{"id":"b6b2a338-6f8f-493e-898d-8b5817ac31fc","arxiv_id":"2607.10623","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LATO.2 factorizes mesh generation into a vertex-generation flow and a vertex-conditioned connectivity flow, beating joint-latent and autoregressive baselines on geometric fidelity and connectivity quality.","lead":"This paper splits 3D mesh creation into two flow-model steps: first generate vertex positions, then generate the connecting faces conditioned on those vertices. The split makes each stage easier to learn and enables higher-resolution part-wise generation and topology-adaptive editing that a single monolithic latent model cannot do.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim — that factorized vertex-then-connectivity flow is what drives SOTA mesh quality — is not supported by any matched joint-latent ablation; Table 2 confounds architecture, data, and capacity, and the ~3% CD margin has no error bars.","rationale":"I read the paper as advancing two intertwined claims: (1) factorized vertex-then-connectivity flow over two dedicated VAEs is a better recipe than joint-latent flow, and (2) this factorization uniquely enables part-wise high-resolution generation and topology-adaptive editing. The first is load-bearing because the second and the SOTA table depend on it. The decisive missing piece is a controlled comparison that isolates factorization. Table 2 cannot do that, and §4.4 never attempts it. The absence of error bars compounds the issue: the CD gap to LATO is 0.0014 absolute (~3.3% relative), which is plausibly within seed variance for end-to-end generative models; Table 3's non-monotonic row (V-VAE T-Flow CD 0.0393 < GT-Verts T-Flow 0.0399) hints at exactly this noise. The paper's own limitation that T-Flow cannot correct vertex-stage errors makes it even more important to know whether the vertex stage's advantage comes from factorization or from V-VAE engineering (offset head, high-resolution aggregation). None of this demonstrates falsity; it means the evidence is conditional. The Reader's formal weakest assumption was T-Flow generalization to out-of-distribution vertex sets, which is a real concern for the two unique applications; I partially agree, but I think the missing matched ablation is more load-bearing for the central SOTA claim, so I mark partial agreement. The verdict should remain CONDITIONAL/UNCHANGED until a matched joint-latent ablation and error bars are provided.","tokens_in":16294,"tokens_out":6423,"duration_ms":67853,"concrete_test":"Train a matched joint-latent control on the same data, using the same V-VAE and T-VAE latents and the same flow-transformer backbone, but replace the two sequential flow models with a single flow over the concatenated [z_v; z_t] latent, with identical conditioning (image features, c_vn, c_g), then decode both stages jointly. Evaluate on the same test split with 3 seeds, reporting paired per-shape CD/HD/NC differences with 95% bootstrap confidence intervals. If the joint control matches LATO.2 within the CI, factorization is not the cause of Table 2 gains; if LATO.2 beats it by more than the CI width (e.g., ≥5% relative CD), the causal claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's causal story is explicit: §1 claims joint latent encoding 'complicates flow learning and manifests as the broken surfaces observed in prior work,' and that factorization eases learning. For the central claim to hold, factorization itself must be responsible for the Table 2 gains (CD 0.0407 vs LATO 0.0421, MeshFlow 0.0455). But Table 2 compares complete systems that differ simultaneously in VAE design, flow transformer, conditioning, training corpus (including 100K procedural meshes), vertex-count control, and inference procedure. §4.4 ablates only within-LATO.2 components: offset head, VDF downsampling, synthetic data, geometric context c_g, and a deterministic topology predictor. No experiment removes the factorization while holding data, capacity, and backbone fixed — e.g., a joint latent that samples [z_v; z_t] with the same V-VAE/T-VAE latents, same flow transformer, and same training budget. The observed ~3% CD improvement over LATO could therefore be due to any of these confounds. The self-stated limitation that 'geometric inaccuracies from the vertex stage cannot be corrected during topology synthesis' sharpens the point: if factorization's benefit is supposed to be easier learning of each homogeneous distribution, that benefit must be isolated from the many other engineering changes. Without a matched control, the central claim is underdetermined. This is not a demonstrated failure, but it is the load-bearing missing experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LATO.2, a factorized flow-matching framework for explicit 3D mesh generation. The method splits generation into a vertex flow (V-Flow) that produces vertex positions from a coarse voxel scaffold and image/point-cloud conditions, and a topology flow (T-Flow) that generates per-vertex topology latents conditioned on the realized vertices, with a dedicated VAE for each stage. The paper claims state-of-the-art geometric fidelity and connectivity quality (Table 2), and claims two advantages unique to the factorization: part-wise high-resolution generation and topology-adaptive editing (§3.3). Experiments