{"id":"a7655f95-816c-4821-9c14-c82ca93ffc0d","arxiv_id":"2505.05901","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MC4AD predicts per-point internal and external corrective force vectors from point clouds and uses their magnitude as an anomaly score, reporting state-of-the-art detection and segmentation on five benchmarks plus a new synthetic dataset.","lead":"MC4AD is a 3D anomaly detection method that frames surface defects as the result of invisible internal and external forces, then trains a network to predict the corrective force that would restore the damaged surface. If the reported results hold, it offers a fast and lightweight alternative for industrial inspection, though the mechanical metaphor is not physically validated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DA-Gen pseudo-anomalies are the entire training signal, yet the closed form in Eq. 5 predicts a zero force at the anomaly center, and no test checks whether the learned 6-channel force score transfers to defects with other geometries, amplitudes, or materials.","rationale":"The reader correctly identifies the DA-Gen distribution as the weakest assumption and the paper can be accepted only conditionally on that assumption being adequate. I agree with that verdict and with the general orientation of the concern. However, I would sharpen the attack in two ways that the reader only gestures at. First, I think there is an explicit mathematical inconsistency at the center of the generated anomaly: Eq. 5 contains a factor pi/max(pi) that vanishes at the center (pi=0), so the pseudo-defect force has zero magnitude exactly at the defect center unless the paper defines pi with an offset or applies the force at a boundary. This is not merely an out-of-distribution concern; it is an internal inconsistency between the generative process and the claim that the model learns to give high anomaly scores at the center of local defects. Second, I would make the concrete test a cross-generation transfer evaluation that directly measures whether the learned score is a calibrated anomaly score for out-of-family defects, rather than a more general robustness check. The paper's own limitation statement (Conclusion, Appendix H) concedes the lack of explicit physical constraints, which supports the conditional recommendation. I do not think the concern invalidates the empirical finding that 6-channel offset prediction with symmetry loss is competitive on several benchmarks; it does mean the mechanical interpretation and the generality of the anomaly scores are not established by the current experiments. The reader's weakest_assumption is the same failure mode I emphasize, so I mark agreement as partial: aligned on the key vulnerability but with a distinct, more specific center-vanishing formulation and a different concrete test.","tokens_in":32891,"tokens_out":2114,"duration_ms":19587,"concrete_test":"Run a controlled cross-generation experiment. (1) Verify the center behavior of Eq. 5 by plotting the displacement magnitude as a function of distance to the patch center: if the magnitude is zero at the center, redesign the check so that the anomaly peak is at the center. (2) Train MC4AD with the published DA-Gen on one dataset (e.g., Anomaly-ShapeNet or Real3D-AD), then evaluate on a held-out test set of real or synthesized defects that are outside the DA-Gen family: defects with amplitude 0.2, defects with non-normal displacement directions (lambda < 0.5 or pure tangential), global scaling/shrinkage defects, and defects consisting of removed points. If O-AUROC drops by more than 10 points relative to the in-family defects, the DA-Gen distribution is the main driver of the reported scores and the mechanical force interpretation is not validated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that CFP-Net learns corrective forces whose norm is a valid anomaly score, and the only supervision that makes a normal point produce zero force and an anomalous point produce nonzero force is the DA-Gen pseudo-anomaly distribution in Eq. 5. This makes DA-Gen the load-bearing component: if real defects are not in the support of that generative distribution, the single trained score is not a calibrated anomaly score. Two concrete weaknesses follow. First, there is a likely inconsistency at the anomaly center. In Eq. 5, the displacement is FD,i = [beta*lambda*nu + (1-lambda)*eta] * gamma * pi/max(pi) * (1 - sigma*|pi|), where pi is the normalized projected distance to the center. The factor pi/max(pi) tends to 0 at the center (pi goes to 0), and the attenuation factor (1 - sigma*|pi|) tends to 1, so the modeled defect force goes to zero at the very point where the defect is strongest. The paper does not state that pi is offset away