{"id":"5048d621-9aba-494b-9c98-c751ba11d8f3","arxiv_id":"2504.14257","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HoLa stores a full B-Rep in surface-only latent vectors, recovers curves from surface pairs via a neural intersection module, and generates valid CAD models from multiple conditions with 82 percent validity.","lead":"This paper introduces HoLa, a representation that encodes an entire CAD solid model into latent codes attached only to surfaces, and then reconstructs the curves and connections with a learned intersection network. The authors train one diffusion model to generate CAD models from text, images, point clouds, or sketches, reporting a validity rate of about 82 percent versus about 50 percent for prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central topology-recovery premise rests on the intersection classifier, whose positive-class precision/recall is unreported and whose loss equations (Eqs. 6-7) are dimensionally inconsistent; this is the load-bearing soft spot.","rationale":"The reader's conditional verdict is well-founded, and I do not see a reason to move it. The strongest empirical anchors—82.68% validity on DeepCAD, point-conditioned topology F-scores, ablations, and the honest failure analysis—are real evidence for the holistic-latent idea, and machine-checked verification is not available in this area. My stress-test focuses on the one link that is necessary for the central claim but not directly evidenced: the neural intersection module is the only mechanism that recovers curves and adjacency from surface latents at inference, and the paper gives only a single 99.99% accuracy figure plus a dimensionally inconsistent loss specification. Under natural class imbalance, near-perfect accuracy is achievable by predicting 'no intersection' for every pair, so the reported number cannot be interpreted without positive-class precision/recall. The coplanar-face and non-manifold representability issue raised by the reader is real, and it does limit the generality of the representation to closed solids, but on the evaluated DeepCAD/ABC benchmark it is secondary; the classifier question tests whether the core mechanism works at all in the evaluated domain. The proposed check is concrete and does not require new baselines or code release: class-balanced precision/recall plus an oracle-label validity ablation would settle whether the intersection module is actually carrying the claimed load. Since the empirical results are strong but this verification gap remains, the verdict stays CONDITIONAL rather than moving to ACCEPT or REJECT.","tokens_in":25448,"tokens_out":10046,"duration_ms":103517,"concrete_test":"Compute precision and recall of the trained intersection classifier on a held-out DeepCAD test set, separately under class-balanced sampling and under the natural imbalanced distribution over all ordered surface pairs; report the positive-class recall in both settings. Then replace the classifier's predicted intersection labels with ground-truth surface-pair intersection labels during decoding while keeping all learned curve reconstruction, and remeasure validity and topology F-scores. If positive-class recall is low, or if oracle labels substantially improve validity and topology metrics, the topology-recovery premise is not supported by the current evidence. If positive-class recall is high and oracle labels leave validity and topology scores essentially unchanged, the concern is resolved and the central claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a full B-Rep can be recovered from surface latents alone because the neural intersection module can identify intersecting surface pairs and reconstruct the shared curve geometry from the pair of latents. This module is therefore the sole mechanism by which curves and adjacency are recovered at inference, and its performance is load-bearing for the 82.68% validity result. The only direct evidence offered is the statement in Sec. 5.5 that intersection accuracy 'remains consistently high (99.99%)'. That number is uninterpretable without positive-class precision and recall: intersecting pairs are rare among all surface pairs, so a classifier that always predicts 'no intersection' can score near 100% accuracy. The paper does not say whether the 99.99% figure was computed on balanced classes or on the natural imbalanced distribution, nor does it report recall on the positive class. The ambiguity is compounded by Eqs. 6-7, which define L_inter using T_SC(i,j) as the label for a surface pair (i,j), even though T_SC is defined in Sec. 3 as the m x n surface-to-curve adjacency matrix. The index ranges and the dimensions therefore do not match a surface-to-surface classifier, so the actual training objective of the intersection module is not specified in the paper. If the classifier misses a substantial fraction of true intersections, the holistic latent does not encode topology in the evaluated setting, and the validity gain must be attributed to something other than the claimed mechanism. If the classifier has high positive recall, the central premise is directly supported. This is the single most load-bearing concern because every other