{"id":"d9be5256-e50e-4060-9feb-6b89f82e9786","arxiv_id":"2607.08674","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":8,"one_line_summary":"Modeling the differential response of real versus AI-generated images under secondary generative transformations yields state-of-the-art cross-generator detection on 19 unseen models.","lead":"This paper detects AI-generated images by applying a second generative transform (like denoising or super-resolution) and measuring how the image responds differently from a real photo. The method could help platforms flag synthetic media from generators never seen during training.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The headline comparison is confounded by backbone choice: GenRes++ uses PE-Core-G14-448 while all baselines use CLIP, and the paper never re-runs baselines with PE-Core to isolate the generative residual contribution from the backbone upgrade.","rationale":"The reader's concern — that the differential-response signal may not generalize to future generator architectures — is valid but generic to all AIGI detection work and not specific to this paper's methodology. The more load-bearing and paper-specific concern is the backbone confound: the headline comparison pits a PE-Core-based method against CLIP-based baselines, and the paper's own ablations (Table 5) show that the backbone alone accounts for a 4.4% gain. The 'no transform' ablation (Table 4, 91.4% ACC) demonstrates that transforms add real value within the GenRes++ pipeline, but the cross-method comparison is fundamentally unfair without re-running baselines on the same backbone. This concern does not push the verdict below CONDITIONAL because: (1) the within-method ablations (NTN vs. simpler fusion, leave-one-out transforms, N-ablation) are genuinely convincing and show the framework's components each contribute; (2) the 4.3% gain from transforms over PE-Core-alone is non-trivial; (3) the compound-perturbation drop to 84.9% (Table 7) is already a known practical limitation. However, the concern does sharpen the CONDITIONAL verdict: the paper should either re-run top baselines with PE-Core or explicitly frame its contribution as 'a method that effectively leverages PE-Core's embeddings for residual-based detection' rather than claiming unqualified superiority over existing methods. The reader correctly identified the compound-perturbation issue and the lack of error bars/code as tempering conditions; the backbone confound is an additional, more specific concern that strengthens the case for CONDITIONAL rather than ACCEPT.","tokens_in":18577,"tokens_out":3058,"duration_ms":203850,"concrete_test":"Re-run the top two baselines (FreLens and C2P-CLIP) with PE-Core-G14-448 as the backbone under identical training conditions (same data, same LoRA rank, same hyperparameters). If either baseline reaches ≥95.0% mACC with PE-Core, the generative residual framework's marginal contribution over existing methods is called into question, and the headline claim of 'outperforming existing methods' would need to be qualified as 'outperforming existing methods that use weaker backbones.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GenRes++ 'outperforms existing methods' due to generative residual learning. However, all baselines (UniFD, C2P-CLIP, FreLens, RINE, etc.) use CLIP-based backbones, while GenRes++ uses PE-Core-G14-448, a larger and more recently pretrained encoder. Table 5 shows that swapping CLIP ViT-L/14 for PE-Core under identical GenRes++ training yields a 4.4% ACC gain (91.3% → 95.7%). Meanwhile, Table 4 shows that PE-Core with zero transforms already reaches 91.4% ACC / 96.1% AP — competitive with several baselines despite using no relational residual modeling at all. The critical missing experiment is: what happens if FreLens or C2P-CLIP is re-run with PE-Core as its backbone? If a strong baseline method like FreLens (95.0% mACC with CLIP) gains even 1-2% from PE-Core, it would match or exceed GenRes++ (95.7%), collapsing the headline claim. The transforms do add value over PE-Core alone (91.4% → 95.7%, a 4.3% gain), but this is a within-method ablation, not a cross-method comparison. The paper never establishes that 'PE-Core + generative residuals' beats 'PE-Core + best existing method.' Without this control, the improvement over baselines could be substantially or entirely attributable to the backbone rather than the proposed framework.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This manuscript proposes GenRes and GenRes++, frameworks for detecting AI-generated images (AIGI) by modeling the differential response of real versus synthetic images to secondary generative transformations (e.g., denoising, super-resolution). The core idea is that applying a second generative model to an AIGI produces characteristic residual discrepancies due to the interaction between the source generator's biases and the transformation network's priors. GenRes models these residuals using a Neural Tensor Network (NTN) on a single transform, while GenRes++ extends this by using a Cross-Attention Aggregation (CCA) module to pool features from N=5 complementary transforms. Both models use a frozen PE-Core vision encoder with LoRA fine-tuning. Evaluated on the UniversalFakeDetect benchmark (trained on ProGAN, tested on 19 unseen generators), GenRes++ achieves 95.7% mACC and 99.1% mAP, outperforming existing baselines. The experimental protocol is standard, and the ablation studies generally support the architectural design choices.","tokens_in":18786,"tokens_out":1232,"duration_ms":249085,"significance":"The problem of cross-generator AIGI detection is