{"id":"cde4ce5f-11c9-4f30-8df1-ca4c63946d3b","arxiv_id":"2608.00695","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A language-free zero-shot anomaly detector that scores patches by their relative distance to normal and anomaly anchors in a source-derived channel-spectral space, trained on auxiliary source categories only.","lead":"FreqAnchorAD is a vision-only system that detects defects in new image categories by measuring how strongly local image patches deviate from stored normal and abnormal references in a frequency-transformed feature space. It reports state-of-the-art anomaly detection and localization results on thirteen industrial and medical benchmarks without using any target-category training data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FDAP's frequency transform is not shown to be load-bearing: Table 7 shows Identity matches DCT, and Table 5 shows a plain MLP nearly matches FDAP, so the paper's central frequency-anchoring mechanism is unsupported.","rationale":"The reader's weakest assumption concerns the transferability of the source-derived channel permutation. My concern is more direct: the paper's own ablations indicate that the frequency transform in FDAP is not load-bearing, and there is no ablation isolating the source-derived channel order from a trivial fixed order. This does not necessarily invalidate the reported SOTA numbers, but it does undermine the stated central contribution and the reader's attribution of the gain to 'channel-spectral anchor projection.' The appropriate response is to require the authors to demonstrate, with a controlled ablation, that the frequency transform and the source-derived order are actually responsible for the reported improvements. If they cannot, the paper should be reframed or rejected for overclaiming the mechanism. I therefore keep the CONDITIONAL verdict but with a sharper condition than the reader's.","tokens_in":17182,"tokens_out":7334,"duration_ms":73166,"concrete_test":"Re-run the FDAP ablations on the full benchmark suite (or at least VisA→MVTec and MVTec→VisA) with four variants: (a) source-derived permutation + DCT (default), (b) source-derived permutation + Identity, (c) random fixed permutation + DCT, and (d) original channel order + DCT. If (b), (c), and (d) all match (a) within about 0.2 AUROC on P-AUROC and I-AUROC, then neither the DCT nor the source-derived channel order is load-bearing, and the central contribution must be reframed as a learned anchor-projection method rather than a frequency-deviation method.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim includes not only the benchmark numbers but also the attribution of the gains to FDAP's channel-spectral anchor projection. That attribution is not supported by the paper's own ablations. In Table 7, replacing the 2D DCT in Eq. (7) with the Identity transform changes P-AUROC by +0.1 (91.6 vs 91.5), P-AP by −0.7, I-AUROC by −0.1, and I-AP by −0.0. Thus the 'spectral' part of the channel-spectral representation contributes essentially nothing on the VisA→MVTec direction. Table 6 similarly shows that a 1D sequence loses only 0.4 P-AUROC and 1.0 P-AP relative to the default 32×32 grid, so the 2D spectral layout is not critical. Table 5 shows that LFCM + Plain MLP achieves 90.9/42.6/91.4/96.1 on VisA→MVTec versus 91.1/42.8/91.6/96.3 for LFCM + FDAP, so the FDAP-specific projection adds only about 0.2 points. This is consistent with the mathematical observation that a fixed orthogonal transform before a learned linear/MLP layer can be absorbed into that layer. The paper never ablates the source-derived channel permutation against a random fixed permutation or the original channel order, so the 'source-derived channel coordinate' could be arbitrary. Because the paper's title and central contribution rest on frequency-deviation anchoring, and the evidence shows the frequency transform is not the active ingredient, the mechanism is not load-bearing as claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FreqAnchorAD, a language-free zero-shot anomaly detection method built on frozen CLIP/DINOv2 backbones. It combines a Local Frequency Compensation Module (LFCM) that injects local DCT statistics into patch tokens, a Frequency-Deviation Anchor Projector (FDAP) that canonicalizes channel order by source-derived variance and applies a 2D DCT plus anchor-relative cosine scoring, and Asymmetric Anchor Supervision (AAS) that trains normal and anomaly anchors. Experiments on six industrial and seven medical benchmarks report state-of-the-art mean image-level and pixel-level metrics relative to text-guided baselines and VisualAD, with consistent gains for both CLIP and DINOv2 backbones.","tokens_in":17562,"tokens_out":3704,"duration_ms":34135,"significance":"If the reported benchmark numbers are reproducible, FreqAnchorAD is a competitive language-free ZSAD method, and the cross-dataset evaluation protocol is appropriate. The paper is largely self-contained, the ablation logic is internally consistent, and the diagnostic observation that anomaly evidence is not universally high-frequency (Fig. 1) is a useful descriptive contribution. However, the paper's central mechanistic claim—that frequency-deviation anchoring is the active ingredient—is not supported