{"id":"337ad032-89b7-4159-973a-9adbbbc9a5c2","arxiv_id":"2512.18953","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Training 3D point-cloud generative models on half-objects and reflecting the generated halves guarantees near-perfect symmetry, at some cost in standard fidelity metrics.","lead":"This paper measures how often 3D generative models preserve mirror symmetry, finds a consistent gap, and fixes it by training on half-objects and mirroring the generated halves. A generalist might care because it shows a simple data-centric trick can enforce a structural prior without changing the model architecture.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline symmetry gains are a construction artifact: full shapes are assembled as H ∪ reflect(H), so mirror-CD is ≈0 by design and cannot measure what the model learned.","rationale":"The reader correctly identified the circularity of the symmetry metric as a primary issue, but framed the weakest assumption as the x=0 plane alignment. My concern is more load-bearing: the headline symmetry improvement is not merely partly circular—it is fully determined by the reconstruction protocol. Any half-object generator, including a random sampler, yields mirror-CD ≈0 after reflection. Thus the central claim that training on half-objects improves symmetry is vacuous as measured. However, the paper still contributes a symmetry auditing protocol and reports independent FPD improvements, so a conditional verdict with a mandatory revision (add a null control and reframe the claim) is appropriate. The reader's CONDITIONAL verdict therefore remains unchanged by my analysis. agreement_with_reader is partial because the reader's stated weakest assumption (plane deviation) differs from my identified concern (metric circularity), although the reader's rationale already lists circularity as a key issue.","tokens_in":12975,"tokens_out":5842,"duration_ms":62123,"concrete_test":"Run a control experiment: train no model; instead, for each evaluation sample, draw a half-object uniformly at random from the training half-object dataset, apply the same reflection/normalization/FPS pipeline as §5.3, and compute mirror-CD, FPD, and 1-NNA against the validation set. If the control achieves the same mirror-CD (~0.004–0.008) as S-PVD/S-LION while FPD/1-NNA are poor, the symmetry improvement is a construction artifact and the paper must be reframed around the non-circular metrics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3 defines the reconstructed full shape as Xtilde_full = Xtilde_+ ∪ Xtilde_−, where Xtilde_− is obtained by reflecting Xtilde_+ across x=0. For any generated half Xtilde_+, the reflected full shape is Xtilde_full itself: reflect(Xtilde_full) = reflect(Xtilde_+) ∪ reflect(reflect(Xtilde_+)) = Xtilde_− ∪ Xtilde_+ = Xtilde_full (exact equality up to FPS tie-breaking and the rescaling step). Therefore the Section 4.1 symmetry metric—CD between a point cloud and its reflected counterpart—is identically ≈0 for every reconstructed sample, regardless of whether the diffusion model has learned any symmetry. The reported drops (0.016→0.004, 0.117→0.005, etc.) are a restatement of the reconstruction rule, not evidence of learned symmetry. A null model that outputs random half-objects would produce identical mirror-CD distributions. Consequently, the paper's central quantitative claim is unsupported; only the FPD results (Table 2) are independent of this artifact, but FPD is not a symmetry metric and the 1-NNA/EMD degradation remains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper audits reflection symmetry in three ShapeNet classes for four point-cloud/3D generative models (PVD, LION, XCube, SLIDE 3D) using a Chamfer-distance-based symmetry score, and proposes a data-centric intervention: train PVD and LION on half-objects (one side mirrored onto the other) and reconstruct full shapes by reflecting the generated half across the x=0 plane. The authors report large reductions in mirror-CD (e.g., airplane PVD 0.016 to 0.004, LION 0.117 to 0.005), improvements in FPD in all reported cells, and a trade-off in 1-NNA/CD/EMD, which they acknowledge. The central claim is that training on half-objects teaches the model a symmetry prior and yields more symmetric, plausible shapes.","tokens_in":13216,"tokens_out":5582,"duration_ms":61896,"significance":"The proposed symmetrization is simple, architecture-agnostic, and the FPD gains in Table 2 are notable if reproducible. The paper is also transparent about the CD/EMD/1-NNA degradation. However, the main quantitative evidence for symmetry improvement is confounded: for the symmetrized models, the symmetry metric is essentially satisfied by construction, not by learned behavior. The FPD results are independent of that artifact, but FPD is not a