M show consistent transfer relationships in 11 of 12 directions, with a significant negative Pearson correlation (r=−0.7174r=-0.7174, p=0.0086p=0.0086).

Source \ Target DomainDF(R)AIGC(R)IMDL(R)DOC(R)DF(ΔM(%))AIGC(ΔM(%))IMDL(ΔM(%))DOC(ΔM(%))
DF—0.356-0.0740.169—-0.217+0.230-1.443
AIGC-0.026—0.3460.199+0.290—-1.863−0.437
IMDL-0.0450.759—0.158−0.510−4.770—−0.490
DOC0.0400.5340.022—−0.373−0.050−0.183—
AuthorsOpenReview ↗

Response to W4,W5

W4: Are RGSA’s performance gains and generalization capabilities reliable?

Yes. We address the reviewer’s concerns point by point below.

  • Gains are small and unverified

Our gains extend beyond the 1.0 pp ACC improvement over SICA. Appendix Tables 13–15 also report gains of 1.5, 3.5, and 3.0 pp in AUC, AP, and F1, respectively. We prioritize ACC for consistency with SICA and further report variance estimates across three random seeds, demonstrating stable performance gains.

Metrics(%)ACCAUCAPF1
RGSA86.27 ± 0.2188.90 ± 0.1782.27 ± 0.2672.84 ± 0.27
SICA85.3787.5478.9869.73
  • RGSA is not best on Doc AUC/AP or IMDL AUC

The Unified FID aim to improve overall performance across all four forensic domains, rather than maximize performance in each individual domain(SICA$^{[1]}$). As discussed in Appendix C, some baselines are specialized for specific domains and may retain advantages on their respective tasks. Nevertheless, RGSA achieves the best overall performance across all four metrics, falling slightly behind only on a few individual metrics.

  • Table 7 ablation results

As stated in Contribution 3, RGSA's core innovation is “complementary update pathways according to the cross-domain sharing scopes of artifacts”, rather than adapter combinations. All variants in Table 7 preserve this design, while variants with weaker Global or Selective role specialization perform worse.

ModelComplementary update pathwaysGlobal RoleSelective RoleMean Acc(%)
G-AnchorOFT+S-OFTTrueBestWorst85.46
G-LoRA+S-LoRATrueWorstBest85.86
G-OFT+S-LoRATrueMedianBest85.96
RGSA(G-AnchorOFT+S-LoRA)TrueBestBest86.37
  • RGSA’s cross-domain underperformance on DeepfakeBench

DeepfakeBench uses single-domain training, without the multi-domain optimization conflicts RGSA targets, limiting its architectural advantages. In contrast, GenImage presents similar cross-generator heterogeneity to Unified FID, where RGSA achieves superior performance. We will include this clarification in the revised manuscript.

W5: Can RGSA’s core innovation be sufficiently validated?

Yes. We address your concerns from two perspectives: our core innovation and its empirical validation.

First, we clarify that our contribution is not merely combining adapters, but systematically quantifying cross-domain optimization interference, revealing conflicting update requirements of heterogeneous artifacts as a key cause of feature space collapse, and accordingly designing role specialized Global–Selective pathways. Based on the results in Tables 2 and 10,

ModelRAcc
LoRA (r = 16)0.2080.854
LoRA (r = 64)0.2090.851
LoRA (r = 144)0.2100.855
LoRA+8 routers (r = 16)0.1880.859

LoRA adapters with different ranks exhibit similar cross-domain harm and detection performance. At comparable capacity, separating the update pathways effectively reduces cross-domain harm and improves performance, suggesting that artifact feature space collapse stems from conflicting update requirements rather than insufficient representation capacity. Thus, RGSA's simple yet effective design lies in identifying and directly addressing the key cause of artifact feature space collapse in Unified FID.

Sec. 4.2 establishes the complementary roles of the Global–Selective pathways, while Secs. 4.3–4.4 detail their role-specialized implementations. Notably, Global is explicitly designed to learn shared representations rather than maximize standalone detection performance, which is not post hoc .

(Response to W5 continues in the next post.)

AuthorsOpenReview ↗

Response to W5,W6

(W5 Continued)

  • shared/specific expert designs are not discussed

Using the trained checkpoints under the configuration in Appendix G.1, we sample 10,000 test images per forensic domain from OpenMMSec (fake : real = 3:1). We compute the relative cross-domain harm R when updating only the Global pathway and compare it with models without pathway separation.

ROnly OFTOFT+Router LoRARGSA
Global Path0.2370.1280.112

The results indicate that the Global pathway learns more cross-domain-compatible shared representations.

