{"id":"f7b2b722-057b-49c3-b243-58ec029efda4","arxiv_id":"2412.17808","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Dora-VAE uses sharp-edge sampling plus dual cross-attention to match XCube-level reconstruction with 1,280 latent codes; Dora-bench adds complexity tiers and a sharp normal error metric.","lead":"Dora-VAE improves 3D shape reconstruction by adding extra training points near sharp edges and using separate attention streams for those points, and it reports parity with a dense baseline while using 8x fewer latent codes. The paper also introduces Dora-bench, a complexity-sorted benchmark with a sharp detail metric, to make VAE evaluation more meaningful.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison to XCube is confounded by different mesh-extraction pipelines: the paper attributes XCube's lower scores to NKSR quantization without controlling extraction, so the claimed 8x-latent parity may not hold for the VAE itself.","rationale":"The reader's conditional verdict is sound and I do not recommend changing it. The method is credibly ablated internally: Table 1 and Figure 6 show consistent gains from SES and DCA, and the improvement over uniform sampling is plausible. The most load-bearing external threat is not an internal inconsistency but the uncontrolled comparison to XCube. The paper's own attribution of XCube's lower scores to NKSR mesh extraction is an unverified assumption, and this assumption is directly load-bearing for the abstract claim of parity at 8x smaller latent. Because Dora-bench's SNE masks and complexity strata are built from the same dihedral-angle sharpness that SES targets, even an independent reproduction on the released benchmark could confirm only a metric-specific advantage. No code or data is released, and no error bars are reported, which makes the requested controlled experiment the right condition for accepting the headline claim. I therefore keep the verdict at CONDITIONAL: release code/data and run the matched-extraction comparison before the parity claim is treated as established.","tokens_in":16188,"tokens_out":5001,"duration_ms":50094,"concrete_test":"Run a controlled extraction experiment on the Dora-bench shapes: decode XCube's occupancy or density field and Dora's occupancy field, then reconstruct meshes from both using one identical iso-surfacing routine (e.g., marching cubes at the same grid resolution, or NKSR applied to both outputs). Recompute F-score(0.01), F-score(0.005), CD, and SNE under this matched protocol. If XCube's numbers move to parity or better, the 8x-latent claim is an extraction artifact; if Dora's margins persist, the central claim survives this confound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim: Dora-VAE reconstruction is comparable to XCube-VAE at 1,280 vs. >10,000 latent codes. This is an end-to-end comparison over decoded meshes, and the two systems use different mesh-extraction routes. In Section 4.4 the authors write that XCube's lower performance is due to 'its use of NKSR [24] for mesh extraction, which introduces additional quantization errors', and Appendix B repeats this attribution. But this attribution is asserted, not controlled: there is no experiment that extracts both occupancy fields with the same iso-surface algorithm, or applies NKSR to Dora's output. If XCube's dense latent is actually more accurate and only its final mesh is degraded by NKSR, the headline 'comparable with 8x smaller latent' does not hold for the VAE itself. The second fragile premise, that Dora-bench and SNE are defined by the same dihedral-angle sharpness that SES explicitly targets, reinforces the metric-specificity of the result, but the F-score and CD gaps are large enough that the extraction confound is the sharper threat to the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents Dora-VAE, a transformer-based 3D shape VAE built on 3DShape2VecSet, which replaces uniform surface sampling with a Sharp Edge Sampling (SES) strategy that concentrates points on edges with dihedral angle above τ=30, and encodes uniform and salient point sets with dual cross-attention. It also introduces Dora-bench, a benchmark that groups test shapes into four complexity levels based on the sharp-edge count N_Γ, and a Sharp Normal Error (SNE) metric computed as normal-map MSE in Canny-detected sharp regions. Experiments compare Dora-VAE with XCube, XCube fine-tuned on the same data, Craftsman, and 3DShape2VecSet on F-score, Chamfer distance, and SNE, and show a downstream image-to-3D application with a latent diffusion model.","tokens_in":16451,"tokens_out":6364,"duration_ms":58616,"significance":"The contribution is potentially useful: SES is simple, adds no extra inference parameters, and the ablation in Table 1 shows consistent gains over the uniform-sampling baseline, supporting the claim that both SES and dual cross-attention matter. The benchmark idea of stratifying by geometric complexity addresses a real gap in existing VAE evaluations. However, the headline comparison with XCube rests on an uncontrolled difference in mesh extraction, and the benchmark's complexity definition is aligned with the method's own saliency criterion. The absence of released code, data, benchmark annotations, and error bars is a further