{"id":"01e2822a-a309-4755-b446-0522fc0f34a9","arxiv_id":"2607.22087","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Training with a scene-composition feature appended to the classifier head and removed at inference improves cross-generator fake-image detection by up to 8.04% on GenImage.","lead":"This preprint proposes FAIR, a training-only method that adds scene-composition features to an AI-image detector's classifier and then discards them at inference. It reports up to 8.04% higher zero-shot accuracy on unseen image generators, with no extra cost when the detector is deployed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SCS domain invariance is asserted but never quantitatively tested; the only table cited (Tab. 4) measures source-label dependence, not cross-domain shift, so the invariant anchor at the core of FAIR's mechanism is unverified.","rationale":"I agree with the reader that the weakest point is the unverified domain invariance of SCS. This is load-bearing because the entire proposed mechanism—learning a smooth boundary by tilting into a stable X* dimension and then slicing it off—requires X* to be stable across generator families. The paper's own evidence (Fig. 4b, Tab. 4) is insufficient: UMAP is qualitative and can mask shifts, and CKA with labels is a different quantity from cross-domain distributional similarity. Moreover, the paper creates an internal tension by claiming I(X*;Y)≈0 while also reporting large gains; if X* were truly label-independent it is hard to see how its addition could improve training. The proposed test would settle both issues by measuring, first, whether X* distributions separate by domain and, second, whether the gains track this separation. Because the reader's verdict is already CONDITIONAL and this concern is the same one, I recommend no change to the verdict.","tokens_in":16746,"tokens_out":13449,"duration_ms":155520,"concrete_test":"Quantify the actual domain shift of SCS features. Using the released (or reimplemented) SCS extractor with N=1024, extract features from a sample of the GenImage SDv1.4 source and from each target generator (e.g., BigGAN, Midjourney, ADM, SDXL, GLIDE). Train a held-out linear classifier (or small MLP) to discriminate source vs. each target from SCS features alone, and report per-target AUC. If AUC significantly exceeds 0.5 (e.g., >0.7 for any target), P(X*|D_source)≠P(X*|D_target) and the invariance premise is false. Complement with a kernel MMD two-sample test between source and target SCS distributions. Finally, correlate per-domain FAIR gains with these shift measures: if gains are concentrated in low-shift domains, the 'invariant anchor' explanation is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FAIR improves cross-generator generalization depends on the premise that SCS features X* form a domain-invariant anchor: §3.4 asserts P(X*|D_source)≈P(X*|D_target). The evidence offered is Fig. 4b (qualitative UMAP) and Tab. 4, which reports CKA of SCS with the real/fake label on SDv1.4 and ProGAN *training* sets. CKA with Y on the training set is a measure of label dependence, not a two-sample comparison of feature distributions across domains; it cannot demonstrate invariance. If X* shifts systematically with the generator (e.g., SD vs. BigGAN), the 'tilting into stable X* dimensions' mechanism collapses: training would embed source-specific SCS statistics into W_X, and discarding X* at inference would remove the very structure the model was regularized against. The observed gains on the five benchmarks could then reflect structural similarity among those particular test generators, not a general invariant prior. Without a quantitative cross-domain shift test, the central mechanism is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FAIR (Feature-Augmented Implicit Regularization), a training-time method that concatenates a hand-crafted Scene Composition Structure (SCS) feature vector to the penultimate features of an AI-generated-image detector, trains the enlarged classification head, and then discards the SCS branch before deployment by slicing the weight matrix. The authors argue that SCS is domain-invariant and label-independent, so the optimizer learns a 'tilted' low-complexity decision boundary that projects back to a smoothed, generalizing boundary in the original feature space. FAIR is applied to two recent detectors, AIDE and PatchCraft, and evaluated on GenImage, AIGCDetect, UnivFD, Fake2M, and DRCT-2M. The headline results are large gains on GenImage (PatchCraft +8.04 mean accuracy, AIDE +4.13) and consistent gains on DRCT-2M and UnivFD, with some degradations on Fake2M and