{"id":"a2d5741b-da2b-4c57-bf35-487c4b28aa9b","arxiv_id":"2504.14884","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MDD-Net combines a dual-decoder reverse distillation network with a class-aware memory module and reports state-of-the-art or near-state-of-the-art anomaly detection and localization on MVTec AD, VisA, Real-IAD, and Uni-Medical.","lead":"This paper proposes a network for detecting defects in images, trained only on normal examples, that combines two reconstruction decoders with a class-aware memory bank. A smart generalist might read it because it reports consistent benchmark gains across industrial and medical anomaly detection, a practical bottleneck for automated inspection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Identity decoder is trained on raw stage-4 tokens but evaluated on Neck-transformed features; RID may add false positives on normal images, undermining the false-positive-reduction claim.","rationale":"The reader's verdict already flagged the identity decoder's training/inference mismatch and the correlated-deficiency premise; this stress test elevates that from a side inconsistency to the load-bearing failure mode of the paper's central false-positive-reduction claim. The paper's contribution is not the memory module alone; the 3.6% improvement and the pixel-level gains are attributed to RID resolving hard-normal false positives. If the identity decoder is off-distribution at inference, RID can be a false-positive generator, and the ablation in Table 3 cannot rule this out because it evaluates the same mismatched pipeline. This is a concrete, addressable concern: a simple inference-only variant (B) would settle it. The paper otherwise has real strengths: multi-benchmark results, ablations, visualizations, and an honest limitations section. The concern does not warrant rejection; it warrants making the code/check available and reporting the variant before the mechanism's claim is accepted. Hence I keep the conditional verdict.","tokens_in":17981,"tokens_out":6125,"duration_ms":57575,"concrete_test":"On MVTec AD, train MDD-Net with the published Algorithm 1, then run two inference variants: (A) the paper's Phase 3 input Neck(F_T^4) for the identity decoder; (B) the training-consistent input F_T^4 for the identity decoder only. If reported P-mAP 66.4 / P-mIoU 48.8 change materially between A and B, or if identity-decoder per-pixel error on normal test images is substantially lower under B, the identity branch has not learned its inference input distribution and RID's false-positive cancellation is unverified rather than demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"MDD-Net's central claim that RID reduces false positives depends on the Section 3.2 assertion that two identically structured decoders have correlated reconstruction deficiencies, so discrepancies cancel on normal regions. The stated protocol does not establish this. In Algorithm 1 Phase 2 step 4, the identity decoder is trained on raw anomaly tokens F_a^4_T, with no Neck. In Phase 3 step 2, inference feeds it Neck(F_T^4), a learned linear transform. The identity decoder is also trained only on synthetic-anomaly images, never on normal-only images, so at test time on a normal image its input distribution is doubly off-distribution. If that shift makes f_I reconstruct normal patches poorly, RID = 1 - cos(F_R, F_I) becomes large on normal pixels - exactly the false positives the method claims to eliminate. The correlated-capacity argument does not save RID, because the two branches receive different inputs (restoration gets Neck(M(F_T^4)), identity gets Neck(F_T^4)) and have different training pathways; correlated capacity does not imply correlated errors under different input transforms. Table 3 ablations run inside the same mismatched protocol and cannot distinguish true cancellation from threshold artifacts, so the reported 3.6% MVTec AD gain and pixel-level improvements are not yet securely attributed to the mechanism claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MDD-Net, a multi-class unsupervised anomaly detection framework that combines a dual-decoder reverse distillation network (DRD-Net) with a class-aware memory module (CMM). A frozen teacher encoder extracts features from normal and synthetic anomalous images; the restoration decoder reconstructs normal features from memory-replaced anomaly tokens, while the identity decoder is intended to preserve anomaly semantics. Anomaly maps are formed by fusing a Restoration-Identity Discrepancy (RID) and a Teacher-Restoration Discrepancy (TRD). The method is evaluated on MVTec AD, VisA, Real-IAD, and Uni-Medical under eight image- and pixel-level metrics, with ablations on MVTec AD. The central claims are that MDD-Net simultaneously mitigates over-generalization and insufficient normal reconstruction, achieving a 3.6% average improvement over prior state-of-the-art