symmetry metric. The manuscript therefore currently does not support its strongest claim, although the underlying idea and the non-circular parts of the evaluation could be salvaged with a substantial rewrite.","major_comments":[{"comment":"The headline symmetry gains are a construction artifact. Section 5.3 defines Xtilde_full = Xtilde_+ ∪ Xtilde_−, where Xtilde_− is the reflection of Xtilde_+. Section 4.1 measures symmetry as CD(S1, reflect(S1)). For any generated half Xtilde_+, reflect(Xtilde_full) = reflect(Xtilde_+) ∪ Xtilde_+ = Xtilde_full, so CD(Xtilde_full, reflect(Xtilde_full)) is identically zero (up to FPS tie-breaking and rescaling). Consequently the reported reductions in Fig. 5 and Sec. 6.4 are restatements of the reconstruction rule and cannot measure what the model learned; a null model generating random halves would yield the same mirror-CD. Please re-evaluate symmetry of the generated halves before reflection, compare against a random-half baseline, and either remove or reframe the mirror-CD comparisons for symmetrized models.","section":"§5.3 and §4.1"},{"comment":"The abstract states that the paper analyzes 'symmetry dynamics during training' and that 'mechanism-inspired diagnostic tests were conducted at the sampling and latent-representation levels.' No such experiments or results appear in the manuscript. These promised analyses are exactly what would be needed to support the claim that the model internalizes a symmetry prior, rather than that the reconstruction rule imposes it. Please either include the missing experiments or revise the abstract and the related claims.","section":"Abstract vs. full text"},{"comment":"All quantitative comparisons are reported as point estimates with no error bars, confidence intervals, or significance tests over generated samples or training seeds. For the baseline-vs-symmetrized comparisons, some differences are large, but the FPD improvements are the only non-circular evidence in favor of the method, and without uncertainty estimates it is unclear whether the differences are stable. Please report variance over multiple generated sets (e.g., bootstrap or multiple sampling runs) and, ideally, over training runs.","section":"§6.4, Tables 1–2"},{"comment":"The method assumes every object is reflectionally symmetric about x=0, but Sec. 4.2 reports non-zero mirror-CD means for ShapeNet (0.011 for airplane, 0.021 for car, 0.019 for chair), so the assumption is only approximate. For objects with a different symmetry plane, the half-object construction pairs two halves that are not true mirrors, and the reconstruction forcibly symmetrizes an asymmetric object. The Discussion acknowledges this as a limitation, but the current evaluation does not quantify how much of the 1-NNA/EMD degradation is due to this imposed symmetry, as opposed to the generative model itself. Please report results on the subset of objects whose measured symmetry error is low, and discuss the effect of the x=0 assumption more concretely.","section":"§5.1 and §7"}],"minor_comments":[{"comment":"The normalization formula 'Xtilde_full = Xtilde_full·σ−μ' is notationally unclear and dimensionally inconsistent; if this is a standardization step, please write it as (X−μ)/σ or the intended equivalent and define σ and μ.","section":"§5.3"},{"comment":"The row 'Reference (Lower bound)' labels the FPD of the validation set as a lower bound, but FPD is not a bound in the mathematical sense; it is the reference value. Please reword.","section":"Table 2"},{"comment":"The sentence 'LION achieves lower Chamfer Distances in the airplane class, with a mean of 0.117' is ambiguous: it is lower than the car and chair means, but not lower than ShapeNet's mean. Please clarify.","section":"§4.3.2"},{"comment":"No code, hyperparameters, or data-preparation details beyond the textual description are provided. A reproducibility appendix with the half-object construction script and training configurations would strengthen the paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The circularity of the symmetry metric is the core issue; it is not a matter of presentation. The paper could become acceptable if the authors reframe the contribution as a reconstruction-enforced symmetrization, add the missing diagnostic experiments, and provide a non-circular evaluation. In its current form, however, the central quantitative claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the headline result doesn't survive contact with the method. The paper reports that S-PVD and S-LION cut mirror-CD from ~0.016 to ~0.004, and so on. But the full shape is constructed as X+ ∪ reflect(X+). Reflecting that whole set gives reflect(X+) ∪ X+, which is the same set. So the mirror-CD for every sample is zero up to sampling noise. The numbers in Figure 5 are not evidence that the model learned symmetry; they restate the reconstruction rule. The stress-test is right: a null model that outputs random half-objects would produce the same mirror-CD distribution.