barrier to independent verification.","major_comments":[{"comment":"The central claim of 'comparable reconstruction to XCube-VAE with an at least 8x smaller latent' is an end-to-end comparison over decoded meshes, but the two systems use different mesh-extraction pipelines. The paper attributes XCube's lower CD and SNE to 'its use of NKSR [24] for mesh extraction, which introduces additional quantization errors' (Section 4.4), yet no experiment controls for extraction. Please run both occupancy fields through the same iso-surface extraction, or apply NKSR to Dora-VAE's output, and report the same metrics. Without this, the '8x smaller latent' parity stated in the abstract cannot be assigned to the VAE itself.","section":"Section 4.4 / Appendix B"},{"comment":"The evaluation framework is built on the same geometric-saliency definition that Dora-VAE is designed to exploit. Dora-bench complexity levels are thresholds on N_Γ computed with τ=30, which is exactly the threshold used by SES to define salient edges, and SNE restricts evaluation to Canny-detected sharp normal regions. Because the training loss is occupancy MSE, this is not a circular fit of the metric, but it does mean the reported margins are conditional on this sharp-edge notion of detail. Please report results on a random or unstratified test set and, if feasible, under alternative complexity definitions (e.g., curvature-based or local-feature-size based) to show the advantage is not specific to the N_Γ/τ=30 definition.","section":"Section 3.2.1 / Eq. (5) / Section 3.3.1"},{"comment":"No error bars, multi-seed statistics, or bootstrap intervals are reported, and no code, trained models, benchmark splits, or metric implementation are released. Several claimed advantages are small (e.g., SNE at Level 4: 1.579 vs 1.639 for XCube†; F-score(0.01) differences below 0.1 for L1–L3), so the quantitative conclusions are underdetermined without variance estimates. Please provide standard deviations across seeds or bootstrap over test shapes, and release the benchmark construction details so that Dora-bench can actually be used by the community.","section":"Section 4.1 / 4.2 / Table 1"}],"minor_comments":[{"comment":"There are typos and grammatical slips: 'seperately' should be 'separately' (Section 3.2.2); Section 4.2 contains 'as as'; the F-score definition sentence is missing a verb: 'which reconstruction accuracy by computing...' should be 'which measures reconstruction accuracy by computing...'.","section":"Section 3.2.2 / 4.2"},{"comment":"The row grouping for 'Ours w/o DCA' and 'w/o SES, DCA' is visually confusing because the LCL values (256 vs 1280) are not aligned with clearly separated model names; please reformat so each configuration is a distinct, labeled row.","section":"Table 1 / Table S2"},{"comment":"The limitations section lists only compression-efficiency challenges; it should also acknowledge the mesh-extraction confound in the XCube comparison and the metric-specific nature of Dora-bench, since both are directly relevant to interpreting the results.","section":"Appendix D"},{"comment":"The SNE pipeline is described only by a schematic; the text should state precisely how the Canny masks are dilated, how occluded or background pixels are excluded from the MSE computation, and whether the 22 viewpoints are the same for all shapes.","section":"Figure 4 / Section 3.3.2"},{"comment":"The paper provides a project-page link but no URL for code, benchmark data, or model weights; please include these so that the proposed benchmark and metric are reproducible.","section":"Project page"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important problem and the core method is well motivated. My main concern is that the headline result against XCube is not yet disentangled from the mesh-extraction pipeline, which is fixable with a controlled experiment. I would also want code/data release and error bars. The alignment between Dora-bench and the method's own saliency criterion is a correctness risk that the authors should address explicitly, but it does not by itself invalidate the method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible, well-ablated 3D VAE paper. The sharp-edge sampling (SES) is a simple adaptation of importance sampling to mesh surfaces, and the authors show it consistently helps; the dual cross-attention is a modest but clean addition. Dora-bench and SNE are a real evaluation artifact that could be useful to the community. The ablations properly isolate each component, and the downstream image-to-3D experiment shows the improvement transfers.\n\nThe soft spots are real but proportionate. The headline claim—comparable to XCube with an 8x smaller latent—is confounded by mesh extraction. XCube outputs an occupancy field that is turned into a mesh with NKSR, while Dora uses a different iso-surface procedure. The paper attributes XCube's lower scores to NKSR quantization, but never runs a controlled comparison with the same extraction on both. That's the sharpest threat to the headline, and it's fixable with an extra experiment.