reported result depends on this module working, and no other experiment isolates its behavior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HoLa, a holistic latent representation for B-Rep CAD models. The representation encodes a full B-Rep (surfaces, curves, vertices, and topological relations) into a latent space defined only over surface primitives. The central design is a neural intersection module that, given a pair of surface latents, predicts whether the surfaces intersect and recovers the geometry of the shared curve; curves, vertices, and topology are then decoded from these surface-pair predictions. On top of this VAE, the authors train a single latent diffusion model for unconditional generation and for conditional generation from point clouds, images, sketches, and text. On the DeepCAD dataset the method reports an unconditional validity of 82.68% versus 47.74% for BRepGen, and improved topology/geometry F-scores in point-conditioned generation. The paper also presents ablations, robustness experiments with imperfect point clouds, and qualitative results for text-, image-, and sketch-conditioned generation.","tokens_in":25616,"tokens_out":5912,"duration_ms":57903,"significance":"If the central claim holds, this is a substantial advance for B-Rep generation: replacing multi-step, primitive-by-primitive generation with a single unified latent space would reduce training complexity and improve generated model validity. The paper's evaluation has notable strengths: validity is checked externally via OpenCascade, the point-conditioned comparison uses disclosed protocol details, the single-run version of the method is reported alongside the multi-run version, and the ablations support the importance of the half-edge structure and spatial resolution in the latent. The claimed validity gain (82.68% vs 47.74%) is large and is supported by an external checker rather than only self-reported metrics. However, the load-bearing mechanism of the paper—the neural intersection module—is not directly evaluated in a way that establishes its precision/recall, and the training objective for this module is not written consistently in the paper. These gaps are fixable and do not invalidate the approach, but they must be addressed before the central claims can be accepted.","major_comments":[{"comment":"The intersection-classifier loss is not well-formed as printed. T_SC is defined in Sec. 3 as the m-by-n surface-to-curve adjacency matrix, but in Eqs. (6)-(7) it is used as a label for a surface pair (i,j), and the double sum runs over n curves rather than over surface pairs. In addition, I_bc is defined as a classifier on R^16, yet it is called with two surface latent vectors (z_s^(i), z_s^(j)). As written, the objective is dimensionally inconsistent and the actual training target of the intersection classifier is unspecified. Please correct the loss by introducing an explicit surface-pair adjacency matrix (or deriving it from T_SC) and by defining the classifier input as the fused feature produced by the cross-attention module.","section":"Sec. 4.2, Eqs. (6)-(7)"},{"comment":"The only direct evidence about the intersection classifier is the statement that intersection accuracy 'remains consistently high (99.99%)'. This number is uninterpretable without reporting precision and recall on the positive (intersecting) class, because intersecting pairs are rare among all surface pairs and a trivial always-negative classifier can achieve almost 100% accuracy on the natural distribution. The paper also does not state whether the 99.99% figure is computed on the balanced training distribution or on the natural inference distribution. Since the decoder's curve and topology recovery depends entirely on this classifier, please report the confusion matrix (or precision/recall) on both balanced and natural pair distributions, and, ideally, the effect of the classifier threshold on final topology F-scores and validity.","section":"Sec. 5.5"},{"comment":"The central prior that 'any curve in a B-Rep must be the intersection between two surface primitives' is not valid for all B-Reps it is claimed to cover. Adjacent coplanar faces (smooth edges), tangent-surface contacts, open shells, and non-manifold configurations do not determine a unique shared curve from the two supporting surfaces alone. The paper does not restrict its claim to closed solids with transverse face intersections, nor does it quantify how often the test-set curves violate the premise. Please either narrow the representational claim to the closed-solid, non-coplanar, non-tangent setting, or provide a dataset-scale analysis: report the fraction of DeepCAD curves for which the two adjacent faces have distinct, non-tangent supporting surfaces, and report topology FE/EV recall separately on the curves that satisfy the premise versus those that do not.","section":"Sec. 1, Sec. 3"}],"minor_comments":[{"comment":"The caption reports '84% vs. ≈50%' for unconditional validity, while the abstract and Table 1 report 82.68% vs 47.74%; please reconcile these numbers.","section":"Fig. 1 caption"},{"comment":"The Ours32 variant samples 32 models per input point cloud and selects by Chamfer