{"id":"6f38524d-8b17-410f-a933-8608b791426e","arxiv_id":"2507.15212","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MeshMamba applies Mamba state space models to dense 3D articulated mesh generation and single-image human mesh recovery, reaching over 10,000 vertices with competitive accuracy and faster inference than transformers.","lead":"MeshMamba is a neural network that uses Mamba state space models to generate and reconstruct detailed 3D human body meshes with over 10,000 vertices, including clothing and hands. The same backbone powers a diffusion generator (MambaDiff3D) and a single-image human mesh recovery model (Mamba-HMR), running at near real-time speed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'outperforms previous approaches' claim rests on a 1.3-point 1-NNA gap against the authors' own DiffSurf baseline with no reported error bars; sampling variability over 200 test meshes could erase it, so the claim is not yet established.","rationale":"Read in good faith, the paper delivers a clear architecture story: Mamba plus structured serialization trains at more than 10,000 vertices and is much faster than transformer baselines; the ablation in Table 3 shows ordering matters and the limitation section honestly scopes the method to registered, fixed-topology meshes. Those parts survive my review. The point that does not hold up as stated is the headline superiority claim in generation: the only quantitative margin is 1.3 points on 1-NNA against a same-group baseline, with no error bars and only 200 test meshes. 1-NNA is a nearest-neighbor accuracy rate over a small test set; such rates are sensitive to sampling and test-set composition. Since the abstract and conclusion say 'outperforms previous approaches' and 'state-of-the-art,' the burden is on the paper to show the gap is outside noise. This is not an accusation of fabrication; it is a request for evidence. The reader's conditional verdict already captures this in its rationale, although the stated weakest_assumption emphasizes the fixed-topology scope rather than the statistical fragility of Table 1. I therefore recommend no change to the verdict: still CONDITIONAL, pending independent verification or error bars.","tokens_in":16053,"tokens_out":3839,"duration_ms":44205,"concrete_test":"Re-run MambaDiff3D and DiffSurf with at least 3 random seeds on the SURREAL training setup and compute 1-NNA on the same 200-mesh test set, reporting bootstrap 95% CIs for the difference. If the interval includes 0, the 'outperforms' claim in Sec. 5.4 is unsupported. If the authors can release generated meshes or code, an independent third-party recomputation of Table 1 would settle the comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 5.4 declares SOTA on 1-NNA based on Table 1: MambaDiff3D 53.1% vs DiffSurf 54.4% on SURREAL. This is the only quantitative support for the generation claim, but no variance, confidence interval, or repeated-seed result is reported and the evaluation uses 200 test meshes (Sec. 5.1.1). The closest vertex-based competitor DiffSurf [85] is from the same group, so the comparison is not independent. Moreover, the qualitative claims of clothing and hand grasping (Figs. 1, 4) are not covered by this benchmark: Table 1 is body-only SURREAL, while CAPE/GRAB dense-mesh results are shown without quantitative evaluation. The fixed-topology/correspondence assumption in Sec. 4 is explicitly stated and is a genuine scope limit, but it does not by itself threaten the claimed results within that scope. The load-bearing weakness is that the headline quantitative comparison could be noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MeshMamba, a Mamba/SSM backbone for dense articulated meshes. The central technical idea is a vertex serialization derived from DensePose body-part labels or template-mesh xyz coordinates, combined with a position-plus-normal output that is integrated through a Poisson solve. Two applications are presented: MambaDiff3D, a diffusion model for unconditional 3D human mesh generation, and Mamba-HMR, a single-image whole-body mesh regressor. The paper claims state-of-the-art generation on SURREAL (1-NNA 53.1% vs DiffSurf 54.4%) and competitive-to-superior whole-body recovery on UBody (PA-MVE 26.3 mm for the all-body setting), with a 6–9x speed advantage over transformer-based diffusion at roughly 10,000 vertices.","tokens_in":16266,"tokens_out":7038,"duration_ms":76378,"significance":"If the empirical claims hold, the paper makes a useful and timely contribution: it adapts selective state-space models to a mesh setting with a simple correspondence-based ordering, demonstrates scalability to more than 10,000 vertices, and includes ablations that isolate the network block and serialization choices. The fixed-template connectivity and known-correspondence assumption is stated explicitly in Section 4 and acknowledged as a limitation in Section 5.7, so it is a genuine scope limit rather than an internal inconsistency. The main weakness is that the quantitative support