from the center or that the force is applied at the boundary rather than the center. If the center displacement is zero, the generated anomalies have a ring-like deformation with an unmoved core, and the trained model may assign low scores to the center of real defects, directly contradicting the maximum-score-as-sample-score rule. Second, the ablation in Table 3 and Appendix E only varies the DA-Gen parameters inside the same family (G, lambda, sigma, gamma). It never tests whether the scoring rule transfers to out-of-family defects, such as global non-local deformations, material/color-only defects, large cracks spanning multiple patches, or defects whose amplitude is outside [0.06, 0.12]. The claim of nine state-of-the-art results therefore rests entirely on an untested and partially unmatched generative distribution. The paper itself flags this in the Conclusion and Appendix H: the framework 'lacks explicit physical constraints to derive damage forces.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MC4AD, a 3D anomaly detection framework that reinterprets defects as the result of internal and external 'damage forces' and trains a network to predict opposing 'corrective forces' whose magnitude serves as the anomaly score. The method contains three main components: DA-Gen, which generates pseudo-anomalies by displacing points along perturbed surface normals; CFP-Net, a Minkowski U-Net with complementary skip connections that outputs a six-channel corrective force field; and a combined loss with symmetry, distance, and directional terms. The paper also introduces an HQC pruning strategy and a new Anomaly-IntraVariance dataset with intraclass variance. Experiments are reported on Anomaly-ShapeNet, Anomaly-ShapeNet-New, MvTec3D-AD, Real3D-AD, and the new dataset, claiming nine state-of-the-art results with low memory and high inference speed.","tokens_in":33312,"tokens_out":5534,"duration_ms":58819,"significance":"If the mechanical framing were valid, the work would offer a conceptually novel perspective on 3D anomaly detection and a practical efficiency-oriented framework. The paper includes public code, extensive comparisons across many categories, a resource analysis, a generalizable hierarchical quality-control module, and a new dataset, which are useful contributions to the community. However, the central theoretical claim is currently not supported: the 'damage force' is defined as the displacement produced by DA-Gen, the losses as written drive the network to reproduce that same displacement rather than its opposite, and the anomaly score is therefore a fitted reconstruction of the authors' own synthetic augmentation. The empirical gains are also close to noise on some benchmarks, and no error bars or out-of-family transfer tests are provided. The dataset and efficiency experiments are solid enough to be of value, but the paper's central claim needs substantial revision.","major_comments":[{"comment":"There is an internal inconsistency at the center of the generated anomaly. In Eq. (5), FD,i is proportional to (pi_i/max(pi_i)) * (1 - sigma_i * |pi_i|), where pi_i is the normalized projected distance to the anomaly center. At the center pi_i = 0, so the first factor is zero and the generated displacement is zero, while the text states that gamma specifies the maximum displacement magnitude at the anomaly center. This means pseudo-anomalies are ring-like with an unmoved core. Since DA-Gen is the sole training signal, a network trained on these samples may assign low scores to the center of real defects, directly undermining the use of the maximum point score as the sample-level anomaly score. The authors should either define pi with an offset so that the center receives the maximum displacement, or explicitly analyze and justify the ring-like deformation.","section":"Section 4.1, Eq. (5)"},{"comment":"The loss terms in Eq. (10) contradict the stated mechanics. The text and Eq. (4) require FC = -FD, i.e., the corrective force opposes the damage force. However, Ldist minimizes ||FD,i - FC,i||, which drives FC toward +FD, and Ldir contains the negative normalized dot product of FD and FC, which when minimized also aligns FC with +FD. Similarly, Lsym contains the negative dot product of F'I and F'E, which aligns the two corrective components rather than making them opposing, contrary to the claim in Section 4.3 that the symmetry loss enforces opposing directions. As written, the combined loss trains the network to reconstruct the DA-Gen displacement field. The authors must correct the signs in Eq. (10) or substantially reframe the method as pure pseudo-anomaly reconstruction without the mechanical corrective-force narrative.","section":"Section 4.3, Eq. (10)"},{"comment":"The core 