{"id":"ba36d025-a94d-4111-a48f-9d1eacf3d0c8","arxiv_id":"2608.07904","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DeCo decouples the shape of an industrial defect from its source product and recouples it with a new product, improving downstream defect detection by 5-8% pixel AP.","lead":"DeCo is a method that generates fake industrial defects on products without using any real images of defects on those products. It works by separating the defect shape from the product it came from, then pasting that shape onto a new product's normal texture.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'new state-of-the-art' claim is not yet established: the most relevant reference-based zero-shot baselines (AnomalyControl [14], AnomaMagic [19]) are cited but never compared, so the reported gains may be against a weaker subset of competitors.","rationale":"I read the paper in good faith: the decoupling/recoupling story is coherent, the ablations isolate each component, and the cross-dataset results on VisA and Real-IAD with fixed hyperparameters are genuine evidence that the mechanism transfers. The reader's weakest assumption (that Q encodes structure while K/V encode texture in SD3 self-attention) is a real mechanism risk, but it is partially mitigated by the QK-Routing ablation (Tab. 8) and by the fact that the text-branch A-LoRA can still modify K/V of the abnormal token. The most load-bearing gap is comparative, not mechanistic: the paper declares state-of-the-art among zero-shot anomaly generation methods while omitting the two reference-based methods it cites as the field's existing approaches. Because the SOTA claim is the central claim, this omission is decisive for acceptance. The MVTec test-set tuning of lambda_PI and omega (Tabs. 9–10) adds a secondary caveat, but the VisA results with fixed settings reduce that concern. I therefore recommend keeping the conditional acceptance, now explicitly conditioned on supplying the missing baselines or a justification of their absence.","tokens_in":15986,"tokens_out":10309,"duration_ms":124867,"concrete_test":"Run AnomalyControl [14] and AnomaMagic [19] on MVTec AD under the exact downstream protocol used for Tab. 2: train the same U-Net on 1,000 generated images per category, evaluate pixel AP and P-F1, and report the same metrics. If either reference-based method reaches at least 68.0 P-AP or 64.5 P-F1, the SOTA claim fails; if code is unavailable, the authors should either provide the comparison or explicitly justify why these two cited methods are excluded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DeCo's central claim is that it 'establishes a new state-of-the-art' in zero-shot industrial anomaly generation. The experiments compare against text-based (AnomalyAny, AnoStyler) and synthesis baselines, but the two reference-based zero-shot methods explicitly discussed in Sec. 2.1 — AnomalyControl [14] and AnomaMagic [19] — are absent from every quantitative table (Tabs. 1–3, Tab. 12). These are the closest competitors because, like DeCo, they consume reference anomaly images. Without them, the margin over 'existing zero-shot methods' is unverified: if either method matches or exceeds 68.0 P-AP / 64.5 P-F1 under the same downstream protocol, the headline SOTA claim is undercut. The problem is compounded by the fact that the MVTec hyperparameters (lambda_PI = 0.1, omega = 2) were selected on the test set (Tabs. 9–10), so the reported 5.1-point improvement may be optimistic. The VisA and Real-IAD fixed-hyperparameter results are encouraging but do not repair the missing comparison. This is a completeness-of-evidence issue, not an internal inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DeCo, a zero-shot industrial anomaly generation method that uses a reference product's anomaly images (with masks) to synthesize anomalies on a target product for which only normal images are available. DeCo trains a frozen Product LoRA (P-LoRA) and a trainable Anomaly LoRA (A-LoRA). Anomaly acquisition is guided by Dual-Routing Flow (DR-Flow), which splits text tokens via AP-Routing and restricts the A-LoRA to the query projection via QK-Routing, and by Product-Invariant Flow (PI-Flow), which subtracts a product-only velocity to define a residual anomaly flow. At inference the two LoRAs are injected jointly, and Product Compatibility Correction (PCC) extrapolates the velocity prediction to compensate for weak anomaly visibility. The paper reports generation-quality metrics on MVTec AD and downstream anomaly-detection