for the headline generation claim is thin: a 1.3-point 1-NNA margin over the authors' own DiffSurf baseline, with no variance or confidence intervals, and the clothing/hand results that motivate the dense representation are not quantitatively evaluated. These issues are addressable with additional experiments and statistics.","major_comments":[{"comment":"The state-of-the-art generation claim is not yet established. MambaDiff3D's 53.1% 1-NNA beats DiffSurf's 54.4% by 1.3 points on a 200-mesh test set, but no error bars, confidence intervals, multiple sampling seeds, or repeated training runs are reported. This margin could easily be sampling noise. Please report bootstrap confidence intervals on 1-NNA over the 200 test meshes and/or variance over DDIM sampling seeds, and preferably results over multiple training runs. Since DiffSurf is also from the same group, an independent implementation or external baseline would make the comparison more convincing.","section":"§5.4, Table 1"},{"comment":"The paper's claims about capturing clothing deformations and grasping hands are not quantitatively supported. Table 1 evaluates body-only SURREAL meshes, while CAPE and GRAB appear only as qualitative examples. Please add quantitative dense-mesh generation metrics, such as 1-NNA or coverage on CAPE/GRAB validation splits, or at least a quantitative distortion/identity measure on the generated dense meshes, so that the central claim of dense clothing and hand geometry can be assessed.","section":"§5.3, Figs. 1 and 4"},{"comment":"The statement in Section 5.4 that Mamba-HMR 'outperforms the SOTA' whole-body HMR approaches is also based on single-run PA-MVE/MVE numbers with no reported variance. Because the UBody protocol and training-data mixes differ across baselines, please clarify the exact comparison protocol (image resolution, backbone, fine-tuning procedure, and evaluation alignment), report uncertainty or repeated-run statistics, or soften the claim to 'competitive' as used in the abstract.","section":"§5.4, Table 2"}],"minor_comments":[{"comment":"There is a typo: 'predifined' should be 'predefined'.","section":"§4.1"},{"comment":"'an12×9 feature map' should be 'a 12×9 feature map'.","section":"§5.1.2"},{"comment":"Please clarify whether the 200-mesh SURREAL test set is the reference distribution for 1-NNA and how many generated samples are compared; the metric definition later refers to a 'validation set', which is ambiguous.","section":"§5.1.1"},{"comment":"Please define λ_t for the cosine variance scheduler and state explicitly how the weighting w_t = e^{-λ_t/2} in Eq. (4) is applied in the v-prediction objective; a derivation or exact citation to the formulation would help reproducibility.","section":"§4.2"},{"comment":"The speed comparison should specify the exact transformer configuration, sequence length, batch size, and GPU/memory settings used for the 6–9x claim, so that the result is reproducible.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The recommendation is driven by the missing statistical support for the 1-NNA SOTA claim and the absence of quantitative evaluation on CAPE/GRAB for the dense clothing and hand results. These are addressable with additional experiments, not unfixable flaws. The close same-group comparison with DiffSurf should be scrutinized during revision, but it is not by itself grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: MeshMamba is a genuine application of Mamba to dense articulated meshes, and the best evidence is the whole-body HMR result, where the margin over Multi-HMR-B and SMPLer-X is respectable. The generation \"SOTA\" claim is weak: a 1.3-point 1-NNA edge over the authors' own DiffSurf baseline on 200 test meshes, with no error bars or code. Read the abstract's \"outperforms previous approaches\" with that in mind.\n\nWhat is new and good: the vertex serialization based on DensePose part-IUV maps and template coordinates is a sensible design, and the ablation showing random ordering fails to train is solid evidence that this is the right thing to attend to. The combination of vertex positions and normals via a Poisson solve is a useful trick for dense generation, and Figure 7 supports the qualitative claim. Mamba-HMR runs at 22 FPS and uses the full 10,475-vertex mesh directly, which is a clear step beyond the 500-vertex token limit of earlier vertex-based methods. The speed comparison (6-9x faster than transformer) is plausible and valuable if reproduced.\n\nSoft spots, in proportion: Table 1 is the main one. The 1-NNA gap is small, the evaluation set is only 200 meshes, and no variance or repeated-seed numbers are given. Since DiffSurf is from the same group, this is not an independent comparison. Also, FID and APD are only reported for MambaDiff3D and NRDF, not for all methods, so the \"diverse yet realistic\" claim is under-supported. The CAPE/GRAB dense-mesh results are qualitative only. These are all fixable in revision. The fixed-topology/correspondence assumption is stated clearly in Sec. 4; it limits scope but does not invalidate claims within that scope.