compare against autoregressive and latent-flow baselines, with ablations of the vertex branch (Table 4) and topology branch (Table 5).","tokens_in":16664,"tokens_out":3986,"duration_ms":43266,"significance":"If the central claim is validated, the factorized vertex-then-connectivity recipe is a meaningful contribution to explicit mesh generation, and the sub-voxel offset head is a clean, effective idea for improving vertex reconstruction. The paper is thorough in scope: it evaluates against eight baselines, provides ablations for both branches, reports reconstruction upper bounds, and releases code. However, the evidence for the factorized design's causal benefit is incomplete. The headline improvement over LATO is small (CD 0.0407 vs 0.0421, a 3% relative difference), the reported numbers have no error bars, and the two 'unique advantages' are only qualitatively demonstrated. The central claim is therefore plausible but not yet proven.","major_comments":[{"comment":"The central claim that factorization itself drives the quality improvement is underdetermined. Table 2 compares complete systems that differ in VAE design, flow transformer, training data (including the 100K procedural meshes), vertex-count control, and inference procedure. The ablations in Tables 4–5 only remove components within the factorized design; no experiment removes the factorization while holding data, capacity, and backbone fixed. A matched joint-latent control — e.g., sampling [z_v; z_t] jointly with the same V-VAE/T-VAE latents, same flow transformer, and same training budget — is needed to attribute the gains to factorization rather than to the many concurrent engineering changes.","section":"§4.4, Table 2"},{"comment":"The reported numbers are inconsistent with the expected error hierarchy and have no uncertainty estimates. GT-Verts + T-Flow (CD 0.0399) should upper-bound V-VAE + T-Flow (CD 0.0393), yet the reverse is observed. This inversion indicates noise at the 0.001–0.005 CD level, which is the same order as the headline margin over LATO (0.0421 vs 0.0407). The authors should report means and standard deviations over multiple seeds or runs and, ideally, significance tests, before claiming that the small differences in Table 2 are meaningful.","section":"Table 3"},{"comment":"The two 'unique advantages' of the factorization — part-wise generation and topology-adaptive editing — are presented without quantitative validation. T-Flow is trained on whole-mesh vertex sets; the part-wise pipeline feeds it union-of-parts vertex sets rescaled to full latent volume, and the editing scenarios feed it rotated, stitched, or otherwise modified vertex sets. No topology-validity metrics (e.g., manifoldness, face/edge validity, self-intersection counts, genus accuracy) are reported for these applications. Given the paper's own limitation that 'geometric inaccuracies from the vertex stage cannot be corrected during topology synthesis,' these generalization claims need direct evidence rather than qualitative figures.","section":"§3.3, Figs. 4–5"},{"comment":"The claim of superior 'connectivity quality' is not directly supported by the metrics used. The only topology-related metric in Table 2 is |NC|, Normal Consistency, which measures geometric surface orientation, not topological validity. For a paper whose core contribution is topology modeling, the evaluation should include quantitative measures of mesh validity — e.g., fraction of valid faces, edge manifoldness, vertex valence distributions, or genus/connectivity accuracy. Without such metrics, the 'connectivity quality' component of the headline claim remains unsubstantiated.","section":"§4.1, Table 2"}],"minor_comments":[{"comment":"The text says 'loss weights are omitted for clarity'; these weights should be specified in the appendix or supplementary material, as they affect the VAE training balance.","section":"Eq. (3)"},{"comment":"The first contribution bullet has a trailing comma after 'scaffold,'. Please proofread.","section":"Abstract/Contributions"},{"comment":"There is a typo: 'the vertice positions' should be 'the vertex positions'. Also, the loop-detection process that recovers faces from predicted edges is not described or referenced; this is an important implementation detail for reproducibility.","section":"§3.1"},{"comment":"Reference formatting issue: 'V oleti' should be 'Voleti' in the SV3D and Objaverse-XL entries.","section":"References"},{"comment":"The row 'w/o Train on Synthetic Mesh' reports ACC/F1/Recall/IOU that are substantially lower than the full model, while the 'w/o Offset Head' row shows identical occupancy metrics to the full model. This is plausible, but the caption could state explicitly that the offset head is intended to affect only sub-voxel localization.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is built heavily on the authors' own LATO system (V-VAE architecture, VDF input, and the edge-prediction MLP are all inherited), and the incremental contribution — the factorization plus offset head — is real but requires a matched ablation to separate it from confounding changes. The missing joint-latent control and the absence of quantitative topology metrics are the main technical obstacles. If the authors can add the matched ablation, report error bars, and either provide quantitative results for the editing/part-wise applications or explicitly reframe them as qualitative demonstrations, the paper would be much closer to accept."