{"id":"b6896ee7-f756-41e7-abb6-e2507458e720","arxiv_id":"2508.15767","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ATLAS decouples skeleton and shape parameters in a parametric human body model, improving fit accuracy and controllability over previous models like SMPL-X.","lead":"The ATLAS model separates the outer body surface from the internal skeleton in a 3D human model, trained on 600,000 high-resolution scans. It reports tighter fits to unseen people and direct control over height, bone length, and soft tissue, unlike previous models that tie these together.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The two-stage registration (Sec. 4.1) derives the skeleton from triangulated surface keypoints, so the learned skeletal basis may encode soft-tissue thickness; if so, the claimed decoupling is only architectural, not in the learned latent spaces.","rationale":"The paper's core contribution is the explicit decoupling of external shape and internal skeleton, with the claim that this enables keypoint fitting independent of soft-tissue characteristics. The model architecture enforces this separation at the level of parameters (Eq. 2), but the learned skeleton space is obtained from a registration procedure that relies on surface keypoints. If those keypoints are biased by fat/muscle thickness, the skeleton latent variables are contaminated with soft-tissue information, and the claimed independence fails in practice. This is a real correctness risk because it directly undermines the main advantage over vertex-regressed models like SMPL-X, and it is not addressed by the paper's experiments: no comparison against ground-truth bone lengths, no statistical independence test, and no ablation isolating the skeleton-only registration from the surface shape fitting. The reader's weakest_assumption identifies exactly this issue, and I agree. The concern does not necessarily invalidate the whole model—the architecture is still novel and the fitting results are strong—but it makes the central claim conditional on retraining or validating with truly internal skeleton data. Hence the verdict remains CONDITIONAL, with no change from the reader's recommendation.","tokens_in":15785,"tokens_out":7712,"duration_ms":89766,"concrete_test":"Compute the canonical correlation between the 16-dim skeleton latent βk and the 128-dim shape latent βs across the training registrations (or on Goliath-Test). If the largest canonical correlation exceeds ~0.3 (i.e., R²>0.1 shared variance), the skeletal and shape spaces are not statistically independent, indicating the skeleton-only fits absorbed soft-tissue information and the claimed decoupling is incomplete. This test requires no new data and directly checks the independence assumption underlying the paper's central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The model architecture in Eq. 2 guarantees joint locations are functions of βk and θ only, not βs. But the central claim that keypoint fitting is 'independent of external soft-tissue characteristics' depends on the learned skeleton basis actually representing internal bone geometry. Section 4.1 states the data were registered by 'first optimizing registrations with only skeletal parameters and poses. Using triangulated keypoints to regularize joints...' Those keypoints are 2D detections triangulated from the outer surface; joint centers inside the body are not directly observed. Any surface landmark (hip, shoulder, elbow, ankle) sits on skin/fat/muscle, so its 3D position includes soft-tissue thickness. When the skeleton-only fit adjusts bone lengths to match these keypoints, it cannot distinguish a long bone from a thick soft-tissue layer. The resulting skeleton parameters—and hence the 76 attributes and the 16-dim skeleton autoencoder—are therefore statistically confounded with body composition. At inference, fitting βk to keypoints of a new subject can produce bone-length estimates that are biased by that subject's fat/muscle, and the posed mesh will inherit an incorrect scale (e.g., longer limbs for a heavier subject). This re-introduces exactly the vertex-joint correlation the paper claims to eliminate (Fig. 2 critique), just one level removed. The paper's ablations (Table 4) show skeleton-only fits achieve 3.17mm on 3DBodyTex, but this does not validate anatomical correctness or independence; it only shows surface keypoint distances can be explained by skeleton parameters. Thus the decoupling claim is not yet supported by the training procedure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ATLAS, a parametric human body model that explicitly separates a surface shape basis (128 components) from a skeletal basis (16 components over 76 attributes: 15 part scales and 61 bone-length offsets). Surface vertices are first customized in the A-pose while the skeleton stays fixed; the mesh is then scaled and posed via LBS, so joint locations depend only on skeleton and pose parameters (Eq. 2). Pose-dependent deformations are modeled by a sparse, non-linear corrective function (Eqs. 3–4). The model is trained on 600k high-resolution scans plus CAESAR/SizeUSA, and evaluated by fitting to 3DBodyTex and a held-out Goliath-Test, reporting lower vertex