on MVTec AD and large pixel-level gains on all benchmarks.","tokens_in":18282,"tokens_out":7266,"duration_ms":65624,"significance":"If the mechanism and the empirical claims are sound, MDD-Net would be a useful step toward a single unified model for multi-class anomaly detection and localization, directly addressing the named failure modes of over-generalization and insufficient normality reconstruction. The paper's strengths include its breadth of evaluation (four benchmarks and eight metrics) and the explicit component-wise ablations in Tables 3 and 4. However, the central mechanism for false-positive reduction is supported by an asserted rather than measured correlation between the two decoders, and the training/inference protocol for the identity decoder contains a concrete mismatch. The reported 'superior over SoTA' claim is also not uniformly supported by the numbers in Table 1, particularly on Real-IAD. These issues are load-bearing for the paper's main conclusions.","major_comments":[{"comment":"The identity decoder is trained on raw stage-4 tokens F_a^4_T (Algorithm 1 Phase 2 step 4 and Eq. 3), but at inference it receives Neck(F_T^4) (Phase 3 step 2). Figure 2's step 5 caption states that anomalous features are routed through Neck and Identity Decoder during training, which contradicts both the main text and Algorithm 1. This mismatch is load-bearing because the paper's central claim (Section 3.2, Section 3.5) is that RID cancels errors on normal regions thanks to the correlated deficiencies of two structurally identical decoders. When the identity decoder sees a different input transform at test time, and when it has never seen normal-only images during training, the correlated-error-cancellation argument does not apply; RID can become large on normal pixels. Please standardize the input pathway between training and inference and report RID statistics separately for normal pixels and anomalous pixels on real test images.","section":"Section 3.2, Algorithm 1, Figure 2"},{"comment":"The correlated-capacity argument is asserted rather than measured. The large improvement from DRD-Net+L over DRD-Net-L in Table 3 (P-mAP 61.91 vs 56.73, P-mIoU 44.90 vs 43.64) may be driven mainly by L_dist, which directly supervises M_RI with the synthetic anomaly mask via Eq. 5, rather than by emergent cancellation of correlated decoder errors on normal patterns. To separate these explanations, please add an ablation that removes L_dist while retaining the identity decoder, and directly measure decoder error correlation on hard-normal patches, for example by computing per-pixel TRD and RID on normal-only test images and reporting their joint distribution.","section":"Section 3.2, Table 3, Eq. 5"},{"comment":"The fusion ratio alpha (0.4, 0.4, 0.1, 0.5 for the four datasets) and memory bank size N (500, 1000, 2000) are reported without a validation protocol. Because alpha directly controls how much weight the proposed RID term receives, tuning it per benchmark on the test set can inflate the apparent gains of the method. The paper should either fix these hyperparameters across datasets, justify them on a held-out validation split, or report sensitivity curves showing that the conclusions are robust to the chosen values. In addition, no repeated-run statistics or error bars are provided, and some reported gains are small enough to fall within typical run-to-run variation.","section":"Section 4.1, Implementation Details"},{"comment":"Table 1 does not support the statement that MDD-Net is superior to current state-of-the-art approaches across all four benchmarks. On Real-IAD, MambaAD achieves 87.0 mAUROC, 85.3 image-level mAP, 98.6 max, 32.4 pixel mAUROC, 38.1 pixel mAP, 91.2 pixel mF1, and 23.9 mAU-PRO, while MDD-Net achieves 86.7, 85.1, 97.7, 30.8, 35.9, 86.5, and 22.3, respectively; MDD-Net is worse on every listed Real-IAD metric except a tie on mF1-max (77.6). On VisA, MDD-Net is also below MambaAD on image-level AUROC and pixel-level mF1. The abstract and introduction should be revised to present per-dataset comparisons accurately rather than claiming uniform SoTA performance.","section":"Table 1 and Section 4.2"},{"comment":"The method is described as 'unsupervised' in the title and abstract, but the CMM training uses class labels y through the classification loss L_cls in Eq. 11. This contradicts the Introduction's statement that training uses 'only normal samples.' The Limitations paragraph acknowledges reliance on category labels, but the framing is still misleading. Please reclassify the setting as label-guided or weakly supervised multi-class anomaly detection, or provide a clear justification for why the availability of class labels during training is compatible with the unsupervised claim.","section":"Section 3.3, Eq. 11, Limitations"}],"minor_comments":[{"comment":"The inner summation is written as 'D=1' to 'D' with the dummy index D colliding