results on MVTec AD, VisA, and Real-IAD, claiming a new state of the art, including a 5.1% P-AP gain over AnoStyler on MVTec AD and an 8.2% P-AP gain on VisA.","tokens_in":16298,"tokens_out":3937,"duration_ms":42058,"significance":"If the reported results are substantiated, DeCo would be a practically useful contribution: it addresses a real cold-start problem, the method is described with complete equations, each component is ablated, the code is promised to be available, and the cross-dataset generalization experiments (VisA and Real-IAD) use fixed hyperparameters without retuning. The structural rather than semantic binding of abstract anomalies is a sensible idea and is more general than the semantic-level decoupling used in prior LoRA-composition work. However, the central claims of state-of-the-art performance and of clean product-invariant decoupling are currently supported by incomplete comparative evidence and by hyperparameters selected on the same benchmark used for the headline numbers.","major_comments":[{"comment":"The two reference-based zero-shot anomaly generation methods discussed in Section 2.1, AnomalyControl [14] and AnomaMagic [19], are never included in any quantitative comparison. Since DeCo is itself a reference-based method that consumes reference anomaly images, these are among its closest competitors. Without them, the paper's claim of a 'new state-of-the-art' is not established; the reported gains might be against a weaker subset of baselines. Please add these methods under the same downstream detection protocol, or explicitly justify their exclusion with concrete reasons.","section":"Sec. 2.1 and Sec. 5.1, Tables 1–3 and 12"},{"comment":"The two main hyperparameters of DeCo, λ_PI = 0.1 and ω = 2, are selected by maximizing downstream metrics on MVTec AD, and the headline MVTec results in Tables 2 and 4 are obtained with these same values. This makes the reported 68.0% P-AP partly a fitted value on the test set. The VisA and Real-IAD results with fixed hyperparameters are a valuable independent check, but they do not certify the MVTec improvement. Please report a hyperparameter-selection procedure that does not use the test set (for example, selection on a validation split), or provide per-seed means and variances across repeated runs, or otherwise quantify the sensitivity of Table 2 to the chosen values.","section":"Sec. 5.1, Tables 9–10"},{"comment":"The definition v_PI = v_t − û_p assumes that the residual after subtracting the product-only flow is exactly the 'pure anomaly' and is product-invariant. This is the central assumption behind the PI-Flow objective in Eq. (13), but the paper provides only qualitative visual evidence (Fig. 3) that the residual excludes product-specific appearance. Since the decoupling claim is load-bearing for the method, please provide a direct test: for example, measure the similarity between v_PI and the product-specific flow û_p across different product pairs, or evaluate whether an A-LoRA trained with v_PI transfers anomalies across products without leaking source-product texture in a quantitative metric.","section":"Sec. 4.3, Eq. (12)"}],"minor_comments":[{"comment":"References [8] and [9] appear to be the same paper (Duan et al., AAAI 2023); please remove the duplicate or replace one with the intended different source.","section":"References [8] and [9]"},{"comment":"The entries '30 + k×11' and '30 + k×12' in Table 5 do not define k; please state that k is the number of anomaly categories/types used for training the A-LoRA.","section":"Table 5"},{"comment":"The sentence 'A-LoRA,stands for Anomaly LoRA' has a typo (missing space after the comma); please correct it.","section":"Sec. 4.1, paragraph after Eq. (6)"},{"comment":"The metric 'IC-L' is introduced as 'IC-LPIPS' in Section 5.1 but never defined in the table caption; please add a one-line definition so Table 1 is self-contained.","section":"Table 1 and Sec. 5.1"},{"comment":"The term 'full-rank A-LoRA' is unusual because LoRA is defined as a low-rank adaptation in Eq. (2); please clarify that full rank here means the update matrices have rank equal to the projection dimension, and state the memory and storage implications of this choice.","section":"Appendix A, 'Architecture and Implementation Details'"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the journal and the core mechanism is plausible, but the 'new state-of-the-art' claim currently rests on an incomplete baseline set and on MVTec hyperparameters that are tuned on the test