{"id":"cadca364-c90f-4587-87c6-b04a4d495307","arxiv_id":"2607.21928","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A surface-centric two-stage diffusion pipeline generates B-reps by diffusing surface adjacency and then geometry, and reports state-of-the-art validity and efficiency on DeepCAD and ABC.","lead":"This paper introduces TG-Diff, a two-stage diffusion model that generates CAD shapes by first creating a surface adjacency graph and then producing the surface geometry, with edges and vertices recovered by intersecting surfaces. Because it decouples topology from geometry, the method reports higher validity and lower computational cost than prior B-rep generation models on DeepCAD and ABC.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validity claim is substantially attributable to the up-to-three-retry angle-adjustment post-process; Table 9 shows validity drops from 83.6% to 62.4% when retries are capped at one.","rationale":"The reader and I converge on the same load-bearing point. The central metric of excellence, validity, is defined at the end of a repair pipeline, and the only quantification of that pipeline's contribution (Table 9) shows a 21.2-point swing between one and three retries. This is not an internal inconsistency; the system as described may well achieve these numbers. But it makes the SOTA claim unverifiable from the paper alone: without code or a retry-stratified breakdown, one cannot tell whether the margin over DTGBrepGen comes from the learned representation or from a more aggressive post-processor. I propose a single test that would resolve this: cap retries at one and re-measure, or stratify the reported metrics by retry count. I do not see grounds to strengthen the verdict beyond CONDITIONAL; the concern is about evidential support, not demonstrated failure.","tokens_in":20014,"tokens_out":7695,"duration_ms":70053,"concrete_test":"Run the official pipeline on DeepCAD-7-30 and DeepCAD-0-30 with the post-processing retry cap set to 1 instead of 3, keeping all random seeds and inference settings fixed; record validity, COV, MMD, JSD and, for every sample, the number of retries and the largest angle perturbation used. If validity drops by more than about 10 points (as Table 9's 83.6 to 62.4 pattern suggests) or if the quality metrics shift materially when retry-repaired samples are separated out, the headline validity advantage is attributable to the angle-adjustment repair loop rather than to the learned diffusion models.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central validity claim is not a property of the two learned diffusions alone. Section 3.4 describes a repair loop that, when surface intersections fail (near-tangent cases, Fig. 7), adjusts the angular parameters of circular surfaces and retries up to three times. Table 9 shows that on DeepCAD-7-30 this choice is worth 21.2 percentage points: validity is 62.4% with one retry, 76.2% with two, and 83.6% with three (86.9% with ten). Thus roughly a quarter of the reported 'valid' outputs are valid only after geometry-modifying repair, not as a direct consequence of the generated topology and surface latents. Section 5 concedes that near-tangent surfaces, circular rings, and spheres can defeat the extension/intersection kernel, and Fig. 8(b,d) show irrecoverable failures. The paper reports no breakdown of how often retries were actually invoked or how much the angular adjustments perturbed the generated geometry, and no quality metrics (COV/MMD/JSD) are stratified by retry count. Since the headline claim is 'excels in the validity metric' and Table 2 compares against baselines whose post-processing does not include this angle-adjustment retry loop, the margin over DTGBrepGen (85.6 vs 75.6 on DeepCAD; 61.4 vs 52.9 on ABC) may substantially reflect a more aggressive repair procedure rather than a better generative model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TG-Diff, a two-stage diffusion framework for unconditional B-rep generation. Topology is represented only as a surface-adjacency matrix and generated by a D3PM discrete diffusion; surface geometry is encoded into latents by a topology-aware VAE and generated by a lightweight topology-conditioned DiT latent diffusion. Edges and vertices are not generated explicitly but are recovered by post-processing that extends surfaces, computes intersections, detects minimum cycles, and trims faces, with up to three retries that adjust angular parameters of circular surfaces. The paper reports state-of-the-art