acceptodds
Under review as a conference paper at ICLR 2027

FakeCueS: Learning Scene-Conditioned Cues for AI-Generated Image Detection

Abstract

Advances in generative models enable high-fidelity image synthesis across diverse scenes, characterized by visual style and subject content, challenging reliable AI-generated image detection. Recent multimodal detectors identify visual evidence for their decisions but do not explicitly model its scene-dependent diagnostic value: textures or structures that suggest AI generation in photographs may reflect ordinary artistic expression in paintings, potentially leading to misclassification. To address this limitation, we introduce FakeCueS, a framework that uses scene context to guide forensic cue generation and evaluation. Predicted scene representations guide the model to generate context-specific cues. An online cue evaluator verifies visual support and groups semantically related cues to assess their source association within each scene. These assessments provide set-level rewards for group-relative policy learning alongside image-level supervision, without reference rationales or region annotations. We further introduce RedScenes, a dataset of 168,688 social-media images spanning 70 scenes. FakeCueS achieves 91.53% accuracy on RedScenes, exceeding the strongest evaluated zero-shot baseline by 12.25%. Without target-domain fine-tuning, it achieves 94.49% and 83.09% accuracy on AnomReason-Deepfake and Chameleon, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, explainability, multimodal learning, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tdttOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper’s premise is that a forensic cue means different things in different scenes. Smooth skin or flawless geometry points to AI generation in a photograph, but is ordinary craft in a painting or a 3D render. Detectors trained mostly on photographs therefore misread stylized content, and the paper argues that scene context must condition both which cues a model generates and how much weight each cue carries. There are two main contributions: 1. Redscenes dataset:168,688 Xiaohongshu images with human-verified Real/Fake labels and Gemini-generated style and subject labels. 2. FakeCueS, a Qwen detector that conditions both source classification and cue generation on a predicted scene representation, and trains open-ended forensic cues with GRPO using a set-level reward combining visual support, scene-conditioned source association, and novelty, without requiring reference rationales.

Critical strengths
  1. Well-motivated problem and useful dataset. Cue validity across styles is a real gap. The paper documents that existing explainable detectors and their training sets are photo-centric. RedScenes offers real-world social-media data with human-verified source labels, a reasonable watermark-screening pipeline, and joint style and subject annotations. It covers Real artwork, design and 3D content.

  2. Strong reported numbers and transparent reporting. In-domain and AnomReason-Deepfake numbers are strong. The appendix candidly reports negative findings, including near-unanimous cue verdicts, cue voting causing a net loss relative to the head, and ablations that do not isolate reward terms.

Critical weaknesses
  1. Dataset is good but I am worried about the potential label leakage in scene annotation. Authors use Gemini 3 Pro to assign style and subject labels with the image and its confirmed source label together. Style boundaries are often ambiguous. For example, a photorealistic AI-generated portrait could reasonably be labeled either “photography” or “3D rendering”. Knowing that an image is Fake may bias the annotator toward “3D”, while a Real image with similar appearance may be labeled “photography”. In some extent, style labels may therefore encode source information. This is supported by the highly skewed source proportions across styles: about 82% of 3D images are Fake, compared with about 22% of Design images. This matters because FakeCueS predicts style and feeds it into source classification. The model could exploit style as a proxy for the source label rather than learning scene-dependent cue validity. The authors should report how scene labels change when the source label is withheld, for example by re-annotating a subset without it and measuring agreement. They should also evaluate the model with source-agnostic scene labels.

  2. Another concern is about whether performance gains can be attributed to the proposed components. Firstly, the paper does not compare variants with and without scene conditioning, scene supervision, or scene-conditioned association under an otherwise identical setup. Rows in Table 3 differ in training steps, prompts and prediction paths, and the authors acknowledge that the table does not isolate individual components. Besides, Classifier with Scene uses no cue generation or RL, yet it reaches 91.73% on RedScenes, essentially the same as FakeCueS (91.53%). The cue generation and reward design therefore contribute little to detection accuracy. Also, all baselines in Table 1 are off-the-shelf or zero-shot, while FakeCueS is trained on RedScenes. The 12.25 pp margin therefore largely reflects in-domain training.

Questions for authors
  1. The ethics statement conditions any release of RedScenes on legal review. Given that the dataset is presented as a primary contribution and that the main results are not reproducible without it, can you state what is likely to be released and in what form (images, URLs plus labels, or labels only)? Will training and evaluation code be released?

  2. The reported predictions come from the source head, which does not use the generated cues. In what sense do the cues explain the model’s decision? On the 790 test images where cue voting and the source head disagree, which output should a user trust, and why?

  3. Figure 5(b) justifies a Real verdict through “conscious artistic conception”, which the cue-generation prompt explicitly forbids. How often does this kind of reasoning occur in test outputs?