errors than SMPL-X, STAR, and SUPR. A single-image fitting pipeline and a runtime comparison are also presented.","tokens_in":16160,"tokens_out":8157,"duration_ms":96295,"significance":"If the decoupling is validated, ATLAS would be a meaningful advance: it offers large-scale training data, a high-resolution mesh (115k vertices), fine-grained skeletal control, and sparse non-linear pose correctives. The architectural formulation is clear and the supplementary material provides detailed training and LBS details. However, the central claim of independence between skeleton and soft tissue needs stronger support, both at the level of what the learned skeletal basis represents and in the statistical reliability of the empirical comparisons.","major_comments":[{"comment":"The abstract and §3.1 claim that keypoint fitting is 'independent of external soft-tissue characteristics,' but the decoupling is only partial. The 76 skeletal attributes include 15 scale parameters that change body-part size (Sec. 3.2), and Eq. (2) applies them through LBS to surface vertices, so β_k can alter soft tissue as well as joints. Additionally, the skeleton-only registrations in Sec. 4.1 are regularized by triangulated keypoints from the outer surface; the resulting skeletal basis may absorb fat/muscle thickness rather than bone geometry. Please validate the skeletal space with an independent anatomical source or show that keypoint-driven β_k changes do not systematically affect body-mass-related surface attributes. Without this, the core decoupling claim is not established.","section":"§3.1–§3.2, §4.1, Eq. (2)"},{"comment":"The main empirical claims are single point estimates without error bars or significance tests. The headline Goliath-Test result (2.34 mm vs 2.78 mm), the 3DBodyTex component curves, the ablation in Table 4, and the monocular fitting results in Table 3 all need per-subject variance and paired statistical tests to support the claim that ATLAS outperforms baselines. Please report mean±std, confidence intervals, and appropriate tests (e.g., Wilcoxon signed-rank) for key comparisons.","section":"§4.2, Fig. 5, Tables 3–4"},{"comment":"The comparison of non-linear vs linear pose correctives reports only a single error reduction (1.82→1.61 mm) on the SMPL dataset. No error bars, number of test sequences, or per-joint error analysis are given, and the non-linear model has strictly more capacity. To support the specific claim that the sparsity mechanism is beneficial, please report statistical significance, parameter-matched baselines, or quantitative sparsity measures (e.g., number of active vertices per joint).","section":"§4.3 (Linear vs Non-Linear Pose Correctives)"}],"minor_comments":[{"comment":"Typo: 'immediate immediate' in the sentence introducing the local neighbor set n(j).","section":"§3.3, Eq. (3)"},{"comment":"Please clarify how the 'number of fitting components' is counted for each model; for ATLAS it appears to be shape+scale, while for baselines it is shape only. Also provide the absolute error values behind the '21.6% lower' claim in a table.","section":"Fig. 5"},{"comment":"Clarify whether the reported runtime includes evaluation of the pose-corrective network and whether it is averaged over a standard pose sequence; also state the exact GPU model beyond 'A100'.","section":"Table 2"},{"comment":"Please describe how Sapiens relative depth is normalized and how the rendered depth is aligned to it; this is important for reproducibility of E_depth.","section":"§3.4"},{"comment":"This ablation is only in the supplement; consider moving it to the main text and adding error bars, since it directly supports the necessity of both shape and skeleton parameters.","section":"Supplementary Table 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll cut to the chase. This is a serious body-modeling paper, probably the strongest industrial entry in a while. The genuinely new thing is the explicit separation of the skeleton basis from the surface shape basis: joint positions are functions of skeleton parameters and pose only, and the surface is customized in the A-pose before skinning. That, plus 76 hand-designed skeletal attributes and the sparse non-linear pose correctives, gives a model that fits scans better than SMPL-X at similar vertex counts and offers clean controllability for edits like shoulder width or arm length. The math in Eqs. 1–4 is coherent, and the ablations (Table 4, linear vs non-linear correctives) support the design choices. The in-the-wild fitting pipeline with depth and mask terms is a nice bonus.\n\nThe soft spots are real but not fatal. First, the central decoupling claim is only architectural until proven. The two-stage registration in Sec. 4.1 optimizes skeleton-only fits regularized by triangulated keypoints. Those keypoints are surface landmarks, so bone-length estimates are at risk of encoding soft-tissue thickness (fat, muscle). The skeleton-only fit achieving 3.17 mm on 3DBodyTex does not prove the skeleton is anatomically independent; it only shows surface distances can be explained by skeleton parameters. So the claim that keypoint fitting is 'independent of external soft-tissue characteristics' is not fully supported by the training data. Second, the empirical comparison is thin: no error bars or significance tests on the 2.34 vs 2.78 mm numbers, and no direct comparison to BLSM or SKEL, which also have skeletal spaces. Without that, the 'first to decouple' narrative is oversold. Third, no code or model release, and the 600k-scan dataset is proprietary, so independent reproduction is impossible.