\n\nVerdict: This is a serious, honest paper. The core architecture idea is solid, the HMR numbers are encouraging, and the paper flags its own limitations. It deserves full peer review. I would urge reviewers to ask for error bars, code/data release, and a more careful framing of the generation results.","headline":"A genuine Mamba-for-meshes architecture with credible whole-body HMR results, but the generation SOTA claim rests on a thin margin over the authors' own baseline with no error bars.","tokens_in":16832,"tokens_out":2448,"would_cite":true,"duration_ms":25862,"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":"MeshMamba shows that Mamba state space models, with a body-aware vertex ordering, can generate and reconstruct dense articulated 3D meshes with more than 10,000 vertices, including clothing and hands.","keywords":["articulated mesh generation","state space models","Mamba","human mesh recovery","diffusion models","vertex serialization","whole-body pose estimation","clothed human meshes"],"falsifier":"Train MeshMamba on a dataset of raw scans or garment meshes that do not share topology, with no template fitting step. If generation or reconstruction quality collapses compared to the template-registered case, the fixed-connectivity ordering assumption is confirmed as the key constraint; conversely, if it survives with a generic spatial ordering, the ordering prior is not the crucial ingredient.","tokens_in":15813,"feed_emoji":"🧍","tokens_out":4892,"duration_ms":49997,"temperature":0.7,"pith_summary":"The paper argues that Mamba state space models, previously used for long sequences in language and vision, can serve as the backbone for dense articulated 3D meshes if the mesh vertices are serialized in an order that respects body structure. The proposed MeshMamba architecture scales to meshes with over 10,000 vertices, which transformer-based vertex approaches could not handle without upsampling, and it supports both unconditional generation of clothed bodies with grasping hands and single-image whole-body mesh recovery. If the claim holds, Mamba offers a near-linear-complexity alternative to transformers for vertex-level 3D shape modeling, making dense whole-body avatars practical in near real time.","feed_headline":"Mamba handles 10,000-vertex body meshes with clothing and hands","feed_subtitle":"A body-aware vertex order lets Mamba beat transformers at dense human shape generation and recovery.","key_machinery":"The load-bearing mechanism is vertex serialization: mesh vertices are sorted into a one-dimensional sequence using DensePose body-part IUV maps or template xyz coordinates (with axis-order variations), so the selective scanning of Mamba can exploit body structure. The architecture stacks Mamba blocks with skip connections, uses the v-prediction diffusion objective for generation, and for smooth surfaces solves a Poisson system that blends generated vertex positions with predicted per-vertex normals.","core_discovery":"MeshMamba demonstrates that a selective state space model can learn to generate and regress full-resolution articulated meshes—SMPL at 6,890 vertices and SMPL-X at 10,475 vertices—provided the vertices are ordered by body-part UV maps or by the sorted 3D coordinates of a template mesh. On the SURREAL benchmark, the diffusion variant MambaDiff3D achieves a 1-NNA of 53.1% versus 54.4% for DiffSurf, and the recovery variant Mamba-HMR reaches 26.3 mm PA-MVE on UBody whole-body reconstruction, below prior parametric and non-parametric baselines. The authors also show that combining vertex positions with surface normals through a Poisson solve removes local noise while preserving global shape, and that Mamba is 6-9 times faster than a transformer at 10,475 tokens during DDIM sampling.","pith_inferences":["The method's generality is tied to having a template with known correspondences; an interesting test would be applying it to meshes with varying topology, such as loose clothing, after an automatic canonicalization step, and measuring how much the body-part prior helps versus a learned ordering.","The reported speed advantage is measured on specific GPUs; on hardware without optimized selective-scan kernels the gap is smaller, so the practical gain depends on the deployment hardware.","The 1-NNA improvement over DiffSurf is only a few points, so the headline value is arguably the reachable resolution and speed rather than a large accuracy jump.","The Poisson post-processing is not end-to-end differentiable with the diffusion model; a fully differentiable version might further improve smoothness and could be a direct follow-up."],"forward_implications":["Transformer-based vertex methods that cap around 500 tokens can be replaced by Mamba backbones, removing the need for upsampling stages and preserving fingers and face detail.","Whole-body mesh recovery from a single image can run at near real time (around 22 FPS) while directly regressing more than 10,000 vertices, enabling interactive