The Selective pathway is designed to route conflicting artifacts to different experts while sharing experts for similar artifacts. To validate this, we use RGSA trained on OpenMMSec to analyze expert activation similarity across manipulation types using six test subsets: Coverage (IML_cm1) and DEFACTO CopyMove Subset (IML_cm2), RTM CopyMove Subset (Doc_cm), DocTamper Splicing Subset (Doc_sp), generative inpainting datasets CocoGlide(IML_gp), and DiffusionForensic (AIGC). Only fake images from the test sets are included in this analysis.

DomainManipulation TypeDatasetExperts Similarity
SameSameIML_cm1 vs IML_cm20.834
SameCrossIML_cm1 vs IML_gp0.589
CrossSameAIGC vs IML_gp0.830
CrossSameIML_cm1 vs Doc_cm0.569
CrossCrossIML_cm1 vs Doc_sp0.277

The results show that expert activation patterns reflect specific manipulation types: similar manipulations exhibit high expert similarity within the same domain and can share experts across domains, while conflicting artifacts activate different experts for specialized representations.

  • what direct evidence supports the global pathway’s artifact absorption claim?

To validate pathway complementarity, using identical trained weights, we compare the performance of (i) features with only Selective activated and (ii) Full − Global features, obtained by subtracting Global-only features (Selective disabled) from Full features (both pathways activated)(Original Table 12, revision Table 16), on the OpenMMSec test set :

Accuracy(%)OFT+Router LoRAAnchor-LoRA+Router LoRAAnchor-OFT+Router LoRA
Global disabled81.1779.5381.48
Global active (Full − Global)85.7385.9086.33

The performance improvement with Global activated supports the complementarity between the two pathways. We will additionally report per-domain results in the Appendix.

  • Why omit gradient-surgery baselines

Our analysis of gradient similarity CC in Sec. 3.3 shows that gradient alignment cannot adequately capture actual cross-domain optimization harm. We evaluate two representative gradient-surgery methods, PCGrad and CAGrad, using DINOv3 with rank-16 LoRA adaptation:

Metric(%)AccAUCAPF1
LoRA-Rank1685.3988.2181.571.37
With PCGrad84.6988.3978.9570.89
With CAGrad84.8288.4379.9071.27

Both methods show an overall performance decline, consistent with our finding that cross-sample optimization interference arises from artifacts with different sharing scopes during training. Such fine-grained conflicts cannot be adequately resolved by gradient surgery alone.

W6: Does ForensicHub setup unfairly penalize high-resolution IMDL/document baselines?

Not entirely. Our experiments follow the standard Unified FID task definition and evaluation protocol, focusing on image-level forgery detection across four forensic domains. A consistent input resolution ensures fair and consistent cross-domain comparisons.

To address the reviewer's concern, we further evaluated each method at its originally adopted input resolution. The results show that higher resolutions benefit the relevant baselines. Meanwhile, RGSA also benefits and maintains the best overall performance. We will include these results in the Appendix.

ModelInput SizeACC(%)AUC(%)AP(%)F1(%)
DFD-HR (Deepfake)44886.5487.1779.6571.06
RGSA(Ours)44887.0190.8085.4975.88
Turfor(IMDL)51279.0573.8561.1355.50
Mesorch(IMDL)51279.2079.9267.3358.75
Relayformer(IMDL)51278.7978.6766.1358.84
  • Why skip localization evaluation and use inconsistent ACC formatting/precision without repeated-run variance?

All experimental results are based on our local reproductions and evaluations, as detailed in Appendix G.1. Differences in ACC decimal precision reflect table formatting considerations, with similar variations also present in SICA. We will adjust the reported precision where necessary to improve clarity.

Reviewer_Rg58OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the problem of cross-domain negative transfer in unified fake image detection and proposes RGSA, which models cross-domain shared and selective shared features through a global OFT branch, an input-conditioned LoRA expert branch, and relational anchoring, respectively. The authors further analyze the interactions between different forensic domains using gradient cosine similarity, relative cross-domain damage, and long-term transfer results.

Critical strengths

The core explanation of the paper involves two levels of inference: (a) the cross-domain update conflicts can reflect the scope of sharing of heterogeneous forgery traces; (b) the performance improvement of RGSA demonstrates the effectiveness of the mechanism for allocating adaptation paths based on the scope of sharing.