distance to the input, which is a best-of-N test-time augmentation not available to the deterministic baselines. This is disclosed in the text, but the table would benefit from a footnote stating that the comparison is not single-sample-equivalent.","section":"Sec. 5.2, Table 2"},{"comment":"The text-conditioned results are entirely qualitative; given that Text2CAD provides a baseline and a quantitative protocol, reporting at least one quantitative metric (e.g., the baseline's own evaluation measure) would strengthen the claim that the method produces 'more plausible models' from text.","section":"Sec. 5.3"},{"comment":"There appears to be a typographical error in the first term of Eq. (6), where 'I_bc(S_i(i), z(j)_s)' mixes a surface feature with a surface latent; this should be part of the correction to the well-formed objective described in the major comment, not a separate issue.","section":"Sec. 4.2, Eq. (6)"},{"comment":"The paper does not state whether code or trained models will be released. Given the heavy training cost (about 1M iterations for the VAE and 1M for the LDM) and the importance of the exact intersection-module training protocol, a reproducibility statement would be valuable.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The dimensional inconsistency in Eqs. (6)-(7) is likely a presentation error, but the missing precision/recall analysis for the intersection classifier is a substantive gap in the evidence for the paper's central claim. If the authors cannot supply that analysis, I would not regard the 82.68% validity result as explained by the stated mechanism. The representational-prior concern about coplanar/tangent faces should also be addressed, though it may be resolvable with a restricted claim or a dataset analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real step forward for B-Rep generation: encoding an entire B-Rep into a latent space defined only over surface primitives, then recovering curves and topology with a learned surface-pair intersection module, is genuinely new and it works. The validity jump from ~48% to 82.68% on DeepCAD, checked externally by OpenCascade, is the kind of result that makes people in the field sit up. Second, the paper is mostly honest and the protocol is disclosed, but the load-bearing intersection classifier is under-reported, and one of its loss equations is misprinted.\n\nWhat's actually new: prior work (BRepGen, SolidGen, ComplexGen, NVDNet) either uses separate latents for surfaces/curves/vertices, autoregressive steps, or explicit curve/vertex encoding. HoLa's bet is that topology can be reformulated as geometry: two surfaces in a closed solid intersect in a curve, so if you can predict which surface pairs intersect and reconstruct the shared curve from their latents, you don't need a separate curve latent. The ablations support this — removing self-attention, spatial resolution, or half-edge orientation degrades validity and reconstruction, which is exactly what you'd expect if the latent is really carrying the directional info needed to recover edges. The point-cloud-conditioned results are also solid, with quantitative comparisons against HPNet/SEDNet+Point2CAD and NVDNet, and the robustness tests on noisy scans are a nice extra.\n\nSoft spots, in proportion. (1) Eqs. 6-7 define the intersection loss using T_SC, the surface-to-curve adjacency, while the classifier operates on surface pairs. As printed, the indices don't line up. This is likely a notation error, but it means the actual training objective of the intersection module is not specified in the paper. A referee should ask for the corrected equation and, more importantly, positive-class precision/recall. The paper notes balanced sampling, so 99.99% accuracy isn't vacuous, but recall on true intersections is the number that verifies the central premise. (2) Text- and image-conditioned results are qualitative only, with no comparison to Text2CAD or Img2CAD even though baselines exist. That's a smaller gap, but it matters for the claim of being the first to do these conditions. (3) The evaluation distribution is filtered to 7-30 surfaces; the paper reports BRepGen's post-filter validity but not its own pre-filter number, so the headline comparison is slightly asymmetric. The filter is disclosed, though, so this is a minor issue. (4) No code or data.\n\nThese are all addressable. The central mechanism — recovering edges from surface latents in closed solids — holds up in the evaluated setting. The coplanar-adjacent-face limitation is real but not relevant to the DeepCAD/ABC solids used here.