'prediction' is circular with respect to the anomaly score. The damage force FD,i in Eq. (5) is defined as the displacement applied by DA-Gen, and the network is trained with a loss that minimizes the difference between FD,i and the predicted corrective force FC,i. At inference, the norm of FC is used as the anomaly score. Thus the score measures how well the network reconstructs the authors' synthetic augmentation field; no independent physical constraint enters. The ablations in Table 9 only vary parameters inside the DA-Gen family (G, lambda, sigma, gamma). To support the claim of generalization to real defects, the paper needs experiments where the training distribution and the test defect distribution are explicitly different in kind (e.g., non-normal deformations, material-level changes, larger displacement magnitudes, or real scanned defects), and an analysis of whether real defect displacements lie in the support of Eq. (5).","section":"Sections 4.1 and 4.3"},{"comment":"The experimental evidence for state-of-the-art claims is weakened by missing variance estimates and close margins. On MvTec3D-AD the O-AUROC is 0.954 versus 0.951 for the second-best method, and on Real3D-AD the P-AUROC is 0.837 versus 0.836; with no error bars or multiple seeds it is unclear whether these differences are significant. In addition, the text in Section 5.2 states that MC4AD outperforms 'M3DM of 94.5% and 90.6%' on MvTec3D-AD, but Tables 16 and 17 report M3DM average values of 0.874 O-AUROC and 0.906 P-AUROC, so the cited 94.5% is inconsistent with the tables. The same paragraph says the comparison spans 12 categories while Table 16 lists 10 categories. These inconsistencies and missing statistical evidence should be resolved before the 'nine state-of-the-art performances' claim can be accepted.","section":"Section 5.2 and Tables 16-18"},{"comment":"The theoretical support is overstated. Proposition A1 shows that the network mapping is differentiable, which is a generic property of convolutional networks with smooth activations and does not establish that the norm of a predicted corrective force is a valid anomaly score. Lemma A2 is cited from a preprint titled 'Reconstruct Anything Model' and is not a formal theorem about U-Nets solving arbitrary inverse problems in anomaly detection. The abstract states that MC4AD 'has been proven effective through theory and experimentation,' but the provided theory neither proves effectiveness nor connects the mechanical force model to detection guarantees. The authors should either provide a genuine theoretical statement linking the force-field objective to anomaly separability, or remove the 'proven by theory' claim.","section":"Appendix B.3, Proposition A1 and abstract"}],"minor_comments":[{"comment":"The heading contains a typo: 'Introductuon' should be 'Introduction'.","section":"Section 1 heading"},{"comment":"The model name is inconsistently written as 'CTF-Net' in the text and 'CFP-Net' in the figure and elsewhere; please unify.","section":"Section 4, first paragraph"},{"comment":"The text says MvTec3D-AD has 12 categories, but Table 16 lists 10 categories; please correct the count in the text.","section":"Section 5.2"},{"comment":"The dataset name is written as 'Mvtec-AD' in one place; use the consistent form 'MvTec3D-AD'.","section":"Appendix G"},{"comment":"The heading 'Limitations and Feature Work' should read 'Limitations and Future Work'.","section":"Appendix H"},{"comment":"The final sentence of Section 4.3 ('with the reconstruction effect') is grammatically incomplete and should be rephrased.","section":"Section 4.3, last sentence"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds on a strong empirical setup but the central contribution is currently a relabeling of pseudo-anomaly offset prediction with a mechanical narrative. The sign inconsistency in Eq. (10) and the center-zero issue in Eq. (5) need to be fixed or the narrative substantially revised. The novelty relative to PO3AD, which also predicts point offsets from pseudo-anomalies, should be clarified explicitly. Given the close margins and missing error bars, a stronger statistical analysis is needed before the state-of-the-art claims can be defended."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick read. MC4AD is a competent, incremental successor to PO3AD: instead of predicting a single 3D offset, it predicts internal and external corrective force vectors (6 channels) and adds a symmetry loss between them. It also introduces a slightly richer pseudo-anomaly generator (DA-Gen) than PO3AD's Norm-AS, a pruned fast model with thresholding (HQC), and a new 16-category dataset with intraclass variance. The empirical work is extensive—five benchmarks, 90 classes, code released—and the efficiency gains (14M params, ~15 FPS on Real3D-AD) are real. The generalization experiments showing MC-Skip helps PointNet/DGCNN etc. are a nice extra. This deserves a serious referee.