\n\nSecond, Dora-bench and SNE are defined using the same sharp-edge notion that SES targets. The training loss is occupancy MSE, so this isn't circular in the strong sense, but the evaluation is clearly aligned with the method's inductive bias. A reader should ask whether the margins hold for details defined differently (e.g., curvature, texture). No code or data is released either, which limits verification. No error bars or multi-seed stats—minor, but worth noting.\n\nNone of this kills the paper. The improvement over Craftsman at equal latent length is consistent across metrics and complexity levels, and the ablations support the components. That part doesn't depend on the XCube comparison.\n\nWho it's for: anyone working on 3D VAEs or latent diffusion for shapes. It deserves a serious referee. I'd send it out, with a request for code/data and a controlled extraction comparison.","headline":"Solid 3D VAE paper with a genuinely useful sampling idea, but the headline XCube comparison is not fully controlled and the benchmark is co-designed with the method.","tokens_in":17002,"tokens_out":3870,"would_cite":true,"duration_ms":37126,"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":"Sharp edge sampling lets a compact 3D shape auto-encoder match a dense one with an 8x smaller latent.","keywords":["3D shape VAE","sharp edge sampling","importance sampling","latent diffusion","occupancy fields","mesh reconstruction","Sharp Normal Error","geometry detail benchmark"],"falsifier":"Build a test set of shapes whose important details are smooth or curved with no dihedral angles above 30 degrees and re-run the benchmark; if a uniform-sampling VAE matches or beats Dora-VAE there, the claim that sharp-edge sampling is what preserves detail would be falsified, and a threshold sweep over the angle and Canny parameters would show whether the benchmark's ranking is stable.","tokens_in":15988,"feed_emoji":"🧊","tokens_out":7979,"duration_ms":69965,"temperature":0.7,"pith_summary":"The paper argues that uniform point sampling, the standard way to feed a 3D mesh to a shape auto-encoder, is the main reason reconstructed shapes lose fine geometry. It proposes Dora-VAE, which mixes uniformly sampled surface points with extra points concentrated along sharp edges and encodes the two groups with separate cross-attention branches before merging them. On the paper's own benchmark, Dora-VAE with 1,280 latent codes reconstructs detail at a level comparable to the dense XCube-VAE, which uses more than 10,000 codes. The paper also introduces Dora-bench and a Sharp Normal Error metric to score reconstruction specifically at salient geometric features. The paper concludes that compact latent spaces can support high-detail 3D diffusion generation without the detail loss normally associated with them.","feed_headline":"Sharp-edge sampling lets a compact 3D VAE match a dense one","feed_subtitle":"Focusing samples on sharp edges lets compact 3D latent codes keep the detail diffusion needs.","key_machinery":"Sharp Edge Sampling (SES): identify mesh edges whose adjacent faces meet at a dihedral angle $\\theta_e > \\tau$ with $\\tau=30^\\circ$, collect their vertices, sample $N_{\\mathrm{desired}}=16384$ salient points from that set (interpolating along edges when vertices are scarce), and union them with uniformly sampled points. Dual cross-attention: compute $C_u = \\mathrm{CrossAttn}(P_s, P_u, P_u)$ and $C_a = \\mathrm{CrossAttn}(P_s, P_a, P_a)$ with separate key/value sets for uniform and salient points, and sum the results, so the encoder can treat smooth and sharp regions differently. Dora-bench and SNE: classify test meshes into four detail levels by the count $N_\\Gamma$ of sharp edges, and score reconstructions by the mean squared error between ground-truth and reconstructed rendered normal maps inside dilated Canny edge masks.","core_discovery":"The central claim is that reconstruction fidelity in vector-set 3D VAEs is bottlenecked by the surface sampling strategy, not by encoder capacity. Dora-VAE adds points sampled from edges whose adjacent-face dihedral angle exceeds 30 degrees, then runs separate cross-attention over the uniform and salient point sets and sums the resulting features before self-attention produces the latent code. With this sampling-plus-attention change, the model matches the dense volumetric XCube-VAE in reconstruction quality at a latent size at least 8x smaller (1,280 vs >10,000 codes), and a diffusion model trained on its latents preserves more geometry in single-image 3D generation than one trained on a uniform-sampling baseline. The paper presents Dora-bench as a sharper evaluation, grouping test shapes by sharp-edge count and measuring normal error only inside Canny-detected sharp regions.","pith_inferences":["Extension: because Dora-bench and SNE define detail with the same sharp-edge geometry that SES targets, the reported margins may shrink if detail is defined differently, such as by curvature or by surface texture boundaries.","Extension: the 8x comparison is against XCube as configured with NKSR mesh extraction; a dense baseline with a different extraction step could change the comparison, since the paper attributes part of XCube's lower scores to that step.","Extension: SES is a general mesh-sampling recipe that could be dropped into other point-based 3D networks besides VAEs; the paper only demonstrates it in the VAE and downstream diffusion setting.","Extension: SNE