with the dimension D; please change the dummy index to c, i.e., sum_{c=1}^{D} exp(\\hat{y}^{(t)}_c).","section":"Eq. (11)"},{"comment":"References [1] and [2] are identical duplicates (both cite the same BMAD paper); one should be removed or replaced with the intended second citation.","section":"References"},{"comment":"In the L_rec term of Algorithm 1, the parentheses are mismatched: 'Dcos(f_R(Neck(M(F_n^4_T)), F_n^i_T)' should be 'Dcos(f_R(Neck(M(F_n^4_T))), F_n^i_T)'. Also, the notation Dcos is used in the supplement but not defined in the main text.","section":"Algorithm 1, Supplementary"},{"comment":"The sentence 'MDD-Net segments anonymous areas more accurately' should read 'anomalous areas.'","section":"Section 4.2"},{"comment":"The caption for step 5 says the anomalous features are reconstructed through Neck and Identity Decoder, which conflicts with Section 3.2 and Algorithm 1 (where the identity decoder receives raw F_a^4_T). Please align these descriptions.","section":"Figure 2 caption"},{"comment":"The sentence 'Divergences across datasets exclusively manifest in memory bank capacity and batch dimension configurations' is inaccurate because the fusion ratio alpha also differs per dataset; please correct it.","section":"Section 4.1"},{"comment":"The abstract states an 'average improvement of 3.6%' over state-of-the-art on MVTec AD; please specify the baseline and the set of metrics used to compute this average, since Table 1 shows different gains for different metrics.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper has substantial experimental breadth and a clear attempt to address two real failure modes, but the identity-decoder training/inference inconsistency and the unverified correlated-capacity assumption directly undermine the claimed mechanism. Additionally, the Real-IAD results in Table 1 contradict the paper's uniform-SoTA framing. These issues are fixable with additional experiments and re-analysis, so I recommend major revision rather than rejection. A version that retrains the identity decoder on the same input features used at inference, reports normal-pixel RID statistics, adds a validation-based hyperparameter protocol, and revises per-dataset claims would be much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before reading this one.\n\nFirst, the strong part: it is a properly engineered paper. The dual-decoder decomposition (restoration vs. identity) with class-aware memory prototypes carrying per-item class probability vectors is a genuinely new arrangement, and the experimental package is unusually thorough — four benchmarks, eight metrics, ablations, component ablation tables, and evaluation on held-out real test sets. The pixel-level gains are large: MVTec pixel mAP 66.4 against 55.4 for ViTAD, and Uni-Medical pixel mAP 57.8 against 48.2. That is a credible SOTA claim, conditional on code and repeated runs (no error bars anywhere).\n\nSecond, the soft spot, and it is the headline story, not a side detail. The paper's mechanism claim is that RID cancels false positives on normal regions because the two decoders have correlated deficiencies. That is asserted, not measured — there is no analysis of RID on normal-only images. The stress-test concern is real and I confirmed it on a careful read: Algorithm 1 Phase 2 trains the identity decoder on raw anomaly tokens with no Neck, but Phase 3 inference feeds it Neck-transformed features. The figure's training diagram even shows the Neck before the identity branch, so the paper is internally inconsistent about its own identity path. And the identity decoder only ever sees synthetic-anomaly images in training, so on a normal test image its input is doubly out-of-distribution. None of this necessarily sinks the empirical result, but it means the reported gains are not securely attributed to the RID cancellation mechanism, and the ablation tables run under the same mismatched protocol cannot settle it.\n\nTwo smaller items: the per-dataset alpha and memory size choices are presented as fixed settings rather than a validation protocol, and CMM uses class labels during training. That label use is disclosed in the limitations, and Table 4 suggests the class-aware part is only worth about 0.1 points over a conventional memory, so the fairness concern is probably minor — but the comparison against label-free baselines deserves a sentence.