benchmark. The revision should prioritize adding AnomalyControl and AnomaMagic comparisons under the same protocol, and clarifying the hyperparameter selection. If those comparisons cannot be produced, the authors should temper the SOTA claim accordingly. The duplicate reference [8]/[9] and the minor presentation issues are easily fixable in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, the core mechanism—restricting the anomaly LoRA to the Query projection while the product LoRA handles Key/Value, plus a residual flow target that subtracts the product velocity—is genuinely new and well motivated. The ablations support each design choice. Second, the \"new state-of-the-art\" claim is not yet established: the two reference-based zero-shot methods from their own related-work section (AnomalyControl and AnomaMagic) are never compared quantitatively, and the MVTec hyperparameters are tuned on the test set.\n\nThe paper does several things right. The problem is clearly stated and the method is described with full equations. The decoupling idea is a real departure from semantic-prior approaches like UnzipLoRA and QR-LoRA, which rely on pre-trained concepts; DeCo instead binds structure to an abnormal text token via routing and a residual constraint. The PI-Flow loss is a sensible way to prevent the anomaly token from absorbing source-product texture, and the Product Compatibility Correction is a simple CFG-style extrapolation that addresses a real failure mode. The cross-dataset results on VisA and Real-IAD with fixed hyperparameters are encouraging and provide some independence from the MVTec tuning.\n\nSoft spots, in order of concern. (1) Missing comparisons. AnomalyControl and AnomaMagic are described as reference-based zero-shot generators, which makes them the closest competitors. Their absence from all tables means the 5.3-point P-AP margin over AnoStyler might not survive against a stronger baseline. This is a completeness problem, not an internal inconsistency. (2) Hyperparameter selection. lambda_PI and omega are chosen on the MVTec test set, so the headline 68.0 P-AP is partly a fitted number. The VisA/Real-IAD results with the same values mitigate this, but they don't fully repair it. (3) No error bars or multiple seeds. For a claimed SOTA with a 5-point margin, this matters. (4) The Q/K/V role assumption is borrowed from natural-image editing, and the paper would be stronger if it probed whether the roles hold for abstract industrial defects. The ablations partly address this, but a dedicated analysis would help.\n\nOverall, the paper is a solid engineering contribution. The mechanism is new, the writing is clear, and the evidence is mostly reproducible—the code is available and the ablations are thorough. It deserves a serious referee. My recommendation: send it to review, but ask the authors to add the missing reference-based baselines, report variance, and soften the SOTA claim until they do.","headline":"Solid, clearly-written anomaly generation paper with a genuinely new LoRA routing mechanism, but the headline SOTA claim is under-cut by missing comparisons to the closest reference-based baselines and test-set hyperparameter tuning.","tokens_in":16800,"tokens_out":3959,"would_cite":true,"duration_ms":43580,"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":"DeCo claims that splitting anomaly structure from product texture, then recombining them, makes zero-shot synthetic industrial defects realistic enough to train downstream detectors.","keywords":["zero-shot anomaly generation","industrial anomaly inspection","anomaly synthesis","diffusion models","LoRA composition","decoupling and recoupling","anomaly detection","MVTec AD"],"falsifier":"Run the same DeCo pipeline with the routing reversed, putting the anomaly LoRA on the key and value projections and the product LoRA on the query projection, on the same MVTec reference-target pairs; if reversed routing matches or exceeds the reported 68.0 P-AP, the claimed query-controls-structure role is not doing the required work.","tokens_in":15809,"feed_emoji":"🏭","tokens_out":6351,"duration_ms":65996,"temperature":0.7,"pith_summary":"The paper claims that synthetic industrial defects look wrong because existing methods either describe the defect in words or copy it from a reference image without separating the defect's geometry from the product it came from. DeCo's answer is to split the job: first decouple the defect's structure from the source product and bind it to a dedicated anomaly token in a diffusion