COV, MMD, JSD, and validity on DeepCAD-0-30 and ABC-0-50, along with lower parameter and GFLOP counts than baselines.","tokens_in":20341,"tokens_out":8900,"duration_ms":75155,"significance":"If the claims hold, the paper offers a clean surface-centric alternative to edge/vertex-level topology generation, with a compelling efficiency story and a natural mechanism for reducing geometric-topological inconsistency. The ablations are unusually complete: they test GT topology, VAE-GNN removal, BLISS, tokenization, single- vs two-stage generation, and the retry schedule, and Section 5 candidly describes failure modes. The central validity claim, however, is not yet established, because Table 9 shows that 21.2 points of validity on DeepCAD-7-30 are purchased by the second and third post-processing retries, and the paper does not decompose how much of the Table 2 margin over DTGBrepGen is due to the learned model versus the repair loop. Since no code is released, the missing breakdown cannot be filled in by reviewers. With the additional analysis requested below, the contribution would be publishable; as it stands, the headline validity result is conditional on an unquantified repair procedure.","major_comments":[{"comment":"The headline validity claim is not a property of the two learned diffusions alone. On DeepCAD-7-30, Table 9 shows validity increases from 62.4% with one retry to 76.2% with two and 83.6% with three (86.9% with ten). The paper does not report how often the retry loop is triggered, how many surfaces per shape are angle-adjusted, or how much these adjustments perturb the generated geometry, and no COV/MMD/JSD numbers are stratified by retry count. Section 5 concedes that near-tangent surfaces, circular rings, and spheres can defeat the intersection kernel and that some failures are irrecoverable; Section 4.2 also reports a 3.1% residual failure rate even after repair. The abstract's claim of excelling 'in the validity metric' therefore conflates generative quality with repair capability. Please report validity as a function of retry budget for the main benchmarks, the distribution of perturbed surfaces, and before/after quality metrics.","section":"Section 3.4 and Table 9"},{"comment":"The validity comparison against baselines uses unmatched post-processing. Table 6 states that BrepGen and DTGBrepGen use three thresholding attempts and 200-step edge-refinement, while TG-Diff uses three retries that adjust the angular parameters of failed surfaces. These are not equivalent repair strategies: TG-Diff's retries modify the analytic geometry of circular surfaces, whereas the baselines' threshold attempts do not change generated surface parameters. The margin in Table 2 (85.6 vs 75.6 on DeepCAD-0-30; 61.4 vs 52.9 on ABC-0-50) may therefore partly measure the aggressiveness of the post-processor. Please provide a matched comparison, such as re-running TG-Diff with one retry and with no angle adjustment, and re-running baselines with the same retry budget, or at least quantify the geometric deviation induced by the angle adjustments.","section":"Section 4.2 and Table 6"},{"comment":"There is an unexplained internal inconsistency in the reported validity numbers. Table 2 reports 85.6% validity for Ours on DeepCAD-0-30, while Table 5 reports 95.6% for 'all' on the same dataset, with a caption that mentions 'raw data only after post-processing'. The manuscript does not define what input 'raw data' refers to. If Table 5 measures how often the post-processor converts uncut raw surfaces into valid B-reps, it is a different quantity and must be labeled as such; if it measures the full generation pipeline, the 10-point gap with Table 2 must be explained. This is load-bearing because the paper's main claimed improvement over prior work is the validity metric.","section":"Tables 2 and 5"}],"minor_comments":[{"comment":"In the paragraph defining the reconstruction loss, 'the reconstructio loss is defined' contains a typo; it should be 'reconstruction loss'.","section":"Section 3.2"},{"comment":"The name of the baseline is written both as 'HOLA' and 'HoLa'; please standardize the spelling.","section":"Table 6 and surrounding text"},{"comment":"The factors '1.76× fewer parameters, 1.59× fewer GFLOPs, and 1.85× reduction in total runtime' are not referenced to a specific baseline; from Table 7, 1.76× appears to be relative to DTGBrepGen in parameters, while 1.59× is relative to BrepDiff in GFLOPs. Please state the reference method for each factor.","section":"Section 4.2, runtime