acceptodds
Under review as a conference paper at ICLR 2027

RED: Reconstruction Evolution Dynamics for Generalizable AI-Generated Image Detection

Abstract

The rapid evolution of image generators calls for forensic cues that generalize beyond known generation mechanisms. Existing detectors often rely on static image representations or endpoint reconstruction discrepancies, leaving the evolution of intermediate reconstruction stages underexplored. We observe that the relative token predictability of real and generated images can reverse across reconstruction scales, suggesting that intermediate stages may expose forensic evidence overlooked by endpoint comparisons. Motivated by this observation, we propose RED (Reconstruction Evolution Dynamics), a framework that captures transferable forensic cues from coarse-to-fine reconstruction evolution. To our knowledge, RED is the first framework to use scale-wise token predictability to guide forensic evidence aggregation across intermediate reconstruction states. It represents the reconstruction trajectory produced by a frozen multiscale VQ-VAE in the shared feature space of a frozen CLIP encoder. To connect the observed predictability variations with visual evidence, RED learns image-adaptive stage weights from scale-wise token negative log-likelihoods provided by a frozen VAR model. A cross-stage evidence aggregation module then jointly models the original-image representation and the weighted reconstruction features, capturing complementary forensic cues through interactions along the reconstruction trajectory. Experiments on six diverse benchmarks demonstrate that RED achieves the highest average accuracy of 92.5% and average precision of 97.5% among the evaluated methods. Further evaluations show strong robustness to common image degradations, supporting the value of reconstruction evolution for generalizable AI-generated image detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, Reconstruction evolution, Domain generalization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TKW1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RED for generalizable AI-generated image detection. RED extracts CLIP representations of the original image and intermediate reconstructions from a frozen multiscale VQ-VAE. Scale-wise token negative log-likelihoods from a frozen VAR model guide image-dependent stage weighting, followed by cross-stage Transformer aggregation. The method achieves strong average performance across six evaluation collections, including five outside the detector’s training distribution.

Critical strengths
  • The design is well motivated: reconstruction features and NLL scores share the same token hierarchy. Ablations support both intermediate states (89.9% → 91.0%) and NLL guidance (91.0% → 92.5%).
  • Strong results across six collections: 92.5% average accuracy and 97.5% AP, with additional robustness evaluations.
  • Source-only model selection and alternative real-image source controls strengthen the evaluation.
Critical weaknesses
  1. Table 9 shows reasonably strong overall accuracy, decreasing from 90.7% to 88.5% when ImageNet real images are replaced with Chameleon real images. However, real-image accuracy and false-positive rates are not reported separately, leaving the absolute false-alarm rate on non-ImageNet real images unclear.
  2. RED applies the same NLL-derived weights before cross-stage attention and again during output pooling. The improvement from 91.0% with uniform weights to 92.5% supports NLL guidance overall, but does not isolate the benefit of applying these weights twice or sharing them between the two locations.
  3. Equation (4) measures the conditional NLL of newly introduced tokens, whereas Equation (2) constructs reconstruction states containing information accumulated across preceding scales. Although the shared hierarchy establishes stage alignment, the rationale for using an incremental score to weight the entire cumulative visual representation remains insufficiently explained.
Questions for authors
  1. Could the authors report real-image accuracy (TNR) and false-positive rates at the fixed threshold of 0.5 for ImageNet, Chameleon, LSUN, and MS COCO, together with the generated-image detection rate?
  2. Could the authors compare input-only weighting, output-only weighting, and weighting at both locations with shared versus separate NLL-based weights, while keeping the remaining architecture fixed?
  3. Why is per-scale conditional NLL preferred over cumulative NLL for weighting cumulative reconstruction features? Could the authors compare it with token-averaged cumulative NLL using the same source-domain standardization and detection architecture?
Reviewer_GMDuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RED detects AI-generated images by analyzing coarse-to-fine reconstruction states produced by a frozen multiscale VQ-VAE. A frozen CLIP encoder extracts visual features, while scale-wise token likelihoods from a frozen VAR model guide image-adaptive stage weighting. A lightweight Transformer then aggregates the original-image and reconstruction features. The method achieves 92.5% average accuracy and 97.5% average AP across six evaluation collections.

Critical strengths

1.The scale-dependent reversal in the relative token predictability of real and generated images is an interesting observation that motivates examining intermediate reconstruction states.

2.The ablations support the value of intermediate reconstruction evidence: accuracy increases from 89.9% with original-image and final-reconstruction features to 91.0% with uniformly weighted intermediate states, and to 92.5% with full RED.

3.The experiments cover diverse generator families and common image degradations. The additional real-image source control in Table 9 provides useful evidence of generalization beyond the main evaluation protocol.

Critical weaknesses

1.Comparisons with uniform and static weights support the effectiveness of the proposed adaptive weighting, but a visual-feature-based adaptive weighting baseline under the same architecture is missing.

2.All six evaluation collections share ImageNet real images as their reference, so detection results may partly reflect source-related differences. Table 9 partially addresses this concern, but uses only one fixed generated-image source and compares three baselines.

3.RED requires ten reconstruction outputs, CLIP feature extraction for eleven images, and VAR likelihood computation, yet inference latency and memory usage are not reported. A small number of trainable parameters does not imply efficient inference.