avatar and AR/VR applications.","Dense cloth deformation and hand grasps, previously only representable via parametric fits, can be generated directly as vertex coordinates.","MambaDiff3D's near-linear scaling suggests generation latency will remain manageable as mesh resolution increases beyond SMPL-X.","The same serialization-then-scan recipe may apply to other articulated objects, as the paper's experiments on Animal3D show."],"supporting_citations":[{"why":"Supplies the selective state space backbone (Mamba) that MeshMamba builds on, the core architecture being adapted.","marker":"[21]"},{"why":"Provides DensePose body-part IUV maps used to derive one of the two vertex serialization orderings.","marker":"[58]"},{"why":"Transformer-based diffusion baseline (DiffSurf) that MambaDiff3D is compared against and outperforms; also the U-ViT design MambaDiff3D is inspired by.","marker":"[85]"},{"why":"Transformer-based vertex regression method (METRO) whose roughly 500-token limit and upsampling need Mamba-HMR addresses; used as a qualitative baseline.","marker":"[39]"},{"why":"Technique for combining vertex positions and surface normals via a Poisson solve, adapted for the denoising step to remove noise.","marker":"[52]"},{"why":"SURREAL dataset used to train and evaluate the unconditional generation model and report 1-NNA scores.","marker":"[75]"},{"why":"SMPLer-X, a strong parametric whole-body baseline that Mamba-HMR is compared with on UBody.","marker":"[10]"},{"why":"Neural localizer field baseline for whole-body reconstruction that Mamba-HMR is compared against.","marker":"[63]"}],"fun_headline_variants":["Mamba SSM generates 10K-vertex 3D human meshes with clothing","Body-part ordering unlocks Mamba for dense 3D mesh generation","Mamba diffusion improves 3D human shape generation over prior","Mamba-HMR recovers whole-body 3D meshes with face and hands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All training and test meshes must share the exact connectivity of a template mesh, and point-to-point correspondences must be known, so the vertex ordering is only defined for registered, template-fitted surfaces.","fun_headline_variants_meta":{"raw":{"variants":["Mamba SSM generates 10K-vertex 3D human meshes with clothing","Body-part ordering unlocks Mamba for dense 3D mesh generation","Mamba diffusion improves 3D human shape generation over prior","Mamba-HMR recovers whole-body 3D meshes with face and hands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000391,"raw_usage":{"total_tokens":2084,"prompt_tokens":1002,"completion_tokens":1082,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":999}},"tokens_in":618,"tokens_out":1082,"duration_ms":10491,"temperature":1.0,"reasoning_tokens":999,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:37:42.980653+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MeshMamba on a dataset of raw scans or garment meshes that do not share topology, with no template fitting step. If generation or reconstruction quality collapses compared to the template-registered case, the fixed-connectivity ordering assumption is confirmed as the key constraint; conversely, if it survives with a generic spatial ordering, the ordering prior is not the crucial ingredient.","supporting_citations":[{"cited_title":"Mamba: Linear-time sequence mod- eling with selective state spaces, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the selective state space backbone (Mamba) that MeshMamba builds on, the core architecture being adapted."},{"cited_title":"Dense- pose: Dense human pose estimation in the wild.arXiv, 2018","cited_arxiv_id":null,"evidence_quote":"Provides DensePose body-part IUV maps used to derive one of the two vertex serialization orderings."},{"cited_title":"Diffsurf: A transformer- based diffusion model for generating and reconstructing 3d surfaces in pose","cited_arxiv_id":null,"evidence_quote":"Transformer-based diffusion baseline (DiffSurf) that MambaDiff3D is compared against and outperforms; also the U-ViT design MambaDiff3D is inspired by."},{"cited_title":"End-to-end hu- man pose and mesh reconstruction with transformers","cited_arxiv_id":null,"evidence_quote":"Transformer-based vertex regression method (METRO) whose roughly 500-token limit and upsampling need Mamba-HMR addresses; used as a qualitative baseline."},{"cited_title":"Efficiently combining positions and nor- mals for precise 3d geometry","cited_arxiv_id":null,"evidence_quote":"Technique for combining vertex positions and surface normals via a Poisson solve, adapted for the denoising step to remove noise."},{"cited_title":"Black, Ivan Laptev, and Cordelia Schmid","cited_arxiv_id":null,"evidence_quote":"SURREAL dataset used to train and evaluate the unconditional generation model and report 1-NNA scores."},{"cited_title":"Neural localizer fields for continuous 3d human pose and shape estimation","cited_arxiv_id":null,"evidence_quote":"Neural localizer field baseline for whole-body reconstruction that Mamba-HMR is compared against."}],"review_version":1}