  4. Have you evaluated the accuracy or usefulness of the generated cues, through a human study, an automatic faithfulness test, or a comparison with explanations from baselines such as FakeVLM or AIGI-Holmes? If not, what evidence supports the claim that FakeCueS improves explainable detection?

4 comments
AuthorsOpenReview ↗

Response to Reviewer tdtt (Part 1 of 3)

Thank you for your detailed comments on scene annotation, attribution of performance gains, and explanation validity. We first apologize for an error in our initial implementation and the additional review burden it caused. Our checks identified that cue clusters and historical statistics were not correctly shared across GPUs during training. We have corrected the implementation, rerun training and evaluation, updated the full-model results, and expanded OOD evaluation from two datasets to nine. We address your concerns below.

W1: Source information in scene annotation and potential label leakage

Thank you for highlighting how source labels could affect ambiguous boundaries such as photography versus 3D. We first clarify the pipeline: scene annotations are used only during training for scene-alignment supervision and scene-specific cue statistics; inference requires neither scene labels nor source labels. The model predicts style and subjects from the image and combines these predictions with visual features to form the scene representation used for source classification and cue generation.

To examine potential bias in the training annotations, we test whether source labels change scene annotations and whether scene-specific source proportions can explain detection performance.

A. Most scene annotations remain unchanged when source labels are withheld.

We sample five images from each of 70 scenes, totaling 350 images, and use the scene-annotation model to annotate each image with and without its source label. All other inputs and decoding settings are held fixed.

Annotation agreementImages in agreementAgreement rate
Style331/35094.57%
All four subject indicators339/35096.86%
Style and all subject indicators320/35091.43%

For the style boundaries you specifically raise, 94.57% of style labels remain unchanged. Joint agreement on style and all subject indicators is 91.43%. Differences primarily involve the addition or removal of subject labels, such as animals, and boundaries between photography and design or between anime and painting. This paired audit shows that most style and subject annotations remain stable across the two source-label conditions.

B. The model distinguishes sources within the same scene; style priors alone cannot explain its performance.

Source proportions do differ between 3D and Design. We therefore construct a baseline that predicts source solely from the training-set AIGC/Non-AIGC proportions of each scene, without reading the image. We compute balanced accuracy within each of the 65 scenes containing both classes and average equally across scenes (scene-macro BAcc.), so the majority class in a scene has no scoring advantage.

MethodRedScenes Acc.Scene-macro BAcc.
Scene-source prior only72.6150.00
FakeCueS Direct92.8989.73
FakeCueS Cue92.8191.07

The prior-only baseline achieves 72.61% overall accuracy but 50.00% scene-macro BAcc., compared with 89.73%/91.07% for FakeCueS Direct/Cue. This gap shows that the model's discrimination extends beyond identifying the style and selecting its majority source class.

AuthorsOpenReview ↗

Response to Reviewer tdtt (Part 2 of 3)

W2: Component contributions beyond in-domain training

We use training comparisons with the same backbone to assess gains beyond in-domain training. The following results reproduce Table 3 of the revised manuscript. All configurations use Qwen3.5-4B, with all except Zero-shot trained on RedScenes. Values are accuracy (%), and Avg. is the mean across nine OOD datasets.

MethodPredictID: RedScenesOOD: Anom.-DFOOD: ChameleonOOD: Avg.
Zero-shotLabel66.7792.3775.4882.69
Pure RLCue89.5593.5081.8284.82
FakeCueSDirect92.8994.5884.7987.32
− Cue LearningDirect91.7392.2681.7784.05
FakeCueSCue92.8194.2381.7385.50
− Scene StatisticsCue90.5494.0779.8685.07

Compared with standard GRPO on Qwen3.5-4B (Pure RL), FakeCueS improves ID accuracy by 3.26 percentage points and mean OOD accuracy by 0.68 points when both use Cue voting. This demonstrates the additional benefit of scene-conditioned cue learning.

The two core ablations assess cue learning as a whole and scene-conditioned statistics. Removing cue learning (− Cue Learning) reduces Direct accuracy by 1.16/3.27 points on ID/OOD. Replacing scene-conditioned statistics with global statistics (− Scene Statistics) reduces Cue accuracy by 2.27/0.43 points. The updated results show that cue learning improves both in-domain detection and external generalization, with the full model outperforming the scene-aware classifier.

To address the advantage of in-domain training, we also trained a standard source classifier without scene or cue modules. Its ID/mean OOD accuracies are 91.56%/85.96%, compared with 92.89%/87.32% for full FakeCueS Direct. The advantage thus persists against a classifier trained on the same RedScenes data.