Questions for authors

1.Under the same aggregation architecture, how does NLL-guided weighting compare with visual-feature-based adaptive weighting? Does randomly shuffling NLL vectors across images substantially reduce performance?

2.Could comparative experiments use the datasets’ original real-image splits where available, or incorporate multiple real-image sources across more generated-image collections?

3.What are the end-to-end inference latency and peak GPU memory usage? How much detection performance is retained when fewer reconstruction stages are used?

The paper provides promising evidence that intermediate reconstruction states improve detection performance. However, the current controlled experiments do not sufficiently establish the distinct contribution of NLL guidance or the claimed breadth of generalization. Stronger adaptive-weighting baselines and broader real-image source evaluation could change my assessment.

Reviewer_S5DWOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

RED investigates reconstruction evolution as a source of transferable forensic evidence for AI-generated image detection. It represents ten coarse-to-fine VQ-VAE reconstruction stages in a shared CLIP feature space, employs VAR token negative log-likelihoods for image-adaptive stage weighting, and aggregates cross-stage evidence through a Transformer. Evaluations on GenImage and five external benchmarks report 92.5% average accuracy and 97.5% average precision, with additional ablations, degradation tests, and real-image source controls.

Critical strengths
  • Well-motivated and technically meaningful formulation. The scale-dependent reversal in real/generated token predictability (Figure 1; Appendix C) provides a concrete motivation for moving beyond endpoint reconstruction discrepancies. Combining intermediate reconstruction representations with NLL-guided stage weighting is a coherent approach to capturing complementary forensic evidence.

  • Strong empirical generalization with informative ablations. RED achieves the highest average accuracy and AP among the evaluated methods across six benchmarks, including diverse unseen generators. Table 4 shows that intermediate reconstruction stages improve average accuracy from 89.9% to 91.0%, while the complete method reaches 92.5%. The degradation experiments and real-image source controls further strengthen the empirical findings.

Critical weaknesses
  • The mechanism underlying NLL-guided adaptation remains incompletely characterized. Although Table 4 demonstrates improvements over uniform and learned static weighting, the aggregate NLL reversal does not directly establish that image-specific predictability identifies the most discriminative stages. The paper lacks analyses of learned stage weights, their variation across generators, and sensitivity to the reconstruction prior. Shuffled-NLL controls or alternative reconstruction backbones would more directly test whether the gains arise from meaningful sample-specific evidence allocation.

  • Computational cost and statistical reliability require further characterization. RED processes ten reconstruction states through a frozen CLIP encoder, alongside VQ-VAE reconstruction and VAR likelihood evaluation, yet reports no end-to-end latency, throughput, or memory comparison. The evaluation also reuses a fixed ImageNet real-image reference pool across generator groups and reports aggregate scores without uncertainty estimates. Although Appendix Table 9 provides useful real-source controls, additional paired confidence intervals and efficiency measurements are needed to assess the robustness and practical value of the reported improvements.

Questions for authors
  • How do learned stage weights vary across real images, generator families, and correct versus incorrect predictions? Does randomly permuting image-level NLL profiles while preserving their marginal distributions reduce performance?

  • What are RED's end-to-end inference latency, throughput, and GPU memory requirements relative to CLIP-based baselines? How much computation is attributable to reconstruction, VAR scoring, and repeated CLIP encoding?

  • Can the authors report paired confidence intervals or multi-seed variability for the principal improvements? How sensitive are the results to the choice of VAR/VQ-VAE backbone and alternative real-image reference distributions?

Comments on the paper

2 comments
Area_Chair_BpMJOpenReview ↗

Author's Responses

Dear Reviewers,

The authors have provided detailed responses to your concerns.

Please review their responses ASAP and respond to them, to try and resolve concerns before the end of the author-review discussion phase (Nov 18, AoE).

Best, AC

AuthorsOpenReview ↗

Response Summary

Dear PC, SAC, AC, and Reviewers,

We sincerely thank all chairs, committee members, and reviewers for their time and dedication. We particularly appreciate the reviewers' constructive feedback and the AC's careful coordination and proactive follow-up throughout the discussion.

Our additional work addresses the main concerns as follows:

  1. NLL guidance: Stage-weight analyses, a matched visual-adaptive baseline, and NLL permutation controls provide further evidence for the contribution of image-aligned NLL guidance.

  2. Design choices: Controlled ablations favor shared weighting at both locations and incremental NLL over cumulative NLL in the evaluated settings.

  3. Real-source generalization: We expanded evaluation to 24 real/generated combinations and reported TNR, FPR, and TPR at the fixed threshold of 0.5, clarifying source sensitivity and detection behavior.

  4. Statistical reliability: Three-seed comparisons and a VAR-d16-to-d20 prior replacement characterize training variability and likelihood-prior sensitivity, with the VQ-VAE backbone held fixed.

  5. Computational cost: End-to-end efficiency measurements, component profiling, and stage-reduction experiments identify decoding as the main bottleneck and quantify the trade-off between detection performance and inference speed.

We appreciate the reviewers' follow-up assessments acknowledging that these additions addressed their main concerns. We will incorporate the results, clarifications, and relevant protocols into the revised manuscript.

Finally, we once again thank all chairs, committee members, and reviewers for their careful work and valuable contributions.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.