\n\nBottom line: this deserves a serious referee. The architecture and benchmark results justify referee time; the weakest section is the mechanism story, and the identity-branch inconsistency must be fixed, with code and a normal-image RID analysis requested. It is a conditional accept, not a reject.","headline":"Solid engineering result with large reported gains on four benchmarks, but the RID mechanism story is under-supported and the identity decoder has a genuine training/inference inconsistency that the authors need to fix.","tokens_in":18806,"tokens_out":7492,"would_cite":true,"duration_ms":62213,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual-decoder network detects anomalies across 15 classes at once, lifting pixel mAP on MVTec AD from 55.4 to 66.4.","keywords":["multi-class anomaly detection","unsupervised anomaly detection","feature reconstruction","reverse distillation","class-aware memory","dual decoder","anomaly localization"],"falsifier":"On a validation set of only normal images, compute per-pixel cosine distance maps between the restoration and identity decoders and measure the spatial correlation of their individual reconstruction errors; if the two error maps are weakly correlated, the RID signal cannot cancel normal-region false positives. As a second check, retrain with the identity decoder using an architecturally different decoder; if performance does not drop, the correlated-deficiency assumption is not load-bearing.","tokens_in":17788,"feed_emoji":"🔍","tokens_out":10804,"duration_ms":89805,"temperature":0.7,"pith_summary":"Reconstruction-based anomaly detection faces two opposite failure modes in multi-class settings: subtle or compositionally similar anomalies are reconstructed too well and escape detection, while intricate normal textures are reconstructed too poorly and produce false alarms. The paper proposes MDD-Net, which attacks both at once by training two structurally identical decoders side by side: a restoration decoder that turns synthetic anomalies back into normal features, and an identity decoder that deliberately preserves anomalous semantics. Because the two decoders share architecture and capacity, the paper argues their errors on hard normal patterns are correlated and cancel when the outputs are compared, leaving the Restoration-Identity Discrepancy as a cleaner anomaly signal. A class-aware memory module replaces anomalous tokens with learned normal prototypes, suppressing anomaly reconstruction directly. On MVTec AD the method reports image-level AUROC of 98.6 and pixel-level mAP of 66.4 against 55.4 for ViTAD, with similar gains on VisA, Real-IAD, and Uni-Medical.","feed_headline":"Twin decoders lift pixel-level anomaly maps from 55.4 to 66.4","feed_subtitle":"Their shared blind spots cancel, so one model can replace per-class detectors.","key_machinery":"The load-bearing object is the Restoration-Identity Discrepancy (RID), the summed cosine distance between the multi-stage feature maps produced by two identically structured ViT decoders. The Restoration Decoder receives teacher features after the class-aware memory module replaces each token with a sparse combination of learned normal prototypes, so it is trained to reconstruct normal features from anomaly-contaminated input; the Identity Decoder receives teacher features without that restoration objective and is trained to reproduce the anomalous semantics. The paper's argument is that because both decoders have identical structure and comparable capacity, their representational deficiencies are correlated, so on normal regions the two outputs fail in similar ways and the difference cancels, while on anomaly regions the restoration branch moves toward normality and the identity branch stays with the anomaly, producing a large discrepancy. The Class-aware Memory Module carries the second half of the mechanism: sparse cosine retrieval from a bank of class-labeled prototypes, a per-token classification loss that disentangles class-specific from shared prototypes, and gradient truncation that keeps anomaly features from contaminating the memory.","core_discovery":"The paper's central claim is that over-generalization and insufficient normality reconstruction are not a trade-off but two sides of the same decoder, and both can be corrected by decoupling what the decoder is asked to do. The restoration branch is trained to output normal teacher features after the memory module has stripped anomaly content from the input tokens; the identity branch is trained to preserve the anomaly semantics present in the teacher features. At inference, the pixel-wise cosine distance between the two branches, fused with the teacher-restoration distance, forms the anomaly map. The paper reports that this scheme improves multi-class anomaly detection and localization on all four benchmarks, with the largest margins at the pixel level: MVTec AD pixel mAP of 66.4 versus 55.4 for ViTAD, pixel mIoU of 48.8 versus 42.5, and a 3.6-point average image-level gain over the previous best method.","pith_inferences":["An implicit consequence of the correlated-deficiency argument is that the same dual-decoder trick should transfer to other dense reconstruction tasks where a student model has systematic blind spots; the paper does not test this.","A controlled experiment that feeds both decoders the same transformed input at training and inference has not been reported; the pseudocode