depends on rendering views and Canny thresholds, so sweeping those thresholds would test whether the benchmark's ranking is stable or an artifact of its parameter choices."],"forward_implications":["Uniform point sampling, not model capacity, is the main source of lost geometric detail in vector-set 3D VAEs; replacing it with saliency-aware sampling recovers the detail.","Latent code length can drop from more than 10,000 to 1,280 without a reconstruction-quality penalty, removing a major obstacle to training latent diffusion models on these codes.","Evaluating reconstruction inside sharp-edge regions (SNE) and separately by complexity level exposes quality differences that global F-score or Chamfer distance hide.","A latent diffusion model built on Dora-VAE produces single-image 3D results with sharper geometry than one built on a uniform-sampling VAE under the same training budget.","Ablations show both components matter: removing SES and dual cross-attention, or removing only the dual attention, degrades reconstruction most on high-complexity shapes."],"supporting_citations":[{"why":"supplies the base transformer VAE pipeline (uniform surface sampling, cross-attention encoding, occupancy-field decoding) that Dora-VAE modifies.","marker":"[63]"},{"why":"the dense volumetric XCube-VAE baseline whose reconstruction quality and >10,000-token latent space Dora-VAE is compared against.","marker":"[45]"},{"why":"the compact-latent Craftsman-VAE baseline, fine-tuned on the same data, that Dora-VAE outperforms on complex shapes.","marker":"[31]"},{"why":"provides the mesh preprocessing, multi-resolution latent training strategy, and DiT-based diffusion setup used in the downstream image-to-3D experiment.","marker":"[65]"},{"why":"the NKSR mesh-extraction method whose quantization the paper blames for XCube's lower measured scores, a load-bearing part of the comparison.","marker":"[24]"},{"why":"Poisson disk sampling, the uniform sampling routine that SES supplements and that the ablation removes.","marker":"[61]"},{"why":"Objaverse, the dataset used to train Dora-VAE on roughly 400,000 filtered meshes and to source the highest-complexity benchmark shapes.","marker":"[16]"}],"fun_headline_variants":["Sharp-edge sampling shrinks 3D VAE latent space 8x","Edge-focused sampling lets compact 3D VAE match dense ones","Dora-VAE: Prioritize sharp edges to cut latent size 8x","Sampling sharp edges preserves 3D detail with 8x fewer codes","Dora-bench: Test 3D VAEs on sharp-edge reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison relies on treating sharp corners (edges where the surface bends more than 30 degrees) as the definition of detail, and on blaming the dense baseline's lower scores on its mesh-extraction step rather than on the auto-encoder itself.","fun_headline_variants_meta":{"raw":{"variants":["Sharp-edge sampling shrinks 3D VAE latent space 8x","Edge-focused sampling lets compact 3D VAE match dense ones","Dora-VAE: Prioritize sharp edges to cut latent size 8x","Sampling sharp edges preserves 3D detail with 8x fewer codes","Dora-bench: Test 3D VAEs on sharp-edge reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1299,"prompt_tokens":940,"completion_tokens":359,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":259}},"tokens_in":556,"tokens_out":359,"duration_ms":3714,"temperature":1.0,"reasoning_tokens":259,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:08:39.006243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a test set of shapes whose important details are smooth or curved with no dihedral angles above 30 degrees and re-run the benchmark; if a uniform-sampling VAE matches or beats Dora-VAE there, the claim that sharp-edge sampling is what preserves detail would be falsified, and a threshold sweep over the angle and Canny parameters would show whether the benchmark's ranking is stable.","supporting_citations":[{"cited_title":"3dshape2vecset: A 3d shape representation for neu- ral fields and generative diffusion models.ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023","cited_arxiv_id":null,"evidence_quote":"supplies the base transformer VAE pipeline (uniform surface sampling, cross-attention encoding, occupancy-field decoding) that Dora-VAE modifies."},{"cited_title":"Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies","cited_arxiv_id":null,"evidence_quote":"the dense volumetric XCube-VAE baseline whose reconstruction quality and >10,000-token latent space Dora-VAE is compared against."},{"cited_title":"Clay: A controllable large-scale generative model for creat- ing high-quality 3d assets","cited_arxiv_id":null,"evidence_quote":"provides the mesh preprocessing, multi-resolution latent training strategy, and DiT-based diffusion setup used in the downstream image-to-3D experiment."},{"cited_title":"Neural kernel surface re- construction","cited_arxiv_id":null,"evidence_quote":"the NKSR mesh-extraction method whose quantization the paper blames for XCube's lower measured scores, a load-bearing part of the comparison."},{"cited_title":"Sample elimination for generating poisson disk sample sets","cited_arxiv_id":null,"evidence_quote":"Poisson disk sampling, the uniform sampling routine that SES supplements and that the ablation removes."}],"review_version":1}