model, then explicitly recouple that structure with the normal texture of a target product. The paper reports that detectors trained on DeCo-generated images improve pixel-level average precision by 5.1 points on MVTec AD and 8.2 points on VisA over existing zero-shot generation, and that its anomaly localization reaches 68.0 P-AP and 64.5 P-F1 on MVTec AD. If correct, this would make zero-shot synthetic defects usable as training data for industrial inspection without collecting real images of each product's failures.","feed_headline":"Zero-shot fake defects beat prior art by 5-8% pixel AP","feed_subtitle":"Training detectors on DeCo-generated defects lifts MVTec pixel AP by 5.1% and VisA by 8.2%.","key_machinery":"Dual-Routing Flow (DR-Flow) with AP-Routing and QK-Routing. AP-Routing separates the abnormal text token from the product token so each is transformed only by its own LoRA; QK-Routing computes image queries with the anomaly LoRA and image keys and values with the frozen product LoRA, forcing the anomaly LoRA to encode structure rather than texture. Product-Invariant Flow (PI-Flow) computes a residual velocity by subtracting the product-only prediction, and Product Compatibility Correction (PCC) extrapolates the final prediction away from the normal-product prediction. Together these mechanisms decide what is learned by which adapter and where the learned defect is injected during generation.","core_discovery":"DeCo establishes text-structure binding in Stable Diffusion 3 by training two LoRA adapters with complementary roles. A frozen Product LoRA encodes the normal product into the key and value projections, while a trainable anomaly LoRA acts only on the query projection, on the premise that queries determine spatial layout while keys and values carry appearance. Dual-Routing Flow (DR-Flow) routes the abnormal text token exclusively through the anomaly LoRA and the product token through the product LoRA; Product-Invariant Flow subtracts the product-only velocity as a residual target so the anomaly token cannot absorb source-product texture. At inference the two LoRAs are merged and Product Compatibility Correction extrapolates the denoising velocity away from the product-only prediction, which the paper says compensates for mismatches when an anomaly learned on one product is transplanted to another. The paper's evidence is that downstream U-Net detectors trained on DeCo images beat all compared zero-shot and synthesis baselines on MVTec AD, VisA, and Real-IAD.","pith_inferences":["If the query/key/value role split holds for abstract industrial defects, the same decouple-recouple recipe could transfer other non-semantic visual structures, such as tool marks, wear patterns, or biological textures, without a text description of the anomaly.","The paper's full-rank anomaly LoRA result suggests that when a diffusion model lacks semantic priors for a target concept, adapter capacity, not just routing, is the binding constraint, so low-rank LoRA composition methods may systematically underperform on abstract industrial structures.","Product Compatibility Correction is an inference-time compatibility dial; a testable extension would be to make its strength per anomaly-product pair, or learned, rather than a global constant.","A stronger decoupling test would be to remove reference masks at inference and check whether the anomaly token still localizes the defect structure, since masks are currently borrowed from the reference set."],"forward_implications":["A defect structure learned from one set of products can be transplanted to a different product category with no real defect images of the target, because structure and texture are bound to separate tokens.","Downstream detectors gain most at the pixel level, where synthetic-data quality matters most: the paper reports +5.1 P-AP on MVTec AD and +8.2 P-AP on VisA.","The same separation makes generated anomalies more controllable: changing the product LoRA changes the surface while the anomaly token keeps its geometry.","Because the method needs only normal images of the target product, cold-start inspection for new product lines becomes feasible without waiting for real defects to be collected."],"supporting_citations":[{"why":"Supplies the Stable Diffusion 3 backbone and MM-DiT attention blocks on which both LoRAs are trained and merged.","marker":"[10]"},{"why":"Defines low-rank adaptation, the parameter-efficient mechanism used to build the product and anomaly adapters.","marker":"[16]"},{"why":"Provides