Scene supervision and scene-conditioned statistics have distinct roles. In the classifier-only comparison, retaining the scene architecture while removing the scene-supervision loss changes mean OOD accuracy from 84.05% to 85.46%. The component gains are therefore supported by cue learning and scene-conditioned statistics, rather than by adding scene-classification supervision alone.

Q1: Dataset and code release

We plan to release the RedScenes images and labels together with the training and evaluation code to support reuse and reproduction of the main results. We are organizing and de-identifying the dataset and plan a near-term release, subject to completion of reviews covering data-use permissions, copyright, and privacy.

Q2: Relationship between generated cues and classification decisions

Direct and Cue are independently usable prediction modes for different application needs. We train Qwen3.5-4B for one epoch on the 151,790 RedScenes training images. Direct provides efficient detection through the source-classification head. Cue generates forensic cues and takes a majority vote over their source verdicts, making the cues the explicit basis of the Cue decision. Cue learning benefits Direct during joint training by improving their shared representations.

The two modes have similar detection performance on RedScenes, and the choice depends on the application. Direct avoids text generation, while Cue provides observations and reasoning that can be inspected. When they disagree, the application should retain its preselected mode: Direct when detection efficiency is the priority, or Cue when inspectable forensic cues are required.

Q3: Frequency of reasoning that violates the prompt constraints

Thank you for pointing this out. We randomly sampled 200 complete test responses containing 800 cues and used an external model to examine observation, summary, and reasoning for source inferences based on creative intent, compositional purpose, or expressive motivation. The audit flagged 21 cues as potential violations, yielding an estimated prevalence of approximately 2.63% among generated test cues. None of the 200 responses consisted entirely of such cues. These findings indicate that prompt constraints do not fully eliminate this form of reasoning, but its estimated prevalence is low.

AuthorsOpenReview ↗

Response to Reviewer tdtt (Part 3 of 3)

Q4: Accuracy and usefulness of the generated cues

Independent pointwise and pairwise comparisons with FakeVLM, AIGI-Holmes, and other methods favor FakeCueS's complete forensic responses.

We compare 1,120 complete responses from eight models on 140 images covering 70 scenes, including incorrect answers from each model. For pointwise evaluation, the separately trained specialist evaluator DeepfakeJudge scores responses from 1 to 5 (Kuckreja et al., 2026). For pairwise evaluation, the external general-purpose multimodal model Gemini Argon compares responses using the Veritas evaluation prompt, and preferences are aggregated into Elo ratings (Tan et al., 2026b).

MethodDeepfakeJudge ↑Elo: source label provided ↑Elo: source label withheld ↑
FakeVLM1.964758.2776.2
FakeShield2.750902.4843.4
AIGI-Holmes2.729938.7917.0
FakeReasoning2.629981.81024.2
UniGenDet3.1071037.71044.3
Veritas3.4431020.21074.3
BusterX++3.4711076.41109.7
FakeCueS3.7571284.61210.9

FakeCueS achieves a DeepfakeJudge score of 3.757, exceeding the two methods you mention, FakeVLM (1.964) and AIGI-Holmes (2.729), and ranks first under both pairwise conditions. Since classification correctness can influence response scoring, we additionally withhold the ground-truth source label so the evaluator compares responses using the image and response content. Under this condition, FakeCueS retains an average preference rate of 77.98%, including 59.05% against the strongest comparator, BusterX++. Its forensic responses therefore retain a quality advantage without providing the evaluator with the correct source label.

AuthorsOpenReview ↗

Kindly Follow-Up: Reviewer tdtt

Dear Reviewer tdtt,

Thank you again for your thoughtful review and constructive suggestions. As the discussion period ends in about 5 hours, we would greatly appreciate it if you could take a look at our response when you have a moment. We would be happy to clarify any remaining questions, and if our response addresses your concerns, we would be grateful if you could take it into account in your final assessment.

Best regards, The Authors

Reviewer_K1a1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FakeCueS combines visual verification and historical cue-report statistics to learn scene-conditioned forensic cues, jointly with image-level classification. It also introduces RedScenes, containing 168,688 social-media images across photographic and non-photographic content.