AIGCDetect.","tokens_in":17093,"tokens_out":5585,"duration_ms":64118,"significance":"If the claimed effect is real, FAIR would be a valuable and cheap addition to AIGC detection: it improves cross-generator transfer with zero inference-time overhead and without changing the deployed architecture. The manuscript has several strengths: it tests two architecturally distinct detectors, covers a large number of target domains, includes a PyTorch implementation, and compares against traditional regularizers and alternative feature priors. The core idea of using a macro-structural prior in a Learning-Using-Privileged-Information style is interesting and the empirical gains on GenImage and DRCT-2M are substantial. However, the central mechanism rests on an unverified domain-invariance assumption, and the reported comparisons are confounded by an unequal training protocol; these issues prevent the paper from being accepted in its current form.","major_comments":[{"comment":"The comparison changes two variables at once. The text states that 'the base model's primary feature extractor remains entirely frozen' during FAIR training, so only the expanded classification head (and the SCS projector) is trained. The 'Base' results appear to come from the standard end-to-end training of the full detector. Freezing the backbone and training only the head (linear probing) is itself a well-known and often strong regularizer for domain shift. The paper does not include the necessary control: the same head-only training on the frozen features without SCS, or with a random/constant auxiliary input. Without this control, the large improvements in Table 1 and the ablation in Table 5 cannot be attributed specifically to the SCS prior. Please add this ablation for both base detectors and both training regimes.","section":"§3.3 (FAIR Implementation), Tables 1 and 5"},{"comment":"The domain-invariance claim is load-bearing but not measured. The mechanism assumes P(X*|D_source) ≈ P(X*|D_target), yet the only quantitative evidence in Tab. 4 is CKA between SCS features and the real/fake label on the source training sets. CKA with the label is a test of label dependence, not a two-sample comparison of SCS distributions across generators. Fig. 4b is a qualitative UMAP projection. If SCS shifts systematically across generators (e.g., SD vs. BigGAN), the 'invariant anchor' does not exist and the tilting explanation in Fig. 1 collapses. Please provide a direct quantitative cross-domain invariance test—for example, MMD, Wasserstein distance, or A-distance between SCS distributions from the source and several held-out generators—and report the numbers.","section":"§3.4, Fig. 4, Tab. 4"},{"comment":"The claim that FAIR 'consistently improves' cross-generator generalization is contradicted by the paper's own aggregate numbers. In Table 1, PatchCraft+FAIR degrades on Fake2M (80.46→78.82) and AIDE+FAIR degrades on AIGCDetect (93.02→92.14). No error bars or multiple seeds are reported anywhere, so for the small aggregate gains (e.g., +0.77 for PatchCraft on UnivFD) it is impossible to tell whether the difference is significant. The headline 'up to 8.04%' is the single largest improvement and is not representative of the overall behavior. Please report mean±std over at least three seeds for the aggregate and per-domain results, and temper the consistency claim accordingly.","section":"Table 1; Tables S9, S10"}],"minor_comments":[{"comment":"The abstract reports 'up to 8.04%' while the contributions list says 'up to 8.10%'. Please reconcile.","section":"Abstract vs. Section 1"},{"comment":"The header 'Lower is more invariant' is misleading: the reported CKA is with the label, which is a label-dependence score, not a domain-invariance measure. Rename the column and explain what the score is actually measuring.","section":"Table 4"},{"comment":"The text equates CKA with I(X*;Y) and calls it a 'non-parametric measure' of 'conditional dependence'. CKA is a representation similarity index and is not a mutual-information estimator; the statement is technically imprecise. Please either define the actual estimator used or rephrase.","section":"§5.3, Tab. 4"},{"comment":"The projection dimensions (N,M) vary across configurations without justification, and no sensitivity analysis is reported. Since M controls the capacity of the auxiliary branch, a small sweep around the chosen values would strengthen the claim that the method is robust to this hyperparameter.","section":"§4 / Table S7"},{"comment":"The 'tilting' and 'projection' language is illustrative. Consider labeling Fig. 1 as a conceptual schematic and moving the formal description of the weight slicing to a numbered equation for