\n\nThe soft spots are mostly about framing and one possible bug. The 'mechanical perspective' is a metaphor—Newton's law is invoked but the integrals in Section 3 don't define a physically consistent force model, and the authors admit in the limitation that they lack explicit physical constraints. That's fine as motivation, but it shouldn't be presented as theory.\n\nMore concretely, the stress-test note is on target about Eq. 5. DA-Gen writes FD,i = [...] * gamma * pi/max(pi) * (1 - sigma*|pi|). If pi is the normalized distance from the patch center, then the displacement is zero at the center, contradicting the text that gamma is the max displacement at the anomaly center. That means generated anomalies are ring-shaped with an unmoved core, and the model may learn to underweight the center of real defects. I don't see any offset or boundary condition in the text that fixes this. This needs a careful correction or an explanation of what pi actually is.\n\nOther issues: no error bars anywhere, and several SOTA margins are thin (0.954 vs 0.951 on MvTec3D, 0.837 vs 0.836 P-AUROC on Real3D). The MvTec3D comparison is restricted to only-3D baselines, which is defensible but should be stated more clearly. And the text says 'M3DM of 94.5%' while the table shows 0.874—sloppy.\n\nBottom line: the central claim is plausible but rests entirely on DA-Gen's ability to cover real defects. The center-zero issue is a load-bearing crack, not a cosmetic one. I'd send it to review with a request to fix or explain Eq. 5, add variance estimates, and correct the internal inconsistencies. If the equation is clarified, this is an acceptable incremental paper with a useful dataset.","headline":"A solid, incremental 3D anomaly detection paper with real efficiency gains and a useful dataset, but the pseudo-anomaly generator has a likely center-zero bug and the 'mechanical' theory is motivating metaphor rather than physics.","tokens_in":33952,"tokens_out":3328,"would_cite":true,"duration_ms":32688,"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":"This paper argues that every 3D surface defect can be modeled as an imbalance of internal and external mechanical forces, and that a network trained to predict the corrective force at each point yields anomaly scores that beat…","keywords":["3D anomaly detection","point cloud","mechanical forces","corrective force prediction","pseudo-anomaly generation","industrial quality control","Anomaly-IntraVariance dataset","MvTec3D-AD"],"falsifier":"Take a defect class absent from DA-Gen's family, such as scratches that displace points tangentially rather than normally, or surface cracks visible only in texture while depth stays flat, and measure per-point AUROC on that class alone. If the corrective-force norm separates such points no better than chance while matching the paper's numbers on bulge and dent classes, the mechanical premise is falsified.","tokens_in":32662,"feed_emoji":"⚙️","tokens_out":8413,"duration_ms":81401,"temperature":0.7,"pith_summary":"Most 3D anomaly detectors describe what a defect looks like, comparing test geometry against memorized normal geometry. This paper claims that the right question is why the defect exists: every anomaly is the trace of an unpredictable defective force, either internal (for instance expanding gas) or external (for instance a collision), acting on the product surface. It proposes MC4AD, which synthesizes pseudo-anomalies along surface normals, trains a lightweight U-Net to predict the corrective force vector that would push each point back to its normal position, and scores each point by the magnitude of that predicted force. The authors report that this mechanical inversion obtains nine state-of-the-art detection and segmentation results across five established benchmarks plus their new Anomaly-IntraVariance dataset, using 14M parameters and the fastest inference among compared methods. If the claims hold, the source of a defect, not just its shape, is enough to find it.","feed_headline":"Every 3D defect is a force imbalance; predicting the fix detects it","feed_subtitle":"A 14M-parameter net predicts corrective forces per point; force magnitude flags bulges, dents, and scratches on five benchmarks.","key_machinery":"The load-bearing object is a pair of theoretical definitions: a defect force $F_D = \\int_{S^+} f_E \\, dS^+ + \\int_{S^-} f_I \\, dS^-$ that deforms a normal manifold $M$ into $M'$, and its opposite corrective force $F_C = -F_D$ that restores it. The whole pipeline is built to predict $F_C$: DA-Gen synthesizes