routes the identity branch differently between the two phases, so the RID signal's dependence on that pipeline difference is untested.","The class-aware memory's visualization suggests cross-class texture prototypes are shared (leather and grid); a label-free clustering variant would clarify how much of the gain comes from class supervision versus from memory replacement alone.","Because the main reported wins are pixel-level, evaluating the method on anomaly benchmarks with finer-grained ground-truth masks beyond the four used here would show whether the localization gain generalizes."],"forward_implications":["A single trained model can replace per-class detectors in multi-class industrial inspection: on MVTec AD, one MDD-Net covers 15 classes with 98.6 image-level AUROC and 66.4 pixel mAP.","The largest reported gains are in localization, so the method is suited to tasks that need pixel-accurate anomaly maps, such as generating weak supervision for segmentation.","The same architecture transfers from industrial textures and objects to heterogeneous medical modalities without per-class retraining, with pixel mAP rising to 57.8 on Uni-Medical versus 48.2 for ViTAD.","Because the class-aware memory learns both class-specific and shared prototypes, class labels during training are part of the method, and the authors identify removing that label dependency as the main limitation.","The two discrepancy signals can be fused by simple linear combination and the losses summed directly, suggesting the method does not require delicate multi-loss tuning."],"supporting_citations":[{"why":"Supplies the MVTec AD benchmark, the primary dataset for measuring multi-class detection and localization.","marker":"[4]"},{"why":"Provides the DINO-pretrained ViT used as the frozen teacher encoder whose features both decoders distill.","marker":"[5]"},{"why":"Contributes the memory bank and sparse cosine retrieval design that CMM extends with class supervision.","marker":"[12]"},{"why":"Serves as the strongest multi-class baseline the method must outperform across the four benchmarks.","marker":"[15]"},{"why":"Supplies the vector-quantized transformer baseline and the category-label dependency that the paper states as an inherited limitation.","marker":"[24]"},{"why":"Establishes the unified multi-class feature-reconstruction paradigm and the identical-shortcuts problem MDD-Net targets.","marker":"[39]"},{"why":"Supplies the synthetic anomaly generation (Perlin noise with DTD textures) and binary masks used to train both decoders.","marker":"[41]"},{"why":"Is the direct ViT-based baseline and the per-category comparison point, especially on the Uni-Medical benchmark.","marker":"[45]"},{"why":"Provides the denoising student-teacher formulation and teacher-restoration discrepancy that TRD scoring follows.","marker":"[48]"}],"fun_headline_variants":["Twin decoders cancel blind spots, lifting pixel mAP by 11 points","One model replaces per-class detectors with twin-decoder anomaly maps","Split decoders untangle anomaly detection from normal reconstruction","Dual-decoder network solves the over-generalization vs under-reconstruction dilemma","Memory-augmented dual decoders refine anomaly maps without per-class models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the claim that two structurally identical decoders with comparable capacity fail in correlated ways on hard normal patterns; if their errors do not co-occur, subtracting their outputs creates false positives on normal textures instead of removing them.","fun_headline_variants_meta":{"raw":{"variants":["Twin decoders cancel blind spots, lifting pixel mAP by 11 points","One model replaces per-class detectors with twin-decoder anomaly maps","Split decoders untangle anomaly detection from normal reconstruction","Dual-decoder network solves the over-generalization vs under-reconstruction dilemma","Memory-augmented dual decoders refine anomaly maps without per-class models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001017,"raw_usage":{"total_tokens":4328,"prompt_tokens":1018,"completion_tokens":3310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":3231}},"tokens_in":634,"tokens_out":3310,"duration_ms":21043,"temperature":1.0,"reasoning_tokens":3231,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:37:42.035612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a validation set of only normal images, compute per-pixel cosine distance maps between the restoration and identity decoders and measure the spatial correlation of their individual reconstruction errors; if the two error maps are weakly correlated, the RID signal cannot cancel normal-region false positives. As a second check, retrain with the identity decoder using an architecturally different decoder; if performance does not drop, the correlated-deficiency assumption is not load-bearing.","supporting_citations":[],"review_version":1}