\n\nWho it's for: anyone working on CAD representation learning or generative modeling of B-Reps. It deserves a serious referee; I'd send it out and expect it to come back with minor-to-major revisions focusing on Eqs. 6-7 and the intersection-classifier metrics. I would cite it.","headline":"A genuinely new B-Rep representation that mostly delivers; the validity gains are real and externally checked, but the intersection classifier's loss equations and missing positive-class metrics need fixing before the core mechanism is fully convincing.","tokens_in":26355,"tokens_out":3558,"would_cite":true,"duration_ms":31230,"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":"This paper introduces HoLa, a latent representation defined only over surface primitives that encodes an entire B-Rep model, and shows that a single diffusion model built on it generates valid CAD models at 82.68% on DeepCAD, nearly…","keywords":["boundary representation","B-Rep generation","holistic latent space","neural intersection","latent diffusion model","CAD","topology learning","conditional 3D generation"],"falsifier":"Run the trained intersection classifier on B-Reps with coplanar adjacent faces or open sheet bodies and check whether each ground-truth shared curve is recovered uniquely and whether the binary intersection predictions on positive pairs have high precision and recall; a low recall on intersecting surface pairs would show that topology is not fully recoverable from surface latents alone. A second, cleaner check is to measure the Euler characteristic of the wireframe graph of generated models: for watertight genus-g solids it must satisfy V - E + F = 2 - 2g, so systematic violations would refute the claim that topology is encoded consistently.","tokens_in":25083,"feed_emoji":"📐","tokens_out":9429,"duration_ms":75731,"temperature":0.7,"pith_summary":"HoLa claims that an entire B-Rep model—surface and curve geometry, vertices, and the discrete adjacency topology—can be stored in a single latent space defined only over the surface primitives. The paper's key move is to treat topology as geometry: because a B-Rep curve is necessarily the intersection of two surfaces, a learned neural intersection module can recover each curve and the surface-to-curve adjacency from pairs of surface latent vectors, so no separate curve, vertex, or topology latents are needed. On this latent space the authors train one diffusion model that generates B-Reps unconditionally or from point clouds, single/multi-view images, sketches, and text prompts. If the claim holds, it removes the multi-step pipelines of prior B-Rep generators, cuts redundancies and incoherences, and raises the validity of generated models on DeepCAD from roughly 50% (DeepCAD and BRepGen) to 82.68%. The same representation also improves point-conditioned reconstruction over fitting-based baselines and remains robust to noisy, cropped, or sparse point clouds.","feed_headline":"A surface-only latent space yields 82% valid CAD models","feed_subtitle":"Tying each curve to two surfaces lets one diffusion model handle text, images, point clouds, and sketches","key_machinery":"The holistic latent (HoLa) space is a per-surface VAE latent, each surface latent having a small spatial resolution (2x2) and feature dimension 8 so that orientation information survives pooling. The load-bearing component is the neural intersection module, a cross-attention network applied to ordered pairs of surface latents that outputs (i) a binary intersection prediction and (ii) a curve feature for the shared intersection; a positional encoding of surface order makes the recovered half-curve orientation consistent, since swapping the surface pair reverses the curve direction. The decoder reconstructs surfaces from the surface latents and curves from the recovered curve features, and these are sewn into watertight B-Reps via a B-spline fitting and wire-loop post-process. A single latent diffusion model then generates or denoises the full surface-latent set conditioned on a 256-dimensional vector extracted from noise, images, point clouds, sketches, or text.","core_discovery":"The central claim is that the topological connection between two surfaces is intrinsically tied to the geometry of their intersecting curve, and therefore a full B-Rep can be encoded and decoded using surface latents alone. The paper builds a variational autoencoder whose encoder fuses surface and curve features through a graph neural network and self-attention into per-surface latent vectors; the decoder then applies a neural intersection module to each pair of surface latents, predicting whether they intersect and producing the feature of the shared half-curve from the pair. The loss combines surface and curve reconstruction, a binary intersection classification, and KL regularization. The authors report that this holistic latent supports a single latent diffusion model that, on the DeepCAD benchmark, produces 82.68% valid watertight B-Reps versus 47.74% for BRepGen and 50.82% for DeepCAD, with improved coverage and distributional metrics; on ABC the valid ratio is 60.46% versus 32.68% for BRepGen. The failures they document are inconsistent surface primitives that lead to inaccurate trimming and non-watertight bodies, plus noise from padding variable-length surface sets to a fixed maximum.","pith_inferences":["If the intersection-recoverability premise generalizes, the same pairwise-intersection inductive bias could be applied to assembly or joint representations, where each connection is also a lower-order geometric feature shared by exactly two components.","A direct stress test would evaluate the intersection classifier separately on coplanar adjacent faces and on open shells; if curves there are not unique, the representation needs an