Critical strengths
  1. The problem is practically important: photographic detection criteria may incorrectly penalize human-created artwork. RedScenes broadens evaluation across five styles and 70 scene combinations.
  2. Joint training shows promising external transfer: AnomReason-Deepfake accuracy improves from 92.26% for the scene-aware classifier to 94.49%. Sparse-statistics backoff and pre-image history-update rules are explicitly documented.
Critical weaknesses
  1. The core scene-conditioned mechanism is insufficiently isolated. On RedScenes, the scene-aware classifier achieves 91.73%, versus 91.53% for the full method. This small difference does not establish inferiority, but does not demonstrate added in-domain accuracy either. Existing configurations differ in training budgets, prompts, and prediction paths; a matched global-statistics versus scene-conditioned-statistics comparison is missing.
  2. Cue validity remains insufficiently established. Figure 4(c) illustrates selected differences in model verdicts rather than independently verified diagnostic relationships. The verifier assesses observations rather than the full forensic reasoning, while low-support reports still enter history. Thus, cue-report associations may reflect model reporting biases. Source-informed scene annotations and splits without author/post grouping introduce additional potential confounds.
Questions for authors
  1. Under matched training and inference conditions, how much improvement comes specifically from scene-conditioned rather than global cue statistics?
  2. Can independent evaluation establish both visual support and scene-dependent diagnostic correctness, including the effect of low-support reports entering history?
  3. Do the conclusions hold under author/post-separated evaluation and source-blind scene-annotation audits?
  4. Could the authors discuss differences from the already-cited LTE [1], ReAlign [2], LEGION [3], and FakeXplain [4], and compare with a suitable subset on compatible shared benchmarks to clarify the benefits of scene-conditioned cue learning?

References: [1] Ji et al. Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images. CVPR 2026. [2] Huang et al. ReAlign: Generalizable Image Forgery Detection via Reasoning-Aligned Representation. CVPR 2026. [3] Kang et al. LEGION: Learning to Ground and Explain for Synthetic Image Detection. ICCV 2025. [4] Ji et al. FakeXplain: AI-Generated Image Detection via Human-Aligned Grounded Reasoning. ICLR 2026.

4 comments
AuthorsOpenReview ↗

Response to Reviewer K1a1 (Part 1 of 3)

Thank you for recognizing the importance of detection across visual styles and for your suggestions on scene statistics, cue validity, and evaluation controls. We first apologize for an error in our initial implementation and the additional review burden it caused. Our checks identified that cue clusters and historical statistics were not correctly shared across GPUs during training. We have corrected the implementation, rerun training and evaluation, updated the full-model results, and expanded OOD evaluation from two datasets to nine. We address your concerns below.

W1, Q1: Benefits of scene-conditioned over global cue statistics

Scene-conditioned statistics improve ID accuracy by 2.27 percentage points and scene-macro BAcc. by 4.58 points. We compare global and scene-conditioned statistics while retaining scene injection and joint supervision. Both configurations use Qwen3.5-4B, 2,372 total updates, and Cue voting. Global statistics pool cue reports across scenes, whereas scene-conditioned statistics compare reporting rates for the two source classes under the same style and subject conditions.

In addition to overall accuracy, we compute balanced accuracy within each of the 65 scenes containing both source classes and average equally across scenes (scene-macro BAcc.). This tests whether the gains persist after balancing scene sizes and source proportions. All accuracy values below are percentages.

ScopeGlobal Acc.Scene-conditioned Acc.Global scene-macro BAcc.Scene-conditioned scene-macro BAcc.
Photo93.6294.4391.0394.20
Anime90.5091.6387.5287.74
3D88.7992.5274.9388.26
Design89.6992.0788.8093.07
Painting87.0491.7789.4792.34
Overall90.5492.8186.4991.07

Both accuracy and scene-macro BAcc. improve in all five styles, with a 13.33% increase in scene-macro BAcc. for 3D. The gains reflect more balanced source discrimination within scenes, supporting the value of learning cue–source associations in their visual context.

Regarding the absence of an additional ID gain in the original results, the corrected full model achieves 92.89% Direct accuracy, compared with 91.73% without cue learning. Their mean accuracies across nine OOD datasets are 87.32% and 84.05%, respectively (Table 3). Cue learning therefore improves ID/OOD accuracy by 1.16/3.27%, with a larger gain on external data.

AuthorsOpenReview ↗

Response to Reviewer K1a1 (Part 2 of 3)

W2, Q2: Visual support, diagnostic correctness, and low-support history

A. Evaluating both visual support and complete forensic reasoning.

To examine the forensic relevance of the scene differences illustrated in Figure 4(c), we evaluate both the visual basis of observations and the reasoning that connects them to source judgments. During training, the verifier receives only the image and observation, and its support score contributes to the cue reward. To test dependence on the specific image, we hold observation text fixed and substitute an image with the same style and subject labels but the opposite source. Across 16,782 image pairs, mean support, averaged equally across the five styles, drops from 0.967 to 0.258. Different images within the same scene thus provide substantially different support for the same observations.

For complete forensic responses, we use pointwise scoring and pairwise comparison. The separately trained specialist evaluator DeepfakeJudge scores individual responses (Kuckreja et al., 2026); the external general-purpose multimodal model Gemini Argon compares response pairs using the Veritas evaluation prompt, with preferences aggregated into Elo ratings (Tan et al., 2026b). Evaluation covers 70 scenes, 140 images, and 1,120 responses from eight models.