clarity.","section":"Fig. 1 and §3.4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Call it a useful empirical trick that partly outruns its own explanation. The new thing here is concrete: take a frozen AIGC detector, concatenate SCS features to the final-layer input during training, then slice the weights and discard SCS at inference. That is a clean, zero-overhead deployment story, and on GenImage the gains are large — +8.04 for PatchCraft, +4.13 for AIDE. The ablation against L1/L2/Dropout and against alternative priors (random, LPIPS, Sobel, Laplacian) is the right comparison and SCS wins on mean accuracy. Credit where due: this is a sensible empirical paper and the training-only design makes the claim falsifiable.\n\nThe soft spots are real but not fatal. First, the \"consistently improves\" language is overstated. On Fake2M, PC drops from 80.46 to 78.82; on AIGCDetect, AIDE drops from 93.02 to 92.14. The detailed tables show several domain-level drops (PC on StarGAN: 88.33 to 79.17; on CogView2: 96.84 to 76.97). So the headline is \"mostly improves, sometimes sharply, occasionally hurts.\" Second, the mechanism in Section 3.4 rests on P(X*|source) ≈ P(X*|target), but the only quantitative support (Tab. 4) is CKA with the label on the source training set. That measures label dependence, not cross-domain shift. The UMAP in Fig. 4b is qualitative. The stress-test note has this right: the invariant anchor is asserted, not demonstrated. If SCS does shift across generators, the \"tilting\" story is unsupported — though the empirical gains would still stand on their own. Third, no error bars or seeds, and the supplementary ships only a wrapper snippet, not the SCS extractor itself. That makes the numbers hard to verify.\n\nMy verdict: the paper deserves a serious referee. The idea is novel in this subfield, the deployment story is clean, and the empirical claim is testable. But the invariance assumption needs a quantitative two-sample test across target domains, and the authors should report seed-averaged results and release the SCS code. I would not cite it yet — I want to see the code and the invariance numbers first. It would be a decent reading-group paper for a discussion of LUPI and what counts as evidence for domain invariance.","headline":"A training-only SCS prior gives real cross-generator gains on some benchmarks, but the central domain-invariance claim is asserted, not shown — worth reviewing, not citable yet.","tokens_in":17516,"tokens_out":2177,"would_cite":false,"duration_ms":24494,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A training-only macro-structural feature, discarded at inference, improves AI-image detectors' cross-generator accuracy by up to 8.04 percent.","keywords":["AI-generated image detection","cross-generator generalization","implicit regularization","Scene Composition Structure","Learning Using Privileged Information","shortcut learning","domain invariance","zero-shot transfer"],"falsifier":"Compute SCS features for many generators (e.g., all GenImage domains plus unseen ones) and train a small classifier to predict the generating model from SCS. If held-out generator identity is recoverable with accuracy well above chance, the prior is not domain-invariant. Alternatively, measure the maximum mean discrepancy (MMD) between SCS distributions of source and target sets.","tokens_in":16664,"feed_emoji":"🕵️","tokens_out":4646,"duration_ms":44747,"temperature":0.7,"pith_summary":"F AIR (Feature-Augmented Implicit Regularization) claims that AI-generated image detectors fail on unseen generators because they overfit to generator-specific textural fingerprints. The paper's fix is to concatenate a domain-invariant 'Scene Composition Structure' (SCS) feature to the detector's latent features strictly during training, then slice it away at deployment. The resulting weight slice yields a smoother, better-calibrated decision boundary without changing the architecture or inference cost. On GenImage, FAIR raises PatchCraft's mean cross-domain accuracy by 8.04 points and AIDE's by 4.13 points; similar gains hold across four other benchmarks.","feed_headline":"Training-only scene prior boosts fake-image detection by 8%","feed_subtitle":"FAIR adds a domain-invariant composition prior at train time, none at inference, yet sharpens zero-shot accuracy.","key_machinery":"The key mechanism is the FAIR training wrapper, a LUPI-inspired augmentation: the final classification layer is trained on the concatenation [X, X*], where X is the base detector's feature vector and X* is the SCS prior projected by a learned GELU MLP. At export, the slice W_X is extracted and X* is never computed