paragraph"},{"comment":"No code or checkpoints are provided. For a systems paper whose central claim depends on a custom post-processor (Section 3.4), a reproducibility statement or release would substantially increase confidence in the reported numbers.","section":"General"},{"comment":"The VLM-based user-study validation uses 20 participants and 110 shapes; please report how the shapes were sampled and the variance of the consistency rates, and consider adding error bars or confidence intervals to the metrics in Table 2, since all experiments are said to be run 10 times.","section":"Section 4.1 and Table 2"},{"comment":"The no-post-processing metrics are computed against a different reference (fitted B-spline patches) than the main results (GT B-rep surfaces), as noted in the text; the table caption should repeat this caveat to avoid misreading.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"To the editor: This is a serious empirical systems paper with thorough ablations and honest discussion of limitations, but the central validity claim is entangled with an unquantified repair procedure. The issues are addressable with additional analysis rather than a conceptual error, so I recommend major revision rather than rejection. The authors should provide a validity decomposition by retry count, quantify the geometric perturbation from angle adjustments, address the Table 2 versus Table 5 discrepancy, and ideally release code so the retry-dependent numbers can be verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read. First, the paper is worth reading: the surface-only representation plus discrete topology diffusion is a genuinely clean design, and the ablation work is unusually careful. Second, the headline validity numbers are not a pure measure of the generative model. The valid numbers in Table 2 come from a pipeline that retries the post-process up to three times by nudging the angles of circular surfaces. Table 9 shows the cost of that choice: 62.4% valid with one retry versus 83.6% with three on DeepCAD-7-30. That is 21 points of the reported validity, and the main comparison tables do not report the retry breakdown. The stress-test note is right about this, and it is the biggest caveat to carry forward.\n\nWhat is actually new: the authors strip B-rep generation down to a surface set plus an adjacency matrix, do the topology with D3PM over a ternary adjacency relation, then feed that topology into a latent diffusion for surfaces via FiLM-modulated attention. That combination does not appear in BrepGen, DTGBrepGen, or BrepDiff. It is an engineering contribution more than a conceptual breakthrough, but a real one, and the efficiency numbers (82M params, 2.2 GFLOPs) make it practical.\n\nWhere the paper is solid: the ablations are informative. Removing BLISS drops validity from 83.6 to 69.8; switching from row-wise tokens to element-wise tokens drops to 72.4; removing the VAE's GNN drops to 79.2. These moves make sense and give me confidence the authors tested their own claims. The without-post-processing metrics in Table 3 also show Ours ahead of BrepDiff and DTGBrepGen on COV*, MMD*, and JSD*, so the model itself is not just riding on the repair step. That is the right kind of evidence to report, even though the reference set there differs from the main table.\n\nSoft spots beyond the retry issue: no code or checkpoints, so nothing is independently reproducible yet. No error bars despite the 10-runs claim in the caption. The VLM-based metrics rely on a 20-person user study, which is fine as a sanity check but not as ground truth. And the manual 500-sample failure analysis, which finds 3.1% of their valid outputs are actually invalid, tells you the validity metric is already lenient.\n\nBottom line: this deserves a serious referee. The core idea is sensible, the experiments are above average for the area, and the retry concern is disclosed internally if you look at Table 9. What I would want before trusting the state-of-the-art claim is code, raw validity before repair, and error bars. With those, I would be fairly convinced.","headline":"Solid two-stage diffusion system for B-rep generation; the topology-first surface-only idea is clean and the raw metrics hold up, but the headline validity gap is inflated by a repair loop and the code isn't out.","tokens_in":20879,"tokens_out":2996,"would_cite":true,"duration_ms":29735,"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 surface-only, topology-first diffusion pipeline beats larger B-rep generators