MethodDeepfakeJudge ↑Elo: source label provided ↑Elo: source label withheld ↑
FakeVLM1.964758.2776.2
FakeShield2.750902.4843.4
AIGI-Holmes2.729938.7917.0
FakeReasoning2.629981.81024.2
UniGenDet3.1071037.71044.3
Veritas3.4431020.21074.3
BusterX++3.4711076.41109.7
FakeCueS3.7571284.61210.9

FakeCueS ranks first in all three evaluations. Providing the ground-truth source label lets the evaluator assess explanations against the correct source. Withholding it reduces the influence of our detection-accuracy advantage on response-quality evaluation. In the latter condition, FakeCueS retains a 59.05% preference rate over the strongest comparator, BusterX++. Together with the scene-macro BAcc. gains in Q1, these results support the diagnostic value of scene-conditioned cues through both independent response evaluation and detection performance.

B. Offline deletion quantifies the effect of low-support history, while current cues are rewarded according to instance-specific support.

Historical statistics estimate cue-reporting frequencies normalized within each source class, while the visual support of the current observation scales its reward contribution. Each image contributes at most one count per cluster, and the current image is scored against history before its own updates, limiting repeated counting and immediate self-reinforcement. A semantic cluster can occur in different images, so a low-support observation does not establish that the entire cluster is consistently invalid.

We performed an offline deletion analysis on saved training trajectories. We identified 316 clusters whose creation-time text–image support was below 0.5, removed their 2,425 source-history records, and recomputed subsequent rewards while holding generated responses, cluster assignments, and novelty rewards fixed.

Across all 593 periodically saved highest-reward responses, the mean absolute reward change is 0.0240, and 81.79% of rewards are unchanged. The average effect is therefore limited within this set of saved responses. Moreover, 122 of 133 subsequently saved cues from these clusters have visual support of at least 0.5. Visual support is instance-dependent: one low-support report does not invalidate a semantic cluster, and weighting the current instance by its support preserves evidence that is supported in other images.

AuthorsOpenReview ↗

Response to Reviewer K1a1 (Part 3 of 3)

W2, Q3: Author/post separation and source-blind annotation audits

A. Removing author overlap leaves detection performance essentially unchanged.

RedScenes contains 168,688 images from 130,344 posting accounts, with only one image retained per post. To remove overlap at the account level, we hold the model and decision thresholds fixed, exclude 5,232 test images (30.96%) whose posting accounts also occur in training, and evaluate the remaining 11,666 images.

OutputAcc.: original → filteredBAcc.: original → filtered
Direct92.89 → 92.8092.89 → 92.76
Cue92.81 → 92.8392.81 → 92.58

The accuracy change is below 0.10 points for both outputs, and the BAcc. change is at most 0.23 points. Performance remains stable after removing nearly one-third of the test images, supporting the conclusion that the results do not primarily depend on author overlap between training and testing.

B. Most scene annotations remain unchanged without source labels.

We sample five images from each of 70 scenes, totaling 350 images, and annotate them with and without source labels while holding all other inputs and decoding settings fixed.

Annotation agreementImages in agreementAgreement rate
Style331/35094.57%
All four subject indicators339/35096.86%
Style and all subject indicators320/35091.43%

Complete scene annotations agree for 91.43% of images, indicating that most annotations remain stable across the two source-label conditions. Differences primarily involve the addition or removal of subject labels, such as animals, and boundaries between photography and design or between anime and painting.

Q4: Relationship to LTE, ReAlign, LEGION, and FakeXplain

A. FakeCueS learns how the source association of a cue varies across visual contexts.

The same visual phenomenon can occur naturally at different frequencies in photographs, paintings, and 3D imagery. FakeCueS uses style and subject content as context, accumulates source associations from cue reports within each scene, and uses these statistics to construct online rewards. Training requires image-level source and scene labels, without per-image reference explanations or region annotations. The distinctions from the cited methods are as follows.

MethodCore mechanism and supervisionDistinction in FakeCueS
LTELocalizes suspicious regions, then examines local crops and the full image; uses region and explanation supervisionLearns cue–source associations from statistics across images within each scene
ReAlignTransfers optimized and verified reasoning text to a lightweight detector through contrastive learningJointly learns open-ended cues and detection representations, with online rewards from visual support and scene-conditioned statistics
LEGIONUses pixel-level artifact masks, categories, and textual explanations to jointly learn localization and explanationUses image-level labels and a frozen verifier, without pixel-level artifact annotations
FakeXplainUses bounding boxes and reference explanations for supervised fine-tuning and staged reinforcement learningLearns source associations from historical cue-reporting rates, without bounding boxes or per-image reference explanations

B. Benchmarking recent VLM detectors with publicly released weights.

At the time of this revision, we could not obtain public weights that directly reproduce the reported detection results of these four methods. The official LEGION repository provides intermediate localization/explanation weights and states that the final detection weights are unavailable. The repository linked by the FakeXplain paper remains empty. We therefore assembled a benchmark of recent VLM detectors with released weights: FakeShield, FakeReasoning, FakeVLM, Veritas, UniGenDet, AIGI-Holmes, and BusterX++. The benchmark covers RedScenes and nine external datasets (Tables 1–2), and Q2 compares their complete response quality.