practically consequential and actively studied. The paper's central premise—that the differential response of images to secondary generative processing provides a generator-agnostic signal—is intuitive and well-motivated. The empirical results are strong, and the ablations (Tables 4-7, Figs. 4-6) provide falsifiable evidence for the contribution of individual transforms, LoRA rank, and fusion strategies. However, the significance of the proposed generative residual framework over a simple backbone upgrade is currently obscured by a major confounding variable in the experimental design, which must be addressed before the central claim can be fully accepted.","major_comments":[{"comment":"The headline comparison in Tables 1 and 2 is confounded by the choice of backbone. GenRes++ uses PE-Core-G14-448, while all baselines (UniFD, C2P-CLIP, FreLens, RINE, etc.) use CLIP-based backbones. Table 5 shows that swapping CLIP ViT-L/14 for PE-Core under identical GenRes++ training yields a 4.4% ACC gain (91.3% to 95.7%). Furthermore, Table 4 shows that PE-Core with zero transforms already reaches 91.4% ACC / 96.1% AP, which is competitive with several baselines despite using no relational residual modeling. The critical missing experiment is: what happens if a strong baseline (e.g., FreLens or C2P-CLIP) is re-run with PE-Core as its backbone? If a baseline gains even 1-2% from PE-Core, it would match or exceed GenRes++, collapsing the claim that the proposed framework outperforms existing methods. The transforms do add value over PE-Core alone (91.4% to 95.7%), but this is a within-","section":null},{"comment":"method ablation, not a cross-method comparison. The paper never establishes that 'PE-Core + generative residuals' beats 'PE-Core + best existing method.' Without this control, the improvement over baselines could be substantially or entirely attributable to the backbone rather than the proposed framework. The authors must isolate the generative residual contribution from the backbone upgrade to validate the central claim.","section":null}],"minor_comments":[{"comment":"§3.2.1: The text states 'd enc = 1536' for the PE-Core-G14-448 encoder, but later in §3.2.2 states 'Z = [z1, ..., zN] in R^{N x d} (d = 256)'. It would help to clarify explicitly that the bottleneck projection layer reduces the dimensionality from 1536 to 256 before the CCA module.","section":null},{"comment":"Table 3: The inference time for GenRes++ is reported as 4625.6 ms per image. This is quite high for a detection task. While the authors acknowledge this limitation in §5, it would be beneficial to briefly discuss in §4.4 whether this latency is dominated by the sequential application of the five transforms or the multi-branch encoding, and whether the transforms could be parallelized in practice.","section":null},{"comment":"§4.5.5 (Leave-One-Out): The text mentions 'removing FFDNet (ACC: 92.3%, AP: 96.4%), indicating that VQ-prior-based face restoration artifacts are highly discriminative.' However, FFDNet is a denoising network, not a VQ-prior-based face restoration model. CodeFormer is the VQ-prior-based model. This appears to be a mislabeling of the transforms in the analysis.","section":null},{"comment":"Figure 3: The caption states 'Average PSNR/SSIM between the original and each transform... for the real and fake samples.' It is unclear if these averages are computed over the entire dataset or just the single real and fake sample shown in the figure. Clarifying this would improve reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern regarding the backbone confound is the primary issue here. The paper's core contribution is the generative residual framework, but by upgrading the backbone to PE-Core while leaving baselines on CLIP, the authors have made it impossible to attribute the performance gains to their method rather than the backbone. This is a load-bearing issue for the central claim. If the authors can show that GenRes++ with PE-Core still outperforms FreLens or C2P-CLIP with PE-Core, the paper would be significantly strengthened. As it stands, the claim 'outperforms existing methods' is not rigorously supported."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The headline comparison is confounded by backbone choice: GenRes++ uses PE-Core-G14-448 while all baselines use CLIP, and the paper never re-runs baselines with PE-Core to isolate the generative residual contribution from the backbone upgrade.","headline":"The headline comparison is confounded by backbone choice: GenRes++ uses PE-Core-G14-448 while all baselines use CLIP, and the paper never re-runs baselines with PE-Core to isolate the generative residual contribution from the backbone upgrade.","tokens_in":19633,"tokens_out":146,"would_cite":false,"duration_ms":173647,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Re-Processing Images Exposes AI Fakes Across All Generator Families","keywords":[],"falsifier":"Construct or discover a generator whose images produce transformed outputs statistically indistinguishable from transformed real images across all five transforms (EnlightenGAN, GFPGAN, Real-ESRGAN, FFDNet, CodeFormer). If the generative residual between original and all five transformed variants is below the detection threshold for such images, the core mechanism fails and the framework cannot detect that generator.","tokens_in":18771,"feed_emoji":"🔍","tokens_out":960,"duration_ms":278710,"temperature":0.7,"pith_summary":"The paper argues that AI-generated images carry a hidden signature that is invisible when you look at the image alone but becomes detectable when you run the image through a second generative