by its own ablations: the DCT basis is nearly interchangeable with Identity, and a plain MLP almost matches FDAP. This substantially weakens the contribution as currently framed, even if the accuracy claims hold.","major_comments":[{"comment":"Under VisA→MVTec, replacing the 2D DCT with Identity changes P-AUROC by +0.1, P-AP by −0.7, I-AUROC by −0.1, and I-AP by −0.0. Since a fixed orthogonal transform before a learned linear/MLP layer can be absorbed into that layer, the paper provides no evidence that the frequency basis itself is load-bearing. The title-level claim of 'frequency-deviation anchoring' therefore needs either a decisive counterexample or a revised framing.","section":"Ablation Studies, Table 7"},{"comment":"LFCM + Plain MLP scores 90.9/42.6/91.4/96.1 versus LFCM + FDAP at 91.1/42.8/91.6/96.3 on VisA→MVTec, a difference of at most 0.2 points on every metric. The Table 3 ablation removes the entire FDAP block (projection head, anchors, and spectral transform), so it cannot isolate frequency-anchoring. The paper needs either a more decisive ablation or a more modest attribution of the gains.","section":"Ablation Studies, Table 5"},{"comment":"The source-derived channel permutation π is never ablated against a random fixed permutation or the native channel order. Table 4 varies only the soft-canonicalization coefficient ρ while keeping the same π, so the transferability of the channel coordinate from source to target categories is untested. This is a load-bearing gap because the method's inference relies on a fixed ρ=1 canonical coordinate computed from source statistics.","section":"Method, Eq. (3)"},{"comment":"X_s,a is defined as all patch tokens extracted from anomalous source images, not only tokens inside anomalous regions. Consequently the variance statistic v_c in Eq. (3) mixes normal-background and anomalous content, which clouds the interpretation of the resulting canonical coordinate. Please clarify whether this is intentional and report whether using only anomalous-region tokens changes the results.","section":"Method, Eq. (2)"}],"minor_comments":[{"comment":"The image-level fusion uses λ_map=0.5 for CLIP and disables the CLS branch for DINOv2 and medical configurations; this detail should appear in the main text rather than only in the experimental setup paragraph.","section":"Experiments, Eq. (17)"},{"comment":"The FE-CLIP row has missing entries for KSDD2 and some metrics; please clarify why these are omitted and how the average is computed when entries are missing.","section":"Table 1"},{"comment":"The per-category sample counts (n=...) and the bracketed percentages are difficult to read at reproduction size; a table or larger font would improve clarity.","section":"Figure 1"},{"comment":"The paper repeatedly refers to the Appendix for implementation details, but the reviewed text does not include it; please ensure that the final version contains all hyperparameters, training schedules, and the full DINOv2 ablation results.","section":"Appendix"}],"recommendation":"major_revision","confidential_remarks":"The benchmark results are potentially valuable, but the central contribution is currently framed around a frequency-anchoring mechanism that the paper's own ablations do not support. A revision should either provide strong evidence that the channel-spectral transform and source-derived permutation matter, or honestly reframe the contribution around channel canonicalization and anchor-based scoring. I recommend major_revision rather than rejection because the accuracy claims, if reproducible, may still be of interest to the ZSAD community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real empirical step in zero-shot anomaly detection, but the title and main story overstate the frequency mechanism. FreqAnchorAD builds on VisualAD, adds a local frequency compensation module, an anchor projection head, and asymmetric supervision, and reports strong numbers on 13 industrial/medical benchmarks with frozen CLIP and DINOv2 encoders, no text, no target data. That is a solid package.\n\nCredit where due: the image-domain analysis in Fig. 1, showing that low- and mid-frequency deviations are common, is a useful descriptive observation and a legitimate motivation. The LFCM component appears to help, especially for localization (Table 5: raw token space -> LFCM token space gains ~7.5 P-AUROC on VisA→MVTec). The experimental scope is broad, the protocol is clean, and the ablations are mostly internally consistent. Removing FDAP (the whole projection) causes the largest drop, and the cross-direction results behave sensibly.\n\nThe soft spot is real, and the stress-test note is right: the specific frequency transform is not load-bearing. In Table 7, replacing the 2D DCT with Identity changes P-AUROC by +0.1, P-AP by −0.7, I-AUROC by −0.1, I-AP by 0.0. That is noise. Table 5 shows a plain MLP within 0.2 points of FDAP. A fixed orthogonal transform before a learned MLP can be absorbed into the MLP, so this result is expected. The paper also never ablates the source-derived channel permutation against a random fixed permutation, so the 'source-derived channel coordinate' remains untested. That matters because the paper's title, abstract, and contribution list all rest on frequency-deviation anchoring.