Across the nine external datasets, FakeCueS Direct/Cue achieve mean accuracies of 87.32%/85.50%, exceeding the strongest comparator, BusterX++ (82.01%), by 5.31/3.49 points. Together with the scene-statistics comparison in Q1, these results demonstrate both the full method's external generalization and the specific benefit of scene-conditioned cue statistics.

AuthorsOpenReview ↗

Kindly Follow-Up: Reviewer K1a1

Dear Reviewer K1a1,

Thank you again for your thoughtful review and constructive suggestions. As the discussion period ends in about 5 hours, we would greatly appreciate it if you could take a look at our response when you have a moment. We would be happy to clarify any remaining questions, and if our response addresses your concerns, we would be grateful if you could take it into account in your final assessment.

Best regards, The Authors

Reviewer_ED5KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FakeCueS proposes scene-conditioned forensic cue learning for AI-generated image detection. It combines style–subject conditioning, visual verification, online cue clustering, and group-relative policy optimization to learn image-specific cues without reference rationales. The authors also introduce RedScenes, a dataset of 168,688 social-media images covering 70 scene combinations. Experiments report 91.53% accuracy on RedScenes and evaluate cross-dataset transfer to AnomReason-Deepfake and Chameleon.

Critical strengths
  • Relevant problem formulation. The paper identifies an important challenge in image forensics: the diagnostic value of visual cues can depend on scene context. Modeling visual style and subject content jointly is a reasonable approach to reducing misinterpretation of legitimate artistic characteristics as generation artifacts.

  • Diverse dataset and empirical evaluation. RedScenes covers photographic and non-photographic content with source, style, and subject annotations. The experiments include multiple detection baselines, cross-dataset evaluation, and cue-level analyses, providing useful evidence about detection across heterogeneous visual content.

Critical weaknesses
  • Insufficient validation of the central cue-learning contribution. The scene-conditioned classifier without cue learning slightly outperforms full FakeCueS on RedScenes (91.73% vs. 91.53%, Table 3). Although joint training improves AnomReason performance, the contributions of visual verification, scene-conditioned association, semantic clustering, and novelty rewards are not individually isolated. More fundamentally, the default detector predicts directly from the input prefix without generating cues at inference. Appendix D.3 reports that 99.96% of valid responses assign identical verdicts to all four cues. Without independent assessment of cue faithfulness and diagnostic specificity, it remains unclear whether the framework learns meaningful forensic evidence or primarily reinforces classification preferences through auxiliary supervision.

  • Dataset confounds and insufficiently controlled generalization evidence. RedScenes uses image-level rather than author-disjoint splits, while Gemini-generated scene annotations are conditioned on ground-truth source labels without subsequent manual verification (Section 3; Appendix A). Substantial class imbalance across styles raises additional concerns about source-prior shortcuts. In-domain comparisons largely involve differently trained or zero-shot baselines, while external comparisons adopt heterogeneous training protocols. FakeCueS also trails TranX-Adapter on Chameleon (83.09% vs. 85.10%). Without source-controlled splits, annotation audits, and matched training baselines, the reported improvements cannot be confidently attributed to scene-conditioned cue learning, limiting the strength of the generalization claims.

Questions for authors
  • Can the authors provide controlled ablations of scene conditioning, visual verification, source association, clustering, and novelty rewards under identical training budgets? In particular, why does the scene-only classifier outperform the full method on RedScenes, and how do these components affect both external benchmarks?

  • Given the 99.96% agreement among cue verdicts, what evidence establishes that the generated cues represent distinct, visually grounded, and genuinely source-discriminative evidence? Can independent human evaluation or controlled cue perturbation verify their faithfulness and scene dependence?

  • Can the authors evaluate author-disjoint splits, source-blind scene annotation, and scene-prior-only baselines? How sensitive are the results to scene-label errors and source–style correlations?

  • How does a standard source-supervised Qwen3.5-4B classifier without scene conditioning or cue learning perform under matched training settings? Can the scene-only baseline also be evaluated on Chameleon to isolate the contribution of cue learning?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ED5K (Part 1 of 4)

Thank you for your suggestions on cue learning, evaluation controls, and data use. We first apologize for an error in our initial implementation and the additional review burden it caused. Our checks identified that cue clusters and historical statistics were not correctly shared across GPUs during training. We have corrected the implementation, rerun training and evaluation, updated the full-model results, and expanded out-of-domain (OOD) evaluation from two datasets to nine. We address your concerns below.