the query/key/value role separation the paper relies on to justify restricting the anomaly LoRA to the query projection.","marker":"[3,4,34]"},{"why":"Supplies blended diffusion, used at inference to inject the generated anomaly only inside the masked region.","marker":"[1]"},{"why":"Provides MVTec AD, the primary benchmark for generated-anomaly realism and downstream detection.","marker":"[2]"},{"why":"Supplies the few-shot anomaly-driven generation protocol that DeCo follows for mask combination and downstream evaluation.","marker":"[13]"},{"why":"Supplies the evaluation protocol of training a U-Net segmentation model on generated image-mask pairs and serves as a comparison baseline.","marker":"[17]"},{"why":"AnoStyler is the main zero-shot baseline that DeCo outperforms by 5.3 P-AP and 3.8 P-F1 on MVTec AD.","marker":"[32]"},{"why":"AnomalyAny is a training-free zero-shot generation baseline compared on generation quality and downstream detection.","marker":"[33]"},{"why":"Provides VisA, the cross-domain dataset used to show that acquired anomaly structures transfer to unseen product categories.","marker":"[41]"}],"fun_headline_variants":["Synthetic defects, real gains: DeCo lifts AP by 5–8%","Zero-shot anomaly generation: DeCo sets new SOTA with +5–8% AP","Decouple and recouple: DeCo's trick for zero-shot defect synthesis","Training detectors on DeCo-generated defects: +5.1% MVTec, +8.2% VisA","DeCo: Fake defects that train better detectors, up to 8% AP gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that in the image-generation model's attention mechanism, one projection sets where things go while the other two set how things look, and that this split survives for abstract industrial defects; if that assumption is wrong for defects, the anomaly adapter can silently copy the source product's texture.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic defects, real gains: DeCo lifts AP by 5–8%","Zero-shot anomaly generation: DeCo sets new SOTA with +5–8% AP","Decouple and recouple: DeCo's trick for zero-shot defect synthesis","Training detectors on DeCo-generated defects: +5.1% MVTec, +8.2% VisA","DeCo: Fake defects that train better detectors, up to 8% AP gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1983,"prompt_tokens":990,"completion_tokens":993,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":876}},"tokens_in":606,"tokens_out":993,"duration_ms":9595,"temperature":1.0,"reasoning_tokens":876,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:41:35.909225+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same DeCo pipeline with the routing reversed, putting the anomaly LoRA on the key and value projections and the product LoRA on the query projection, on the same MVTec reference-target pairs; if reversed routing matches or exceeds the reported 68.0 P-AP, the claimed query-controls-structure role is not doing the required work.","supporting_citations":[{"cited_title":"In: Forty-first International Conference on Machine Learning (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the Stable Diffusion 3 backbone and MM-DiT attention blocks on which both LoRAs are trained and merged."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies blended diffusion, used at inference to inject the generated anomaly only inside the masked region."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Provides MVTec AD, the primary benchmark for generated-anomaly realism and downstream detection."},{"cited_title":"In: European conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Supplies the few-shot anomaly-driven generation protocol that DeCo follows for mask combination and downstream evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the evaluation protocol of training a U-Net segmentation model on generated image-mask pairs and serves as a comparison baseline."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence (2026)","cited_arxiv_id":null,"evidence_quote":"AnoStyler is the main zero-shot baseline that DeCo outperforms by 5.3 P-AP and 3.8 P-F1 on MVTec AD."},{"cited_title":"In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference","cited_arxiv_id":null,"evidence_quote":"AnomalyAny is a training-free zero-shot generation baseline compared on generation quality and downstream detection."},{"cited_title":"a <product> with an <anomaly>","cited_arxiv_id":null,"evidence_quote":"Provides VisA, the cross-domain dataset used to show that acquired anomaly structures transfer to unseen product categories."}],"review_version":1}