\n\nOther soft spots are minor by comparison: no error bars or multiple seeds, and several SOTA differences vs VisualAD are under one AUROC point; the FE-CLIP exclusion from medical benchmarks gets a one-line justification; and the Fig. 1 diagnostic is not described in enough detail to reproduce (band definitions, normal-calibration procedure, per-category statistics). These are addressable.\n\nWho it's for: anyone working on language-free zero-shot AD. It's a useful empirical baseline and a reasonable extension of VisualAD. It deserves serious peer review, but not as-is; the authors should be pushed to either demonstrate the frequency basis does real work or rewrite the contribution around what actually moves the needle: LFCM, the learned anchor-relative projection, and AAS. I'd also require code and seeds.","headline":"Solid empirical package with an overplayed frequency story: the DCT and channel-permutation mechanism ablates to noise, so the paper needs a claim reset or stronger evidence before acceptance.","tokens_in":18102,"tokens_out":3432,"would_cite":true,"duration_ms":30283,"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 language-free, frequency-aware zero-shot anomaly detector claims state-of-the-art image- and pixel-level results across thirteen industrial and medical benchmarks by organizing channel responses along a source-derived coordinate.","keywords":["zero-shot anomaly detection","frequency-domain modeling","channel canonicalization","frequency-deviation anchor projector","defect localization","CLIP","DINOv2","anchor supervision"],"falsifier":"A cross-source permutation test: compute the FDAP channel permutation from one auxiliary source (e.g., VisA), then evaluate on a target using a permutation derived from a different source (e.g., MVTec) while keeping everything else fixed. If P-AUROC and P-AP fall to the plain-MLP level, the source-derived coordinate is target-specific rather than transferable, and the paper's central mechanism fails.","tokens_in":16851,"feed_emoji":"🔍","tokens_out":6879,"duration_ms":55383,"temperature":0.7,"pith_summary":"The paper tries to establish that a zero-shot anomaly detector can beat prior state of the art on both image-level recognition and pixel-level localization without any target-domain training, text prompts, or test-time adaptation, by making frequency-wise deviations—not just high-frequency edges—the explicit basis of discrimination. It argues that local defects shift responses across low, middle, and high frequency bands, so a model should organize evidence across all bands rather than assume anomalies are high-frequency. The proposed FreqAnchorAD sorts frozen visual features along a channel order derived from source categories, applies a 2D DCT over that order to form a channel-spectral anchor space, and scores patches by relative cosine similarity to normal and anomaly anchors. Across thirteen industrial and medical benchmarks, it reports the best mean image- and pixel-level scores, with the largest ablation drop when the frequency-deviation anchor projector is removed.","feed_headline":"Frequency-anchored zero-shot detector beats prior art on 13 benchmarks","feed_subtitle":"It sorts frozen visual features by source-channel variance and scores against normal and anomaly anchors—no target training.","key_machinery":"The load-bearing mechanism is the Frequency-Deviation Anchor Projector (FDAP), built on source-derived channel canonicalization. From source patch tokens at each layer, the method computes per-channel variance $v_c$ and the ascending permutation $\\pi$ (Eq. 3); at inference every target token is reordered by $\\pi$ (full canonicalization, $\\rho=1$, while training uses soft interpolation $\\rho=0.6$), reshaped into a balanced $H_c\\times W_c$ grid (e.g. $32\\times32$ for $D=1024$), and transformed by a 2D DCT over that channel coordinate, not over image space. After LayerNorm and a residual MLP projection, each patch is scored $\\cos(z_i, a_a) - \\cos(z_i, a_n)$ against learnable normal and anomaly anchors, and Asymmetric Anchor Supervision aligns normal patches strongly while weakly regularizing real and synthetic anomaly patches and separating the two anchors.","core_discovery":"Local anomalies are not universally high-frequency phenomena; they induce deviation patterns in low-, mid-, and high-frequency bands relative to normal references. The paper's central discovery is that these multi-band deviations can be exploited for zero-shot anomaly detection by (1) compensating frozen patch tokens with local spatial-frequency cues, (2) canonicalizing channels according to source-derived variance so that a spectral transform operates on a stable coordinate rather than arbitrary channel indices, and (3) measuring anomaly evidence as relative similarity to normal and anomaly anchors in the resulting channel-spectral space. On the cross-dataset protocol (VisA as source for MVTec, MVTec as source for VisA), the CLIP-based model reports I-AUROC 93.4 and P-AUROC 91.5 on MVTec, and the DINOv2 model reports P-F1 48.7 and P-AP 47.5 on MVTec, with average gains over VisualAD baselines across industrial and medical benchmarks.","pith_inferences":["Editorial: The paper's ablations always use the source dataset's own