the pseudo-defect deformation, CFP-Net (a differentiable MinkUNet-style U-Net with MC-Skip complementary feature mixing) emits the six-channel per-point corrective force, and the combined loss $L_{\\text{comb}}=L_{\\text{dist}}+L_{\\text{dir}}+L_{\\text{sym}}$ constrains it to match the synthesized deformation and to keep the internal and external components anti-symmetric. The Hierarchical Quality Control strategy adds a pruned 60%-smaller network that pre-screens obviously normal samples before the full network runs.","core_discovery":"The paper's central discovery is that anomaly detection can be reframed as force-equilibrium restoration. In the Mechanics Complementary Model, a normal point is in equilibrium: both defect force $F_D$ and corrective force $F_C$ are zero. A defective point is subject to a nonzero resultant force $F_D = F_E + F_I$ (external and internal components), which deforms the surface by a nonlinear map $\\phi(F, M) = \\nabla M$; the corrective force is exactly its opposite, $F_C = -F_D$. Training therefore does not rely on normal features alone: DA-Gen deforms training patches by perturbing surface normals with controlled magnitude $\\gamma \\in [0.06, 0.12]$, normal-dominance $\\lambda \\in [0.95, 1]$, and stretching $\\sigma \\in [0, 0.08]$, producing realistic pseudo-defects; CFP-Net, a differentiable sparse U-Net with complementary MC-Skip connections that mix internal and external feature streams, predicts a six-channel corrective force (3D external plus 3D internal) per point. The combined loss enforces that the predicted force reconstructs the known pseudo-defect ($L_{\\text{dist}}$, $L_{\\text{dir}}$) and that internal and external corrective components are anti-symmetric in anomalous regions ($L_{\\text{sym}}$), which the ablation singles out as the largest single contributor. At inference the norm $\\|F_C\\|$ is the point anomaly score and its maximum is the object score. Reported numbers include O-AUROC 90.9% on Anomaly-ShapeNet, 95.4% on MvTec3D-AD, 78.6% on Real3D-AD, 88.8% on Anomaly-ShapeNet-New, and 76.1%/62.5% on the two groups of the new Anomaly-IntraVariance dataset.","pith_inferences":["The mechanical vocabulary is likely an interpretative scaffold: the same arithmetic, predicting a per-point residual offset and scoring by its norm, is at heart what offset-based detectors do; the distinct contribution here is the internal/external decomposition and the anti-symmetry constraint. A direct control that trains a two-channel residual model without the force framing would separate the ","A targeted experiment the paper does not run: apply the trained CFP-Net to a physical simulation, such as finite-element stress on a scanned part, and compare predicted corrective-force directions with principal strain directions. Agreement would evidence the mechanical interpretation; disagreement would mean 'force' is only a naming convention.","Because DA-Gen only displaces points near the surface normal with magnitude in [0.06, 0.12], defects without geometric displacement, such as color or texture changes, subsurface voids, or material-phase changes, are outside the training distribution; a dedicated test set of such non-geometric defects would reveal whether the premise extends beyond morphology.","The paper itself concedes in Section 6 and Appendix H that it lacks explicit physical constraints to derive damage forces, so the force vocabulary is learned implicitly from synthesized deformations and the mechanical interpretation rests on how well the pseudo-anomaly distribution matches real defects."],"forward_implications":["The corrective-force magnitude can replace expensive memory-bank or reconstruction comparisons entirely, since inference is a single forward pass over the point cloud.","Splitting the score into external and internal channels is not decorative: Figure 3 shows that fusing both channels lifts O-AUROC from about 0.855/0.843 to 0.912, which suggests the complementary force channels carry partly independent signal.","The symmetry loss is a real lever: removing $L_{\\text{sym}}$ hurts O-AUROC by 13.4% (from 0.786 to 0.652), so the mechanical constraint is doing genuine work.","A pruned 60%-smaller model can serve as a first-pass filter in HQC, speeding up inference (up to 26.4 FPS on Anomaly-ShapeNet) with essentially no accuracy loss, which fits cost-limited assembly-line inspection.","Benchmarking on Anomaly-IntraVariance shows that intraclass variance is hard: Group 2 (four subspecies per class) drops all methods to 0.625 or below, framing a realistic target for future work."],"supporting_citations":[{"why":"The offset-prediction baseline (PO3AD) whose pseudo-anomaly generator MC4AD replaces in the V4 ablation, and whose 34M-parameter model