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nLATO.2 is a serious piece of mesh-generation engineering. The two-stage factorization — vertex flow over a sparse VDF latent, then a connectivity flow conditioned on the realized vertices — is genuinely new in the flow-matching line (Nexus is concurrent and different), and the part-wise generation and topology-adaptive editing demonstrations are the kind of applications that make people pay attention. The paper also does some things right: the V-VAE sub-voxel offset head is a simple fix that clearly helps (Table 4), the ablation on geometric context c_g is sensible, and the main numbers in Table 2 are internally consistent with the reported pipeline.\n\nThe soft spots are real though. The central causal claim — that joint latent encoding causes broken surfaces and factorization fixes it — is never tested directly. Table 2 compares complete systems that differ in VAE design, flow backbone, training data, and inference; nothing removes the factorization while holding everything else fixed. Without a matched joint-latent ablation, the ~3% CD margin could come from any of these changes. The paper's own Table 3 is not reassuring: V-VAE vertices plus T-Flow (0.0393) beats GT vertices plus T-Flow (0.0399), which is only consistent with noise at the 0.001–0.005 level. There are no error bars anywhere, so we can't tell whether the headline differences are significant. Also, the intro says the topology latent is trained with a contrastive objective, but Eq. 7 uses focal loss plus KL. That's a small but embarrassing mismatch.\n\nThe paper's self-stated limitation — that vertex-stage inaccuracies can't be corrected during topology synthesis — is fair, and it matters: the T-VAE reconstruction with GT vertices accounts for roughly 86% of the full-pipeline error, so the autoencoder side is the real part of the story. The editing and part-wise results are purely qualitative; no topology-validity stats. Those are addressable gaps, not a demonstrated failure.\n\nWho's the audience? Anyone working on explicit mesh generation, production pipelines, or latent flow matching. It deserves a serious referee — the factorization idea is worth engaging, and the engineering is solid — but the referee should push hard for a matched ablation and error bars. I'd send it to review, and I'd cite it for the factorization recipe even while doubting the causal claim.\n\nRegards.","headline":"LATO.2 is a genuinely useful factorized mesh-generation recipe, but the central claim that factorization explains the gains is underdetermined without a matched joint-latent ablation.","tokens_in":17246,"tokens_out":2460,"would_cite":true,"duration_ms":23798,"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 claims that generating mesh vertices and connectivity in two separate flow stages, anchored to a shared coarse scaffold, beats joint-latent generation and unlocks part-wise scaling and topology-adaptive editing.","keywords":["3D mesh generation","flow matching","mesh topology","vertex-conditioned generation","latent diffusion","part-wise generation","topology-adaptive editing","sparse voxel scaffold"],"falsifier":"Take a held-out mesh, rotate one part, stitch together vertex sets from two different meshes, or generate a part-wise result, then run the topology flow and measure topology validity directly (count non-manifold edges, self-intersections, and boundary gaps, and compare face counts against ground truth). If connectivity quality on these vertex sets degrades far below whole-mesh generation, the claimed unique advantages collapse even if Table 2's small metric gains hold.","tokens_in":16069,"feed_emoji":"🧊","tokens_out":5500,"duration_ms":55228,"temperature":0.7,"pith_summary":"LATO.2 tries to establish that the reason prior flow-based mesh generators produce broken surfaces is that they encode vertex geometry and discrete connectivity together in one latent space, forcing a single flow to model two statistically dissimilar signals at once. The paper proposes a factorization: a vertex flow generates high-resolution vertex positions from a coarse voxel scaffold, then a topology flow, conditioned on those realized vertices, samples connectivity as per-vertex features decoded into edge probabilities. The claim is that each stage now models a homogeneous distribution, easing learning, and that the factorization makes topology adaptive to any vertex-level change, enabling controllable resolution, part-wise generation at full latent capacity, and editing where moving vertices automatically regenerates valid connectivity. If correct, this offers a practical recipe for explicit, production-oriented mesh generation without long autoregressive token streams.","feed_headline":"Two-stage flow yields cleaner, more accurate 3D meshes","feed_subtitle":"Separating vertex synthesis from vertex-conditioned connectivity unlocks controllable density and one-shot topology-adaptive edits.","key_machinery":"The central machinery is the factorization itself: a two-stage flow matching pipeline over two dedicated latent spaces, connected by a shared coarse voxel scaffold. The vertex stage (V-VAE + V-Flow) compresses vertex