channel statistics; a direct test of whether a permutation computed from an unrelated source (e.g., using a VisA-derived permutation while evaluating with MVTec as source) preserves performance would separate true transferability from a source-specific artifact.","Editorial: Because the DCT operates on a reordered channel grid, the same anchor-relative scoring could in principle work with other orthogonal bases or with a learned permutation from a small target-normal sample; such variants would probe how much of the gain comes from the ordering versus the fixed spectral basis.","Editorial: The multi-band finding suggests the frequency-deviation representation may transfer to medical modalities where anomalies are subtle texture or boundary changes; the reported medical gains (e.g., OCT17 I-AUROC 95.0 versus 88.9 for the VisualAD baseline) point to a testable extension on X-ray or fundus datasets."],"forward_implications":["Zero-shot anomaly detection does not require a language branch: purely visual, frozen backbones suffice, which simplifies deployment to unseen categories.","Frequency-aware organization can improve pixel-level defect localization more than image-level recognition, since the largest gains in ablations are in P-AP and P-F1.","The source-derived channel coordinate transfers across industrial and medical domains, suggesting that channel variance ordering carries domain-independent structure useful for anomaly scoring.","Because FDAP uses a frozen encoder, the framework can be re-instantiated with new backbones (CLIP and DINOv2 both work) without retraining the encoder.","Anomaly evidence is spread across spectral bands, so systems that only emphasize high-frequency detail may miss a substantial share of defects; in the paper's analysis, 60.25% of anomaly-localization sensitivity is not high-frequency dominant."],"supporting_citations":[{"why":"VisualAD is the direct language-free baseline that FreqAnchorAD extends and compares against in both CLIP and DINOv2 variants.","marker":"Hou et al. 2026"},{"why":"Frozen CLIP ViT-L/14@336px is the primary visual backbone for the industrial and medical experiments.","marker":"Radford et al. 2021"},{"why":"Frozen DINOv2 ViT-L/14 is the second backbone used to demonstrate backbone generality.","marker":"Oquab et al. 2024"},{"why":"FE-CLIP is the frequency-aware ZSAD baseline whose feature-enhancement-only use of frequency cues motivates FDAP's explicit channel-spectral anchoring.","marker":"Gong et al. 2025"},{"why":"Supplies the 2D DCT basis applied over the source-derived channel coordinate in FDAP.","marker":"Ahmed, Natarajan, and Rao 1974"},{"why":"MVTec AD is the principal industrial target benchmark and one of the two datasets in the bidirectional cross-dataset protocol.","marker":"Bergmann et al. 2019"},{"why":"VisA is the source dataset for all non-VisA targets and the target when MVTec is the source.","marker":"Zou et al. 2022"}],"fun_headline_variants":["Zero-shot anomaly detection with multi-band frequency anchoring","FreqAnchorAD: frequency deviation anchors for zero-shot defect detection","Zero-shot anomaly detection via frequency deviation, not pixels","Multi-band frequency deviations reveal zero-shot anomalies","FreqAnchorAD: frequency anchoring for zero-shot defect detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the channel-variance order computed from source categories is a stable, transferable coordinate system for unseen target categories, because the method fixes that permutation at inference and scores every patch in the reordered space.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot anomaly detection with multi-band frequency anchoring","FreqAnchorAD: frequency deviation anchors for zero-shot defect detection","Zero-shot anomaly detection via frequency deviation, not pixels","Multi-band frequency deviations reveal zero-shot anomalies","FreqAnchorAD: frequency anchoring for zero-shot defect detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3331,"prompt_tokens":1033,"completion_tokens":2298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":2220}},"tokens_in":649,"tokens_out":2298,"duration_ms":14735,"temperature":1.0,"reasoning_tokens":2220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:17:43.211333+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A cross-source permutation test: compute the FDAP channel permutation from one auxiliary source (e.g., VisA), then evaluate on a target using a permutation derived from a different source (e.g., MVTec) while keeping everything else fixed. If P-AUROC and P-AP fall to the plain-MLP level, the source-derived coordinate is target-specific rather than transferable, and the paper's central mechanism fails.","supporting_citations":[{"cited_title":"Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month=","cited_arxiv_id":null,"evidence_quote":"VisualAD is the direct language-free baseline that FreqAnchorAD extends and compares against in both CLIP and DINOv2 variants."},{"cited_title":"Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month=","cited_arxiv_id":null,"evidence_quote":"FE-CLIP is the frequency-aware ZSAD baseline whose feature-enhancement-only use of frequency cues motivates FDAP's explicit channel-spectral anchoring."}],"review_version":2}