{"id":"42acf2e2-ba2f-46a1-a710-ffe051834335","arxiv_id":"2608.10602","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A bi-level field-optimization framework where surface-anchored Gaussians supervise SDF geometry achieves lower chamfer distance and cleaner meshes on NeRF Synthetic and OmniObject3D.","lead":"This paper introduces a surface reconstruction method that anchors 3D Gaussian rendering blobs onto an evolving mesh and optimizes a signed distance field with a bi-level training loop. The result is cleaner and more complete object meshes from multi-view photos, with fewer floating artifacts and better low-resolution reconstruction than prior Gaussian and NeRF baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Outer-loop photometric gradients may be nullified by inner-loop appearance adaptation; only qualitative support (Fig. 7b) is provided.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: geometry gradients flow through a photometric objective after the inner loop has already adapted appearance. This is the correct point of maximum leverage because the paper's headline results (mean CD 9.09e-3 on OmniObject3D, 1.16e-2 at 128 resolution on NeRF Synthetic) are attributed to this bi-level optimization, and the ablation in Table 3 only removes the constraints, not the inner-loop adaptation itself. The opacity constraint and the scale/distribution constraints reduce the appearance-absorption channel but do not eliminate color adaptation or in-face center movement, so the residual loss after inner-loop optimization can be small for geometrically wrong surfaces. The paper provides no quantitative evidence that the outer-loop gradient direction correlates with geometric error; Fig. 7(b) is a single qualitative illustration. Because the concern is empirical and testable rather than a detected contradiction, the appropriate outcome is the same conditional accept as the reader's verdict, with the diagnostic above as the required condition. Secondary issues, such as the undisclosed mean-over-successful-scenes rule in Table 1 and the absence of error bars, further support keeping the verdict conditional rather than upgrading it.","tokens_in":14430,"tokens_out":3343,"duration_ms":50230,"concrete_test":"Run a controlled diagnostic on a NeRF Synthetic scene (e.g., Chair): (1) take a converged SDF and apply a known local deformation, such as shifting a patch of surface vertices outward by 5% of object size; (2) freeze the SDF and run the 1000-iteration inner loop; (3) measure the outer-loop gradient norm with respect to the SDF and its cosine similarity to the gradient of the true Chamfer distance with respect to the SDF at the same point; (4) compare final reconstruction CD when the outer loop uses the photometric loss versus the same loss augmented with a small depth or occupancy supervision. If the cosine similarity is near zero, or if the photometric-only run fails to correct the induced deformation while the render loss stays low, the central claim that rendering supervision faithfully drives geometry is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism in Section 3.3 is that after the inner loop optimizes the detached Gaussian scene G̃ to minimize L_RGB on the fixed surface S, the outer loop copies the optimized parameters back and uses the same photometric loss to update S. This is only a faithful geometry signal if the inner loop's appearance parameters cannot absorb the rendering error. But the inner loop freely adjusts per-Gaussian color and barycentric center weights (Eq. 1), while opacity is fixed near 1 (Eq. 11) precisely to keep Gaussians visible. With K Gaussians per face and per-vertex colors, the inner loop can drive L_RGB toward zero on the current, possibly incorrect mesh; the copied-back residual then contains little information about how the surface should move. The opacity constraint prevents hiding bad geometry with low opacity, but color and in-face position adaptation remain as powerful free variables. The paper's only evidence that gradient isolation helps is Fig. 7(b), a qualitative comparison with no scalar metric. If the outer-loop gradient is not correlated with true geometric error, the reported Chamfer-distance improvements could come from initialization, masking, and regularization rather than from the claimed field-optimization mechanism. No code or supplementary material is provided to check this directly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Gaussian Sculpting, an end-to-end framework for multi-view surface reconstruction that combines an MLP-based signed distance field with 3D Gaussian splatting. Gaussians are anchored to a mesh extracted from the SDF by a differentiable iso-surface extraction module, and their parameters are constrained so that the Gaussians act as geometry-aware rendering proxies. Training is formulated as a bi-level optimization: an inner loop fits Gaussian appearance on a detached copy of the scene for a fixed surface, and an outer loop uses the same photometric loss to update the SDF. The method also introduces an octree-like adaptive subdivision scheme for resolution control. Experiments on OmniObject3D and NeRF Synthetic report lower mean Chamfer distances than several NeRF-based and Gaussian-based baselines, together