on DeepCAD and ABC validity, quality, and diversity.","keywords":["boundary representation","B-rep generation","diffusion models","discrete diffusion","surface-centric representation","CAD","surface trimming","topology generation"],"falsifier":"Feed ground-truth B-rep surfaces and adjacency from DeepCAD-0-30 through the Section 3.4 post-processor unchanged: if validity is far below 100% on inputs whose geometry and topology are already correct, the intersection-and-trim step itself is the bottleneck and the reported validity numbers measure the repair pipeline rather than generative quality.","tokens_in":19824,"feed_emoji":"📐","tokens_out":9751,"duration_ms":75970,"temperature":0.7,"pith_summary":"The paper argues that generating a CAD boundary representation (B-rep) becomes easier and more accurate if the model never tries to produce vertices, edges, and surfaces together, and instead generates only surfaces plus their adjacency. It claims that a two-stage diffusion framework—a discrete diffusion model for the surface adjacency graph, followed by a latent diffusion model for surface geometry conditioned on that graph—delivers the best coverage, quality, distribution matching, and validity on the DeepCAD and ABC benchmarks among compared diffusion pipelines. The result matters because it suggests the full vertex–edge–surface hierarchy that previous generators model explicitly is redundant for generation, and that deciding topology before geometry stabilizes the harder surface-synthesis step. The reported pipeline runs at 82.18M parameters and 2.2 GFLOPs, far below the diffusion baselines it compares against.","feed_headline":"Surface-first B-rep diffusion sets validity records at 82M params","feed_subtitle":"Discrete topology diffusion plus topology-guided surface diffusion beats larger baselines on DeepCAD and ABC.","key_machinery":"The load-bearing object is the surface-centric B-rep representation $B=(S,T)$, where $S$ stores each face as a $16\\times 16$ masked UV grid with coordinates and a validity mask, and $T$ is an $N\\times N$ categorical adjacency matrix with states for adjacent, not adjacent, and padding. The machinery that carries the argument is the two-stage diffusion pairing: a D3PM discrete diffusion over $T$ with an absorbing MASK state and row-wise tokenization for the topology, and a latent diffusion over VAE-encoded surface latents whose FiLM-modulated self-attention is conditioned on $T$. The final watertight solid is assembled by extending surfaces, computing pairwise intersections, extracting minimum cycles as face loops, and trimming, so vertices and edges are never generated as independent random variables.","core_discovery":"On its own terms, the central claim is that B-rep generation can be decomposed into a surface-only topology $T\\in\\{0,1,2\\}^{N\\times N}$ and a set of masked UV-grid surfaces $S\\in\\mathbb{R}^{N\\times 16\\times 16\\times 4}$, with edges and vertices recovered afterward by intersecting and trimming the generated surfaces. The topology is sampled by a D3PM discrete diffusion whose row-wise tokenization and symmetry constraint produce valid, diverse adjacency graphs; the surface latents are then denoised by a DiT-based network in which the predicted adjacency modulates attention through FiLM scales and biases. Under this design, every edge in the final solid lies on its parent surfaces by construction, which the paper identifies as the reason it avoids the cracks, misalignments, and orphan entities that plague joint generation. The paper reports validity of 85.6% on DeepCAD-0-30 and 61.4% on ABC-0-50, together with the best COV, MMD, and JSD scores, and attributes the gain to the surface-centric representation and topology conditioning rather than to any single architectural trick.","pith_inferences":["A testable extension: because topology enters only as an attention modulation, the same two-stage design should support conditional generation (text, sketch, or image) by conditioning the surface diffusion on an externally supplied hint; the paper only demonstrates unconditional generation.","A testable prediction: since edges are always constructed from surface intersections, any measured generation failure should be traceable either to a wrong adjacency/geometry or to a post-processor defeat (tangency, rings, spheres); classifying failures this way on the paper's 500-sample manual study would cleanly separate generative error from repair error.","The retry-sensitivity of validity suggests a practical evaluation norm: report