W1, W2, Q1, Q4: Component contributions, training baselines, and OOD generalization

A. Improvements over standard GRPO and contributions of the two core components.

The following ablations reproduce Table 3 of the revised manuscript. All configurations use Qwen3.5-4B. Values are accuracy (%). Avg. is the mean across nine OOD datasets.

MethodPredictID: RedScenesOOD: Anom.-DFOOD: ChameleonOOD: Avg.
Zero-shotLabel66.7792.3775.4882.69
Pure RLCue89.5593.5081.8284.82
FakeCueSDirect92.8994.5884.7987.32
− Cue LearningDirect91.7392.2681.7784.05
FakeCueSCue92.8194.2381.7385.50
− Scene StatisticsCue90.5494.0779.8685.07

Compared with standard GRPO on Qwen3.5-4B (Pure RL), FakeCueS improves in-domain (ID) accuracy by 3.26 percentage points and mean OOD accuracy by 0.68 points when both use Cue voting. This demonstrates the additional benefit of scene-conditioned cue learning.

The two core ablations assess cue learning as a whole and scene-conditioned statistics, respectively. Removing cue learning reduces Direct accuracy by 1.16/3.27 points on ID/OOD. Replacing scene-conditioned statistics with global statistics reduces Cue accuracy by 2.27/0.43 points.

For the standard source-supervised classifier you requested, we also evaluated a model with the same backbone and training data but without scene or cue modules. It achieves 91.56% on RedScenes, 84.07% on Chameleon, and 85.96% on average across the nine OOD datasets, compared with 92.89%, 84.79%, and 87.32% for full FakeCueS Direct. Thus, the gains extend beyond training the same backbone on in-domain classification data.

B. Scene conditioning accounts for the different meanings of visual cues across styles.

Practical AIGC detection must distinguish image sources across photographs, paintings, and 3D imagery. The same visual phenomenon can occur naturally at different frequencies across these styles, changing its diagnostic value. Within each scene, FakeCueS estimates the fraction of images in each source class for which a cue is reported, learning its contextual source association. Here, Real/Fake correspond to Non-AIGC/AIGC.

We compare global statistics pooled across scenes with scene-conditioned statistics, retaining scene injection, joint supervision, and Cue voting. In addition to overall accuracy, we compute balanced accuracy within each of the 65 scenes containing both source classes and average equally across scenes (scene-macro BAcc.). This measures discrimination within scenes while balancing their source proportions.

Statistics used for Cue predictionRedScenes Acc.Scene-macro BAcc.Anom.-DFChameleon
Global statistics90.5486.4994.0779.86
Scene-conditioned statistics92.8191.0794.2381.73

Scene-conditioned statistics improve overall accuracy by 2.27 points and scene-macro BAcc. by 4.58 points. Within 3D, scene-macro BAcc. increases from 74.93% to 88.26% (+13.33 points), and both external benchmarks also improve. The gains persist when scenes and source classes receive equal weight, supporting more balanced Real/Fake discrimination in scenes with uneven source distributions.

AuthorsOpenReview ↗

Response to Reviewer ED5K (Part 2 of 4)

W1, W2, Q1, Q4: Component contributions, training baselines, and OOD generalization

C. Generalization across nine datasets, with model and training-data scale comparisons.

The same model trained on RedScenes is applied directly to nine external datasets without target-domain fine-tuning. Direct and Cue achieve mean accuracies of 87.32% and 85.50%, respectively, compared with 82.01% for the strongest listed baseline, BusterX++. Direct ranks first on six datasets (Table 2).

For Chameleon, the original manuscript used the best configuration reported for TranX-Adapter: LLaVA-1.6-Mistral-7B trained on the full GenImage training set, achieving 85.10%. We now include its 4B configurations below. Results are from Table 2 of the original paper, and image counts follow the official GenImage training split.

MethodBackboneTraining data and image countChameleon Acc.
TranX-AdapterLLaVA-1.6-Mistral-7BFull GenImage, approximately 2.58M85.10
TranX-AdapterQwen3-VL-4BFull GenImage, approximately 2.58M83.60
TranX-AdapterQwen3-VL-4BGenImage SD v1.4 subset, approximately 0.32M72.60
FakeCueS (Direct)Qwen3.5-4BRedScenes, 151,79084.79

FakeCueS (4B) exceeds the best reported 4B TranX-Adapter result by 1.19 points, using approximately 1/17 as many training images.