again. SCS itself is computed by recursive axis-parallel partitioning of the image into segments that maximize SSE reduction; the normalized cumulative gains form the N-dimensional prior. The paper's mechanistic evidence includes weight-norm compression (‖W‖₂ drops from 24.95 to 18.42), reduced weight-update magnitudes, lower target-domain loss, and suppression of overconfident mis","core_discovery":"The central claim: generalization in AIGC detection is a geometric problem, not a parameter-count problem. By tilting the separating hyperplane into an orthogonal, class-agnostic structural coordinate system (SCS) during training, FAIR forces the optimizer to discover a low-complexity boundary in the primary feature space. The prior is discarded at inference, and the sliced weights outperform the unregularized baseline on unseen diffusion and GAN domains. The paper argues that the prior's two properties — near-zero dependence on the label (I(X*;Y)≈0) and near-invariance across source and target domains (P(X*|D_source)≈P(X*|D_target)) — are what prevent shortcut learning while providing a sta","pith_inferences":["A testable extension: replace SCS with other macro-structural priors (depth maps, segmentation layouts) and check whether the same invariance holds; the paper's own Sobel/Laplacian ablation suggests local edge priors fail, but global layout priors may succeed.","The invariance premise implies a sharp falsification test: measure the SCS distribution shift between a large set of generators; if a linear classifier can recover the generator identity from SCS, the anchor is weaker than claimed.","If the mechanism is as general as argued, FAIR could regularize other texture-biased classification tasks (medical imaging, deepfake video) where a domain-invariant macro-structure exists.","The paper itself notes one failure mode: PatchCraft+FAIR drops below baseline at JPEG quality 50, where severe blocking artifacts deceive the structural anchor."],"forward_implications":["If FAIR is correct, any AIGC detector can be upgraded for cross-generator robustness by a training-only wrapper, with no architecture change or inference cost.","AIDE+FAIR reaches 91.01% mean accuracy on GenImage, the best reported on that benchmark, and PC+FAIR gains up to 18.55 points under mild JPEG compression.","The ablation implies that conventional L1/L2/Dropout regularization cannot produce these gains; the structural anchor is doing the work.","FAIR's effect is architecture-agnostic — it helps both a CLIP/DCT hybrid (AIDE) and a texture-patch model (PatchCraft).","The method preserves performance on source domains while improving targets, indicating reduced rather than catastrophic forgetting."],"fun_headline_variants":["A prior you throw away after training sharpens AIGC detection","Zero-overhead scene prior boosts zero-shot fake-image detection","8% better cross-generator fake-image detection with a train-only prior","FAIR's secret: a label-free scene prior that's dropped at inference","Implicit regularization via scene prior lifts fake-image detection by 8%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole argument rests on the claim that the SCS feature distribution is essentially the same for every generator, real and fake — supported only by a UMAP projection and CKA scores computed on training data; if SCS shifts across unseen generators, the invariant anchor disappears and the regularization argument collapses.","fun_headline_variants_meta":{"raw":{"variants":["A prior you throw away after training sharpens AIGC detection","Zero-overhead scene prior boosts zero-shot fake-image detection","8% better cross-generator fake-image detection with a train-only prior","FAIR's secret: a label-free scene prior that's dropped at inference","Implicit regularization via scene prior lifts fake-image detection by 8%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000857,"raw_usage":{"total_tokens":3562,"prompt_tokens":755,"completion_tokens":2807,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":2714}},"tokens_in":499,"tokens_out":2807,"duration_ms":20919,"temperature":1.0,"reasoning_tokens":2714,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:48:39.306623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute SCS features for many generators (e.g., all GenImage domains plus unseen ones) and train a small classifier to predict the generating model from SCS. If held-out generator identity is recoverable with accuracy well above chance, the prior is not domain-invariant. Alternatively, measure the maximum mean discrepancy (MMD) between SCS distributions of source and target sets.","supporting_citations":[],"review_version":1}