\n\nThat said, the paper isn't empty. Its audit of symmetry in existing generators (PVD, LION, XCube, SLIDE-3D) is a legitimate use of the same metric on generated shapes and shows those models are noticeably asymmetric relative to ShapeNet. That's a real, if modest, observation. The FPD numbers in Table 2 are independent of the circular metric and show that S-LION and S-PVD sometimes produce point clouds closer to the real feature distribution than the baselines—though this comes with consistent degradation on 1-NNA, CD, and EMD, which the authors note.\n\nThe weaker parts are numerous. The global x=0 plane assumption is untested; ShapeNet's canonical alignment is not guaranteed to put the symmetry plane there, and for objects that aren't true mirrors, the half-object dataset teaches the model to complete halves that don't correspond to any natural full shape. The paper does mention this in the limitations section, but it's not just a minor caveat—it's part of the method. There are no error bars or significance tests anywhere, no code or data, and no comparison against existing symmetry-aware generators. The FPD improvements might not survive a proper reimplementation with multiple seeds.\n\nSo the central claim—that half-object training improves symmetry in the learned model—is not supported. What remains is a simple data-augmentation trick plus some interesting FPD findings. That could be a short workshop paper or a base for a revised submission, but as it stands the main quantitative argument is a tautology.\n\nThis paper is for a reader who wants a cautionary example of evaluation design and a cheap audit tool for symmetry in point-cloud generators. It deserves a serious referee only in the sense that the audit and FPD results are worth discussing; the current version needs major restructuring before the symmetry claim can be taken at face value.","headline":"The symmetry gains are hardwired into the reconstruction rule, so the headline numbers don't measure learned behavior—but the audit of existing generators and the FPD results give the paper something to work with.","tokens_in":13702,"tokens_out":3951,"would_cite":false,"duration_ms":41378,"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 shows that training point-cloud diffusion models on mirrored half-objects, then reflecting the generated half back, reduces the mirror-symmetry error of generated shapes from roughly 0.02–0.17 to 0.004–0.008 across three ShapeNet","keywords":["3D point cloud generation","reflection symmetry","diffusion models","data-centric intervention","symmetry evaluation","ShapeNet","half-objects","Chamfer distance"],"falsifier":"Train the same half-object pipeline on a class where shapes have a consistent but off-axis symmetry plane (or where symmetry is weak), or estimate the symmetry plane per object during evaluation; if the mirror-CD improvement disappears or the generated halves are judged implausible, the effect is an artifact of the x=0 assumption rather than a learned structural prior.","tokens_in":12837,"feed_emoji":"🪞","tokens_out":5737,"duration_ms":56329,"temperature":0.7,"pith_summary":"This paper identifies a symmetry gap in 3D point-cloud generative models: although standard metrics (1-NNA, CD, EMD) look competitive, the generated shapes are markedly less mirror-symmetric than real ShapeNet objects. It then proposes a simple data-centric fix: split each training object at the x=0 plane, reflect the left half onto the right, and train the model on these half-objects. During generation, the model samples a half and the full shape is recovered by reflecting it across the plane. Across PVD and LION and the Airplane, Car, and Chair classes, this intervention lowers the mirror Chamfer distance from 0.016–0.173 to 0.004–0.008 and improves Fréchet Point Cloud Distance in five of six configurations, while slightly worsening 1-NNA and raw CD/EMD. The paper argues that symmetry-aware evaluation should become a standard complement to existing benchmarks.","feed_headline":"Mirror-half training cuts symmetry error in 3D generators","feed_subtitle":"PVD and LION trained on one mirrored half of each object cut mirror Chamfer distance from ~0.17 to ~0.005, with better FPD.","key_machinery":"The central mechanism is the half-object transformation: each point cloud is split at the plane x=0, the left half is reflected to the right side, and the two are concatenated, yielding a 'right-half' cloud with the same 15K point count as the original. The generative model is trained on these half-clouds, and at inference the sampled half is reflected