model — a super-resolver, denoiser, or face restorer. Real photos and synthetic images respond differently to this secondary processing because the synthetic image already encodes the statistical biases of its source generator, and those biases interact with the assumptions of the transform in a measurable way. The authors call this differential response a generative residual. They build a detector around it: encode the original and its transformed variants with a frozen vision transformer, model the multiplicative cross-feature interactions between original and transformed embeddings with a neural tensor network, and use cross-attention to aggregate signals from five complementary transforms. The resulting system detects images from 19 generator families it never saw during training, achieving 95.7% accuracy and 99.1% average precision, surpassing all prior methods on the standard cross-generator benchmark.","feed_headline":"Re-Processing Images Exposes AI Fakes Across All Generator Families","feed_subtitle":"Running a second generative model on an image reveals whether it was AI-made, enabling detection of 19 unseen generators at 95.7% accuracy.","key_machinery":"The architecture has four components: (1) a frozen PE-Core vision encoder with LoRA rank-6 adaptation producing 1536-dim embeddings projected to 256 dims; (2) five fixed generative transforms — EnlightenGAN, GFPGAN, Real-ESRGAN, FFDNet (σ=25), and CodeFormer — applied to every input; (3) a cross-attention aggregation module treating the original embedding as query and the five transformed embeddings as keys/values; (4) a neural tensor network computing f = tanh(z₀ᵀ W g_agg + [z₀; g_agg]ᵀ V + b) with W ∈ ℝ^{256×256×128}, producing a 128-dim relational vector fed to a binary classification head. Training uses only ProGAN images; the LoRA adapters, projection, CCA, NTN, and head are the only训练d","core_discovery":"The central discovery is that the generative residual — the discrepancy between an image and its output under a secondary generative transform — is a cross-generator detection signal. Unlike generator-specific spectral or upsampling artifacts, this differential response arises from the interaction between two generative priors (the source generator and the transform) and is empirically present across both GAN and diffusion families. By explicitly modeling multiplicative cross-feature dependencies between original and transformed embeddings via a neural tensor network, and aggregating across five diverse transforms with attention, the detector generalizes to unseen generators without learning","pith_inferences":[],"forward_implications":["If generative residuals are truly a universal property of synthetic images, then any future generator — regardless of architecture — should be detectable without retraining, as long as the transform set exposes the residual. This would shift the detection arms race from matching generators to selecting transforms.","The cross-attention aggregation mechanism implies that not all transforms are equally informative for every image. An adaptive transform-selection policy could reduce the five-transform computational overhead while preserving detection power.","If the residual signal generalizes to video, temporal consistency of generative residuals across frames could provide a stronger and more robust detection cue than per-frame analysis, since temporal artifacts would compound.","The framework's reliance on fixed, manually selected transforms creates a known attack surface: a generator designed to minimize residual discrepancy under these specific five transforms could evade detection, motivating adversarial or learned transform sets."],"fun_headline_variants":["Generative Residuals Expose AI Images From Unseen Generators","Image Transform Gap Flags AI Fakes Across Generator Families","Dual-Transform Residuals Detect AI Images Without Generator-Specific Training","Cross-Generator Residual Signal Catches AI Fakes From Unseen Models","Relational Features Between Original and Transformed Images Expose AI Fakes"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that the differential response of synthetic images under secondary generative processing is a universal property that will hold for any future generator architecture, not just the 19 in the benchmark. If a new generator produces images whose statistical structure is indistinguishable from natural images under all five transforms, the detection signal vanishes.","fun_headline_variants_meta":{"raw":{"variants":["Generative Residuals Expose AI Images From Unseen Generators","Image Transform Gap Flags AI Fakes Across Generator Families","Dual-Transform Residuals Detect AI Images Without Generator-Specific Training","Cross-Generator Residual Signal Catches AI Fakes From Unseen Models","Relational Features Between Original and Transformed Images Expose AI Fakes"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":601,"prompt_tokens":525,"completion_tokens":76,"prompt_tokens_details":null},"tokens_in":525,"tokens_out":76,"duration_ms":59094,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T03:11:14.019839+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Construct or discover a generator whose images produce transformed outputs statistically indistinguishable from transformed real images across all five transforms (EnlightenGAN, GFPGAN, Real-ESRGAN, FFDNet, CodeFormer). If the generative residual between original and all five transformed variants is below the detection threshold for such images, the core mechanism fails and the framework cannot detect that generator.","supporting_citations":[],"review_version":1}