\n\nThose caveats aside, the paper is clear, well-structured, and the qualitative results are compelling. It deserves serious peer review. I'd send it to CVPR/ICCV referees without hesitation, and I'd expect the reviewers to push for error bars and a BLSM comparison. If I were building on body models, I'd cite it.","headline":"Strong industrial body model with a real architectural decoupling, but the keypoint-based skeleton fitting makes the decoupling claim conditional until proven on anatomical data.","tokens_in":16735,"tokens_out":1845,"would_cite":true,"duration_ms":19706,"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":"ATLAS is a parametric 3D human body model that keeps the internal skeleton and external soft tissue as separate, independent control axes.","keywords":["parametric human body model","skeleton-shape decoupling","linear blend skinning","pose-dependent deformations","3D human mesh fitting","single-image human reconstruction","skeletal attributes"],"falsifier":"Take a set of subjects, measure their true bone lengths with MRI, CT, or dual-energy X-ray, and compare against ATLAS bone lengths recovered from the scanning protocol used for training. If the ATLAS skeleton changes systematically with body-fat fraction or soft-tissue thickness at fixed height, the decoupling has failed; stable bone lengths across weight changes would confirm it.","tokens_in":15687,"feed_emoji":"🦴","tokens_out":6205,"duration_ms":63739,"temperature":0.7,"pith_summary":"ATLAS is a 3D human body model that treats the internal skeleton and the outer soft tissue as two separate parameter sets, instead of deriving joint positions from surface vertices as earlier models do. The surface is first reshaped with a linear basis while attached to a fixed template skeleton; then 76 skeletal attributes—15 body-part scales and 61 bone lengths—scale and pose the whole mesh through linear blend skinning. Because the skeleton never changes when the surface changes, an artist can widen shoulders or lengthen arms without disturbing body weight, and a fitting algorithm can use keypoints to fix the skeleton without hallucinating fat or muscle. Trained on 600k high-resolution scans, ATLAS reports lower mean vertex error than the leading vertex-centric model on held-out scans (2.34 mm vs 2.78 mm), and its single-image fitting pipeline predicts meshes more accurately than the previous fitting approach. The result matters because expressive, controllable human models underpin avatars, animation, and image-based human digitization.","feed_headline":"New 3D body model decouples skeleton from soft tissue for sharper fits","feed_subtitle":"ATLAS fits 3D scans more accurately and lets creators edit arm length or weight without side effects.","key_machinery":"The load-bearing object is the decoupled parameterization: a fixed template skeleton, a linear surface-shape basis that edits soft tissue in the rest pose, and a separate skeletal basis of 76 attributes (15 body-part scales and 61 bone-length offsets) that scales and poses the mesh via linear blend skinning. The argument is carried by the identity that makes joint positions independent of the surface shape components. A secondary mechanism is the sparse, non-linear pose corrective, a per-joint-group MLP followed by a geodesic-initialized, L1-regularized sparse mask, which localizes pose-dependent deformations to nearby vertices.","core_discovery":"The paper proposes replacing the vertex-centric paradigm, where joint centers are regressed from the customized surface, with a skeleton-grounded representation. In ATLAS, joint locations are a function only of the skeletal basis and pose, not of the surface shape basis: the external shape basis modifies soft tissue in the rest pose, then a separate skeletal basis of 76 controllable attributes scales and poses the mesh through linear blend skinning. The paper also introduces sparse, non-linear pose correctives: each joint group is processed by a small MLP and then mapped to vertex offsets through a geodesically initialized, L1-regularized sparse mask, so deformations stay local around actuat","pith_inferences":["The clean separation suggests a downstream use the paper does not develop: varying bone length and soft tissue independently in ergonomic or medical simulation, where stature and body composition are separate inputs.","A direct test of the decoupling would compare ATLAS's recovered bone lengths to MRI or CT bone geometry for the same subjects; if the skeleton estimates drift with body-mass index, the independence is only as good as the keypoint-based registration.","The sparse non-linear corrective recipe should transfer to other articulated structures such as hands or quadruped bodies, since it relies only on local joint neighborhoods and geodesic masks.","Because the skeletal basis is only 16 