W1, Q2: Visual grounding, diagnostic value, and decision explanations

A. Independent evaluations show higher-quality complete forensic responses.

Following related work, we use pointwise scoring and pairwise comparison to assess 1,120 complete responses from eight models on 140 images covering 70 scenes. For pointwise evaluation, DeepfakeJudge, a separately trained specialist evaluator, scores each response from 1 to 5 (Kuckreja et al., 2026). For pairwise evaluation, the external general-purpose multimodal model Gemini Argon compares responses using the Veritas evaluation prompt, and the preferences are aggregated into Elo ratings (Tan et al., 2026b). Pairwise evaluation is conducted both with and without the ground-truth source label (Table 5).

MethodDeepfakeJudge ↑Elo: source label provided ↑Elo: source label withheld ↑
FakeVLM1.964758.2776.2
FakeShield2.750902.4843.4
AIGI-Holmes2.729938.7917.0
FakeReasoning2.629981.81024.2
UniGenDet3.1071037.71044.3
Veritas3.4431020.21074.3
BusterX++3.4711076.41109.7
FakeCueS3.7571284.61210.9

FakeCueS ranks first in pointwise evaluation and both pairwise conditions. Because FakeCueS also has higher detection accuracy, we withhold ground-truth source labels from the evaluator to reduce the influence of answer correctness and focus evaluation on visual grounding and reasoning quality. Under this condition, FakeCueS retains an average preference rate of 77.98%, including 59.05% against the strongest comparator, BusterX++. The advantage persists without supplying ground-truth source labels.

AuthorsOpenReview ↗

Response to Reviewer ED5K (Part 3 of 4)

W1, Q2: Visual grounding, diagnostic value, and decision explanations

B. Differences in cue content, visual support, and diagnostic value.

Content differences. The cited 99.96% measures agreement among source verdicts. We analyzed the final model's 16,832 valid test responses using the text-embedding clustering rule employed during training. The approximately four cues per response form 3.74 semantic clusters on average, and 97.99% of responses contain at least three clusters. Agreement on source therefore does not imply identical cue text. The painting example in Figure 5 examines brushwork, paper texture, body proportions, and the appearance of a casual sketch, illustrating different observations supporting the same verdict.

Visual support. During training, image–text consistency checks incorporate the image's support for each observation into the reward. The verifier receives neither the source label nor the cue verdict. To test whether observations depend on the specific image, we hold their text fixed and replace the image with one having the same style and subject labels but the opposite source. Across 16,782 image pairs, mean support, averaged equally across the five styles, falls from 0.967 to 0.258 under the same verifier. Thus, visual support changes substantially with the image even when its scene category is held fixed.

Diagnostic value. FakeCueS estimates source associations from cue-reporting rates within each source class and scene. Replacing global statistics with scene-conditioned statistics improves scene-macro BAcc. from 86.49% to 91.07%, showing that interpreting cue–source associations in context improves detection.

W2, Q3: Data splits, scene annotations, and source priors

Thank you for identifying these potential confounds. We examine author overlap through filtered evaluation, annotation sensitivity through paired source-label conditions, and the explanatory power of scene priors through a prior-only baseline.

A. Removing author overlap leaves detection performance essentially unchanged.

RedScenes contains 168,688 images from 130,344 posting accounts. Holding the model and decision thresholds fixed, we remove 5,232 images (30.96%) from the 16,898-image test set because their posting accounts also occur in training, and reevaluate on the remaining 11,666 images.

Full FakeCueS outputAcc.: original → filteredBAcc.: original → filtered
Direct92.89 → 92.8092.89 → 92.76
Cue92.81 → 92.8392.81 → 92.58

After removing nearly one-third of the test images, the accuracy change is below 0.10 points for both Direct and Cue, and the BAcc. change is at most 0.23 points. Removing author overlap does not materially alter detection performance, supporting the conclusion that the results do not primarily depend on features of recurring authors.

B. Scene annotations remain largely stable when source labels are withheld.

We sample five images from each of 70 scenes, totaling 350 images, and annotate each image with and without its source label while keeping all other inputs and decoding settings fixed.

Annotation agreementImages in agreementAgreement rate
Style331/35094.57%
All four subject indicators339/35096.86%
Style and all subject indicators320/35091.43%

Complete scene annotations agree for 91.43% of images. Most annotations thus remain stable across the two source-label conditions. Differences primarily involve the addition or removal of subject labels, such as animals, and boundaries between photography and design or between anime and painting.

Scene labels enter training through auxiliary supervision and the grouping of cue statistics. At inference, the scene representation is formed from image features and predicted style and subjects; no external scene annotation is required. During training, scene–cue combinations with limited statistical support rely more heavily on global statistics, reducing reward fluctuations from sparse groups.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.