displacement fields into a sparse latent and decodes high-resolution vertices through coarse-to-fine subdivision, pruning, and a per-vertex drift vector that corrects quantization error; the topology stage (T-VAE + T-Flow) encodes connectivity as per-vertex latent features using adjacency-masked attention, and decodes edge probabilities with a pairwise MLP, with faces recovered by loop detection. The topology flow is conditioned on realized vertex positions and a coarse geometri","core_discovery":"The central claim is that the discreteness of mesh connectivity is an artifact of modeling it jointly with geometry: conditioned on realized vertices, connectivity reduces to relations among known points in space and can be represented as continuous per-vertex features that decode into pairwise edge probabilities. The paper builds two autoencoders—a vertex VAE that recovers vertices at sub-voxel precision using a per-vertex drift vector, and a topology VAE that learns a per-vertex latent from adjacency-masked attention—and trains two flow matching models over these latents, with a shared coarse voxel scaffold anchoring both stages. Quantitatively, the full pipeline reports lower Chamfer dist","pith_inferences":["A testable extension is to train the topology flow on vertex sets from external sources (scans or deformed meshes) to see whether connectivity prediction generalizes to geometry beyond the training distribution; the paper's editing results suggest it might, but no quantitative topology-validity metrics are reported for those cases.","Since the paper concedes that vertex-stage inaccuracies cannot be corrected during topology synthesis, the practical ceiling of the whole pipeline is set by vertex precision; improving vertex regression (e.g., continuous coordinate prediction instead of offset-from-voxel-center) could yield disproportionate gains.","The quadratic O(N^2) topology decoding will likely become the bottleneck at very high face counts; a hierarchical or blockwise topology flow would be a natural next step and could be evaluated against the current full-scoring decoder.","If the factorization's generalization holds, it could decouple geometry and connectivity research entirely: vertex generators and topology generators could be developed, trained, and swapped independently, sharing only a coarse scaffold."],"forward_implications":["Vertex-count controllability: specifying a target vertex count steers mesh density while preserving shape, giving users direct control over resolution.","Part-wise generation: partitioning the coarse scaffold and generating each part at full latent capacity yields meshes with substantially more faces and finer detail than a monolithic latent permits.","Topology-adaptive editing: user edits, rotations, or stitching of vertex sets propagate to connectivity simply by re-running the topology flow, avoiding hand-crafted topology rules.","Error localization: the paper's ablations indicate the V-VAE reconstructs vertices near-perfectly and the main error comes from the generative flows, so improving the flow stages directly improves final mesh quality.","The paper argues that modeling connectivity as a sampled conditional distribution beats deterministic feed-forward edge completion, supporting the use of latent flow for topology."],"fun_headline_variants":["Separate vertex and topology flows improve 3D mesh quality","Two-stage mesh flow: vertices first, then conditioned connectivity","Factorized mesh generation boosts fidelity and enables part-wise output","Vertex-then-connectivity flow yields cleaner 3D meshes","LATO.2 splits mesh generation for better geometry and topology"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the topology flow learns a conditional distribution that generalizes to vertex sets far outside those it trained on—specifically unions of parts, edited or rotated vertices, and stitched components—and that vertex-stage errors, which the paper states cannot be corrected during topology synthesis, remain small enough that the second stage can still recover valid connectivity.","fun_headline_variants_meta":{"raw":{"variants":["Separate vertex and topology flows improve 3D mesh quality","Two-stage mesh flow: vertices first, then conditioned connectivity","Factorized mesh generation boosts fidelity and enables part-wise output","Vertex-then-connectivity flow yields cleaner 3D meshes","LATO.2 splits mesh generation for better geometry and topology"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1306,"prompt_tokens":720,"completion_tokens":586,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":510}},"tokens_in":464,"tokens_out":586,"duration_ms":6470,"temperature":1.0,"reasoning_tokens":510,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:11:28.059841+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out mesh, rotate one part, stitch together vertex sets from two different meshes, or generate a part-wise result, then run the topology flow and measure topology validity directly (count non-manifold edges, self-intersections, and boundary gaps, and compare face counts against ground truth). If connectivity quality on these vertex sets degrades far below whole-mesh generation, the claimed unique advantages collapse even if Table 2's small metric gains hold.","supporting_citations":[],"review_version":2}