components and the surface basis 128, the model offers a low-dimensional factored latent space that generative models or neural avatars could condition on separately for skeleton and tissue."],"forward_implications":["Body editing becomes deterministic: a single skeletal attribute changes shoulder width or arm length, and a surface component changes weight, without the two interfering.","Keypoint-based fitting no longer needs to distort soft tissue, because skeleton parameters are optimized against keypoints while shape parameters are optimized against silhouettes and depth.","Sparse non-linear pose correctives improve fitting over both sparse linear and dense non-linear alternatives, with gains concentrated around elbows, knees, and shoulders.","The 115k-vertex high-resolution mesh skins in about 5.4 ms on an A100, and the model supports standard lower-resolution topologies for compatibility with existing pipelines."],"supporting_citations":[{"why":"Defines the vertex-centric body model whose joint-from-vertex regression is the paradigm ATLAS replaces.","marker":"[33]"},{"why":"The expressive body model and main quantitative baseline; ATLAS adopts its pose prior and fitting objective while removing the vertex-joint coupling.","marker":"[39]"},{"why":"Supplies the sparse, geodesic-initialized linear pose-corrective formulation that ATLAS generalizes to non-linear correctives.","marker":"[38]"},{"why":"The dense non-linear pose-corrective model ATLAS contrasts with; its lack of sparsity motivates the new corrective design.","marker":"[59]"},{"why":"An earlier decoupled bone-length and shape model that lacks pose correctives, framing the new contribution.","marker":"[54]"},{"why":"A biomechanical-skeleton model derived from surface-mesh models that still inherits a vertex-based shape space, motivating a direct skeleton basis.","marker":"[23]"},{"why":"Provides the relative-depth and mask predictions used in the monocular fitting pipeline.","marker":"[24]"},{"why":"The 3DBodyTex benchmark on which ATLAS reports lower fitting error at matched component counts.","marker":"[49]"}],"fun_headline_variants":["ATLAS untangles skeleton from soft tissue for precise 3D bodies","Skeleton-grounded ATLAS: edit height without bulging side effects","Decoupling skeleton and shape, ATLAS fits tough poses better","ATLAS gives creators separate dials for bone length and weight","Skeleton-first approach in ATLAS improves pose fitting accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The decoupling holds only if the two-stage registration recovers the true internal skeleton from keypoints measured on the outer skin; if flesh thickness leaks into those keypoint fits, skeleton and shape are not fully independent.","fun_headline_variants_meta":{"raw":{"variants":["ATLAS untangles skeleton from soft tissue for precise 3D bodies","Skeleton-grounded ATLAS: edit height without bulging side effects","Decoupling skeleton and shape, ATLAS fits tough poses better","ATLAS gives creators separate dials for bone length and weight","Skeleton-first approach in ATLAS improves pose fitting accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1586,"prompt_tokens":738,"completion_tokens":848,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":769}},"tokens_in":482,"tokens_out":848,"duration_ms":9832,"temperature":1.0,"reasoning_tokens":769,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:40:51.980071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of subjects, measure their true bone lengths with MRI, CT, or dual-energy X-ray, and compare against ATLAS bone lengths recovered from the scanning protocol used for training. If the ATLAS skeleton changes systematically with body-fat fraction or soft-tissue thickness at fixed height, the decoupling has failed; stable bone lengths across weight changes would confirm it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the vertex-centric body model whose joint-from-vertex regression is the paradigm ATLAS replaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The expressive body model and main quantitative baseline; ATLAS adopts its pose prior and fitting objective while removing the vertex-joint coupling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sparse, geodesic-initialized linear pose-corrective formulation that ATLAS generalizes to non-linear correctives."},{"cited_title":"GHUM & GHUML: Generative 3D human shape and articulated pose models","cited_arxiv_id":null,"evidence_quote":"The dense non-linear pose-corrective model ATLAS contrasts with; its lack of sparsity motivates the new corrective design."},{"cited_title":"Blsm: A bone-level skinned model of the human mesh","cited_arxiv_id":null,"evidence_quote":"An earlier decoupled bone-length and shape model that lacks pose correctives, framing the new contribution."},{"cited_title":"Sapiens: Foundation for human vision mod- els","cited_arxiv_id":null,"evidence_quote":"Provides the relative-depth and mask predictions used in the monocular fitting pipeline."},{"cited_title":"3dbodytex: Textured 3d body dataset","cited_arxiv_id":null,"evidence_quote":"The 3DBodyTex benchmark on which ATLAS reports lower fitting error at matched component counts."}],"review_version":1}