across x=0 and merged to form the full shape. The paper's symmetry metric is a mirror-Chamfer distance: the Chamfer distance between a shape and its reflection across the x=0 plane, computed via a Householder reflection. This metric both exposes the symmetry gap and quantifies the intervention's effect.","core_discovery":"The paper's central claim is that the persistent asymmetry in generated 3D shapes is not inherent to diffusion-based generation but a consequence of the model never being forced to respect the reflection prior. Auditing PVD, LION, XCube, and SLIDE 3D with a mirror-Chamfer protocol shows mean symmetry errors of 0.016–0.39, far above ShapeNet's 0.011–0.021. Training PVD and LION on a half-object dataset, where every point cloud is a right-half with the left half mirrored onto it, and then reconstructing full objects by reflecting generated halves, reduces the mirror error to 0.004–0.008—below the real-data baseline—while FPD improves in most cells. The cost is a modest degradation in 1-NNA and","pith_inferences":["A large part of the symmetry gain is mechanical: the reconstruction step reflects the generated half, so any generated half yields a mirror-symmetric full shape. The substantive claim is therefore about the quality of the half-object distribution, which the FPD gains support, but a direct comparison of generated halves to real halves would make this explicit.","The method's success hints that data-centric augmentation with structural priors can regularize generative models more broadly; analogous half-object splits for rotational or translational symmetries could be tested on categories with those properties.","Applying the same half-object protocol to XCube and SLIDE 3D—as the authors propose—would reveal whether the benefit transfers across architectures that operate on voxel hierarchies or sparse latent points, not just point-diffusion models.","Because the half-object construction discards one side of each object, the improved FPD could reflect reduced intra-class variance rather than the symmetry prior per se; ablating with randomly chosen half-planes would isolate the symmetry-specific effect."],"forward_implications":["If correct, the approach shows that reflection symmetry can be instilled into any point-cloud generator by changing only the training data, not the architecture or loss.","Standard fidelity metrics (1-NNA, CD, EMD) miss structural properties; mirror-CD audits reveal failures those metrics hide, so benchmarks should report symmetry alongside the usual scores.","The FPD gains suggest that enforcing symmetry also moves the generated distribution closer to the real-feature distribution, making symmetry a practical proxy for realism.","The reliance on a fixed x=0 plane limits the method to classes with a dominant global reflection axis; extending it to per-object estimated symmetry planes would widen its scope.","The observed drop in 1-NNA and CD/EMD indicates a trade-off between local accuracy and global regularity, so the intervention is not a free lunch for all downstream uses."],"fun_headline_variants":["Mirror-half training fixes 3D generator symmetry","3D generators get symmetry boost from half-object training","Audit reveals symmetry gap, reflection training closes it","Half-mirror data makes 3D point clouds more consistent"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The approach assumes every object in the training set has its true reflection symmetry plane exactly at x=0, and that the half-object dataset therefore represents a natural partial geometry; if real objects' symmetry axes deviate from that plane, the half-objects are artificial and the reported symmetry improvement is imposed by the protocol rather than learned.","fun_headline_variants_meta":{"raw":{"variants":["Mirror-half training fixes 3D generator symmetry","3D generators get symmetry boost from half-object training","Audit reveals symmetry gap, reflection training closes it","Half-mirror data makes 3D point clouds more consistent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1298,"prompt_tokens":796,"completion_tokens":502,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":436}},"tokens_in":540,"tokens_out":502,"duration_ms":6020,"temperature":1.0,"reasoning_tokens":436,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:48:19.062489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same half-object pipeline on a class where shapes have a consistent but off-axis symmetry plane (or where symmetry is weak), or estimate the symmetry plane per object during evaluation; if the mirror-CD improvement disappears or the generated halves are judged implausible, the effect is an artifact of the x=0 assumption rather than a learned structural prior.","supporting_citations":[],"review_version":2}