validity at one retry as the raw pipeline number, and treat the gap to three or ten retries as the contribution of the repair engine, not of the generative model."],"forward_implications":["On the DeepCAD-0-30 and ABC-0-50 benchmarks the paper claims the best COV, MMD, JSD, and validity among compared diffusion baselines, with validity 85.6% and 61.4% respectively (Table 2).","Generated topologies are valid and diverse: the D3PM route reaches 99.8% and 96.5% valid-topo on the two benchmarks and roughly doubles DTGBrepGen's uniqueness on DeepCAD (Table 4).","Reported validity is pipeline-level, not model-level: on DeepCAD-7-30 the same generator yields 62.4% validity at one post-processing retry, 83.6% at three, and 86.9% at ten (Table 9).","The surface-only representation cuts compute substantially: 82.18M parameters, 2.2 GFLOPs, and lower intermediate memory than all compared baselines (Table 7).","Ablations show the representation and conditioning choices carry the gains: removing BLISS canonical labeling drops validity from 83.6% to 69.8%, and switching row-wise tokens to per-element tokens drops it to 72.4% on DeepCAD-7-30 (Table 8)."],"supporting_citations":[{"why":"Supplies the D3PM discrete diffusion formalism used to sample surface adjacency.","marker":"[Austin et al. 2021]"},{"why":"Introduces the absorbing MASK state in the discrete transition matrix used by the topology diffusion.","marker":"[Gu et al. 2022]"},{"why":"Provides the DiT backbone that the surface denoising network extends with topology modulation.","marker":"[Peebles and Xie 2023]"},{"why":"Provides the velocity-prediction objective used to train the surface latent diffusion.","marker":"[Salimans and Ho 2022]"},{"why":"Establishes the masked UV-grid surface representation, the DeepCAD/ABC preprocessing splits, and the main baseline.","marker":"[Xu et al. 2024a]"},{"why":"Is the topology-first autoregressive baseline whose redundant vertex-edge-surface topology motivates the surface-only representation.","marker":"[Li et al. 2025b]"},{"why":"Is the single-stage surface diffusion baseline compared on post-processing and validity.","marker":"[Lee et al. 2025]"},{"why":"Supplies the minimum-cycle basis algorithm used to extract closed face loops from intersection curves.","marker":"[Kavitha et al. 2007]"},{"why":"Provides the DeepCAD dataset and the autoregressive baseline of the same name.","marker":"[Wu et al. 2021]"},{"why":"Provides the ABC dataset used for the second benchmark.","marker":"[Koch et al. 2019]"}],"fun_headline_variants":["Lightweight diffusion splits topology from geometry for B-rep success","Two-stage B-rep diffusion: surface adjacency guides geometry","82M-param diffusion wins B-rep generation with surface-centric design","Discrete topology plus surface diffusion beats larger B-rep models","TG-Diff: surface-only topology and guided geometry for B-rep"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's validity rests on the post-processing premise that extending the generated surfaces, intersecting them, and extracting minimum cycles will reconstruct a correct watertight solid whenever the topology and geometry are good; the paper concedes that near-tangent surfaces, circular rings, and spheres can defeat this step, and Table 9 shows the retry budget largely determines the reported validity rate.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight diffusion splits topology from geometry for B-rep success","Two-stage B-rep diffusion: surface adjacency guides geometry","82M-param diffusion wins B-rep generation with surface-centric design","Discrete topology plus surface diffusion beats larger B-rep models","TG-Diff: surface-only topology and guided geometry for B-rep"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000583,"raw_usage":{"total_tokens":2807,"prompt_tokens":1070,"completion_tokens":1737,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":1648}},"tokens_in":686,"tokens_out":1737,"duration_ms":10777,"temperature":1.0,"reasoning_tokens":1648,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:29:32.828137+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed ground-truth B-rep surfaces and adjacency from DeepCAD-0-30 through the Section 3.4 post-processor unchanged: if validity is far below 100% on inputs whose geometry and topology are already correct, the intersection-and-trim step itself is the bottleneck and the reported validity numbers measure the repair pipeline rather than generative quality.","supporting_citations":[],"review_version":2}