Critical weaknesses
  1. Based on phenomena such as cross-domain gradient interaction and differences in expert activation patterns, the paper infers that forgery traces exhibit different scopes, including globally shared, partially shared, and domain-specific traces. However, since expert activation varies across domains, it may also reflect differences in input distributions or domain labels, and does not necessarily indicate that the experts have learned interpretable forgery traces.
  2. The paper supports the collapse analysis using changes in accuracy, SIL, and visualizations between single-domain and joint training. Current results show that RGSA can mitigate some of the declines in metrics, but performance varies across domains. For example, while joint RGSA improves accuracy for Deepfake and Doc compared to single-domain models, it reduces accuracy for AIGC and IMDL. Its average SIL also remains lower than that of single-domain training results.
Questions for authors
  1. Could the authors provide additional validation at the feature level, such as identifying the forged regions or frequency cues that different paths focus on, and testing whether these cues transfer across similar domains?
  2. Does collapse specifically refer to a decrease in class separability, confusion of domain features, or a decline in detection performance?
3 comments

We sincerely thank the reviewer for the positive and constructive feedback. We especially appreciate your recognition of our cross-domain interaction analysis and the effectiveness of the mechanism for allocating adaptation paths based on the scope of sharing. We would like to address your remaining concerns below.

Q1 & W1: Do the experts learn forgery-specific features, and are these features transferable across domains?

We address this question from two perspectives: (1) the experts learn forgery specific traces rather than merely capturing domain distributions; and (2) these learned forgery traces are transferable across domains.

As stated in Sec. 4.4, “Cross-domain samples with similar artifacts can reuse experts, while differentiated activation can reduce competition among incompatible artifacts.” Accordingly, different expert activation combinations may correspond to distinct forgery traces. However, the combinatorial space across 24 layers is too large to directly identify trace specific patterns. We therefore use different manipulation types as proxies for distinct forgery traces and analyze expert activation similarity among samples of different manipulation types.

We use RGSA trained on OpenMMSec to analyze expert activation similarity across manipulation types using six test subsets: Coverage (IML_cm1) and DEFACTO CopyMove Subset (IML_cm2), RTM CopyMove Subset (Doc_cm), DocTamper Splicing Subset (Doc_sp), generative inpainting datasets CocoGlide(IML_gp), and DiffusionForensic (AIGC). Only fake images from the test sets are included in this analysis.

DomainManipulation TypeDatasetExperts Similarity
SameSameIML_cm1 vs IML_cm20.834
SameCrossIML_cm1 vs IML_gp0.589
CrossSameAIGC vs IML_gp0.830
CrossSameIML_cm1 vs Doc_cm0.569
CrossCrossIML_cm1 vs Doc_sp0.277

The results indicate that the artifact features learned by the experts correspond to specific manipulation types and exhibit cross-domain transferability. We will include additional image pairs from different domains with similar expert activation patterns in the Appendix.

Q2&W2: Definition of Artifact Feature Space Collapse and Evidence of Its Mitigation

SICA 1^{1} explicitly defines artifact feature space collapse as the compression of “domain-specific, high-dimensional artifact features into a unified or domain-indiscriminative, low-dimensional representation space.” It further identifies its direct evidence as “involving extra training data from other subdomains will degrade model performance on the current subdomain.” Accordingly, SICA's Sec. 5.3 (Feature Space Reconstruction Analysis), evaluates feature space reconstruction by examining whether the performance degradation from single-domain to unified training is mitigated.

As shown in the table below(Table.11), RGSA exhibits less degradation than LoRA in both ACC and SIL across all four domains and on average, indicating better preservation of domain-specific discriminative information and more effective mitigation of artifact feature space collapse. We will move Table 11 to Sec. 5.4 to more clearly present the evidence for mitigating artifact feature space collapse.

Acc-DegenerationDeepfakeAIGCIMDLDocMean
LoRA Degeneration-0.38%-4.25%-1.82%+0.55%-1.48%
RGSA Degeneration+3.43%-2.49%-1.76%+1.73%+0.22%
SIL DegenerationDeepfakeAIGCIMDLDocMean
LoRA Degeneration-0.022-0.144-0.056-0.010-0.058
RGSA Degeneration+0.061-0.076-0.046-0.001-0.015

[1] Du et al. Can We Build a Monolithic Model for Fake Image Detection? SICA: Semantic-Induced Constrained Adaptation for Unified-Yet-Discriminative Artifact Feature Space Reconstruction. ICML, 2026.

Thanks for your time and effort in reviewing our paper again, it seems that your reply to our response has been posted in Reviewer LEJ4's discussion thread.

We sincerely appreciate your recognition of our cross-domain interaction analysis, the effectiveness of allocating adaptation pathways according to artifact-sharing scopes in mitigating feature space collapse, and the experts' ability to learn transferable forensic artifacts.

Thank you for the feedback. Your response addresses my concerns, and I will retain my original score.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your time and effort in reviewing our paper.

As the discussion period is drawing to a close, we hope that our responses have adequately addressed your current concerns. If you have any further questions or concerns about current responses, we would be more than happy to discuss them further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.