explicit manifoldness guard rather than just the closed-solid cases in DeepCAD.","Since validity is measured through OpenCascade's checker, part of the reported gain could be sensitive to the checker's tolerance; an independent topological audit of generated B-Reps (for example, the Euler characteristic of the wireframe graph) would pin down how much of the validity gap is topology consistency versus geometric fitting.","The per-surface latent still carries a fixed maximum primitive count through padding; compressing the whole B-Rep to a single global latent, as the paper suggests as future work, would remove that noise source and may further improve validity."],"forward_implications":["One diffusion model replaces the separate surface, curve, and vertex generators of prior pipelines; changing the input modality only retrains the diffusion model, not the VAE.","On the filtered DeepCAD benchmark, unconditional validity rises to 82.68% (versus 47.74% for BRepGen and 50.82% for DeepCAD), while coverage increases to 78.87% and MMD and JSD improve.","Point-conditioned generation reaches 98.23% validity when 32 test-time samples are selected by Chamfer distance, and it outperforms fitting-based pipelines (HPNet+Point2CAD, SEDNet+Point2CAD) on topology F-scores.","The same latent handles text, single-view, multi-view, and sketch conditioning, with multi-view inputs giving the best Chamfer distance among image conditions.","Ablations show that the spatial resolution of the surface latent and the half-curve orientation are necessary: removing them drops validity from 94.18% to 79.62% while degrading precision and recall."],"supporting_citations":[{"why":"BRepGen is the main multi-step baseline with separate surface and curve generators; it provides the 47.74% validity number that HoLa is compared against.","marker":"[Xu et al. 2024a]"},{"why":"SolidGen is the prior autoregressive B-Rep synthesis model that generates vertices, curves, and surfaces sequentially, defining the multi-step paradigm HoLa replaces.","marker":"[Jayaraman et al. 2023]"},{"why":"DeepCAD provides the dataset and the sketch-and-extrude baseline (50.82% validity) used for unconditional and point-conditioned evaluation.","marker":"[Wu et al. 2021]"},{"why":"ABC is the second unconditional generation benchmark, supplying the 32.68% BRepGen baseline for validity comparison.","marker":"[Koch et al. 2019]"},{"why":"OpenCascade supplies the validity checker and the watertight conversion used to define the headline validity numbers.","marker":"[OpenCascade 2021]"},{"why":"NVDNet is the point-conditioned baseline with a Voronoi intermediate representation, used to compare geometry and topology metrics on point-cloud inputs.","marker":"[Liu et al. 2024a]"},{"why":"Text2CAD provides the text-conditioning dataset and baseline used to evaluate text-to-B-Rep generation.","marker":"[Khan et al. 2024]"},{"why":"DDPM is the diffusion formulation that underlies the latent diffusion model trained on the holistic latent.","marker":"[Ho et al. 2020]"}],"fun_headline_variants":["Surfaces alone encode full B-Reps: 82% valid CAD","Holistic latent boosts B-Rep validity to 82%","Curve geometry from surface pairs: valid CAD at 82%","One diffusion model, many inputs: holistic surface latent","Topology via geometry: B-Rep generation from surface latents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every curve in a solid model of the kind used here must be recoverable from exactly two supporting surface primitives; this holds for closed solids whose neighboring faces are not coplanar, but it fails for adjacent faces lying in the same plane, for open shells, for thin sheet shapes, and for edges where more than two faces meet, since two surfaces then do not define a unique shared curve.","fun_headline_variants_meta":{"raw":{"variants":["Surfaces alone encode full B-Reps: 82% valid CAD","Holistic latent boosts B-Rep validity to 82%","Curve geometry from surface pairs: valid CAD at 82%","One diffusion model, many inputs: holistic surface latent","Topology via geometry: B-Rep generation from surface latents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000837,"raw_usage":{"total_tokens":3709,"prompt_tokens":1060,"completion_tokens":2649,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":2560}},"tokens_in":676,"tokens_out":2649,"duration_ms":16946,"temperature":1.0,"reasoning_tokens":2560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:54:33.337277+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained intersection classifier on B-Reps with coplanar adjacent faces or open sheet bodies and check whether each ground-truth shared curve is recovered uniquely and whether the binary intersection predictions on positive pairs have high precision and recall; a low recall on intersecting surface pairs would show that topology is not fully recoverable from surface latents alone. A second, cleaner check is to measure the Euler characteristic of the wireframe graph of generated models: for watertight genus-g solids it must satisfy V - E + F = 2 - 2g, so systematic violations would refute the claim that topology is encoded consistently.","supporting_citations":[],"review_version":1}