MC4AD is compared against for speed and size.","marker":"[12]"},{"why":"Real3D-AD, the few-shot real-scan benchmark used for the main detection and segmentation tables, ablations, and HQC speed experiments.","marker":"[5]"},{"why":"Anomaly-ShapeNet and its IMRNet reconstruction baseline; provides the synthesis protocol that DA-Gen extends and the 40-category evaluation.","marker":"[2]"},{"why":"MvTec3D-AD dataset whose depth modality is used for the only-3D comparison.","marker":"[43]"},{"why":"MinkUNet, the differentiable sparse-convolution backbone that provides voxelization and the convolutional primitives of CFP-Net.","marker":"[41]"},{"why":"PatchCore, the memory-bank detector used as a comparison baseline and as a target for the HQC generalization.","marker":"[22]"},{"why":"M3DM multimodal anomaly detection baseline on Anomaly-ShapeNet and MvTec3D, and in the HQC speed comparisons.","marker":"[31]"},{"why":"Universal approximation theorem cited to justify CFP-Net's capability to fit the corrective-force inverse mapping.","marker":"[58]"},{"why":"U-Net universal inverse-generative claim cited alongside [58] in the theoretical support for CFP-Net.","marker":"[59]"},{"why":"Prior 3D anomaly synthesis (Perlin-noise based) that DA-Gen is positioned against as more realistic pseudo-defect generation.","marker":"[40]"}],"fun_headline_variants":["Defect force? Predict the fix, find the flaw","Force imbalance: predict the fix, find the 3D flaw","Predict corrective forces, catch 3D defects at SOTA speed","MC4AD: force equilibrium restores 3D defect detection","3D defects as force imbalance; predict the fix to spot them"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every real defect is assumed to be a small local displacement of surface points roughly along the surface normal, with magnitude between 0.06 and 0.12 and a specific spatial attenuation profile, so that for every anomaly there is a corrective force whose magnitude is a reliable score; defects outside this deformation family are outside the model's training signal.","fun_headline_variants_meta":{"raw":{"variants":["Defect force? Predict the fix, find the flaw","Force imbalance: predict the fix, find the 3D flaw","Predict corrective forces, catch 3D defects at SOTA speed","MC4AD: force equilibrium restores 3D defect detection","3D defects as force imbalance; predict the fix to spot them"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002007,"raw_usage":{"total_tokens":7984,"prompt_tokens":1253,"completion_tokens":6731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":869,"completion_tokens_details":{"reasoning_tokens":6642}},"tokens_in":869,"tokens_out":6731,"duration_ms":47389,"temperature":1.0,"reasoning_tokens":6642,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:53:08.152427+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a defect class absent from DA-Gen's family, such as scratches that displace points tangentially rather than normally, or surface cracks visible only in texture while depth stays flat, and measure per-point AUROC on that class alone. If the corrective-force norm separates such points no better than chance while matching the paper's numbers on bulge and dent classes, the mechanical premise is falsified.","supporting_citations":[{"cited_title":"Real3d-ad: A dataset of point cloud anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Real3D-AD, the few-shot real-scan benchmark used for the main detection and segmentation tables, ablations, and HQC speed experiments."},{"cited_title":"Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network,","cited_arxiv_id":null,"evidence_quote":"Anomaly-ShapeNet and its IMRNet reconstruction baseline; provides the synthesis protocol that DA-Gen extends and the 40-category evaluation."},{"cited_title":"4d spatio-temporal convnets: Minkowski convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"MinkUNet, the differentiable sparse-convolution backbone that provides voxelization and the convolutional primitives of CFP-Net."},{"cited_title":"Towards total recall in industrial anomaly detection,","cited_arxiv_id":null,"evidence_quote":"PatchCore, the memory-bank detector used as a comparison baseline and as a target for the HQC generalization."},{"cited_title":"Multimodal industrial anomaly detection via hybrid fusion,","cited_arxiv_id":null,"evidence_quote":"M3DM multimodal anomaly detection baseline on Anomaly-ShapeNet and MvTec3D, and in the HQC speed comparisons."},{"cited_title":"3D-PNAS: 3D Industrial Surface Anomaly Synthesis with Perlin Noise","cited_arxiv_id":"2504.12856","evidence_quote":"Prior 3D anomaly synthesis (Perlin-noise based) that DA-Gen is positioned against as more realistic pseudo-defect generation."}],"review_version":1}