with qualitative and quantitative evaluations of mesh regularity and ablations of the opacity, scale, and distribution constraints.","tokens_in":14725,"tokens_out":5680,"duration_ms":61059,"significance":"If the empirical claims hold, the paper makes a useful contribution to Gaussian-based surface reconstruction: it introduces a clean surface-anchored Gaussian parameterization, a set of geometric constraints, and a practical bi-level optimization scheme that appears to improve both reconstruction accuracy and mesh quality. The ablation study in Table 3 supports the value of each constraint, and the mesh-quality analysis in Figure 9 is a genuinely useful addition. The explicit discussion of limitations in Section 5 is also to the authors' credit. However, the manuscript as submitted is not yet reproducible: core implementation details are deferred to a missing supplement, no code is provided, and all benchmark numbers are single-run point estimates without statistical support. The central mechanism of the method, the geometry signal in the outer-loop photometric gradient, currently rests on a qualitative comparison only, which is insufficient for a claim that the field-optimization mechanism is what drives the reported gains.","major_comments":[{"comment":"The central geometry-learning signal can be neutralized by the inner loop's appearance adaptation. Since the inner loop freely optimizes per-Gaussian color and barycentric center weights on a detached copy, it can drive LRGB to low values on any fixed, possibly incorrect surface; the outer-loop gradient then reflects the residual after appearance fitting rather than a direct geometric error. The fixed high opacity prevents hiding errors by lowering opacity, but color and in-face position remain strong free variables. The only support for the gradient-isolation mechanism is the qualitative comparison in Fig. 7(b). Please provide a quantitative validation that the outer-loop photometric gradient is correlated with true geometric error, for example by measuring the agreement between gradient-based surface vertex updates and the direction of decreasing Chamfer distance, or by reporting a scalar-metric ablation of joint versus isolated training.","section":"Sec. 3.3, Eqs. (12)-(14)"},{"comment":"The reported mean Chamfer distances in Table 1 exclude failed baseline scenes without disclosure. For example, the NeuS mean of 20.21 is computed over 11 successful scenes rather than 12, and the GSDF mean of 14.68 over 9 scenes, as indicated by the '—' entries. Because Ours has no failures, the headline 'best mean' comparison is not apples-to-apples. The authors should report the number of failures per method, compute means over the common subset of scenes that all methods solve, or include a penalty for failures in the mean.","section":"Table 1, Sec. 4.2"},{"comment":"The paper defers core implementation details to a missing supplement: the exact mechanics of gradient isolation in the bi-level training, the octree-like subdivision consistency rules and vertex remapping, and the precise values of K, λ_b, λ_d, λ_c, τ, δ_target, ε, θ, and λ are all referred to as supplementary material. No supplement is provided with the arXiv submission, and no code is released. As a result, the proposed method cannot be reproduced or independently checked. The authors should provide the supplement and, ideally, code with the revision.","section":"Sec. 3.2, 3.3, 4.1"},{"comment":"All quantitative results are single-run point estimates without error bars, seeds, or significance tests. Given the small differences used to support ranking claims (for example, Ours at 6.10 versus GSDF at 8.14 on Teapot in Table 1, and Ours128 at 1.16 versus 3DGSR at 1.21 in Table 2), the reported improvements could be within run-to-run variation. The authors should report variance over multiple runs or at least explicitly state that only a single run was performed and interpret the results accordingly.","section":"Tables 1-3, Sec. 4.2-4.4"},{"comment":"The resolution-controllability claim is incompletely supported because Table 2 reports only Ours128 for the proposed method. Despite the introduction of a multi-resolution octree-like subdivision scheme, no high-resolution result for Ours is reported, making it impossible to assess whether the subdivision scheme preserves fine details or how the method compares at higher resolutions. The authors should add Ours-high results or an explicit resolution sweep.","section":"Table 2, Sec. 4.3"}],"minor_comments":[{"comment":"The photometric loss in Eq. (12) is the standard 3DGS loss from Kerbl et al. (2023), but it is attributed to Wang et al. (2004); the citation should be corrected.","section":"Sec. 3.3, Eq. (12)"},{"comment":"Setting θ→1 makes σ^{-1}(θ) unbounded; the actual finite opacity value used in the implementation should be stated explicitly.","section":"Sec. 3.1, Eq. (11)"},{"comment":"The connection between λmax(Σ̃_k)≤1 and containment in the minimum enclosing ellipse is stated without proof or derivation; the cited ellipsoid-inclusion result alone does not justify the specific covariance normalization chosen.","section":"Sec. 3.1, Eqs. (4)-(5)"},{"comment":"The text says mesh quality is statistically analyzed using maximum angle, minimum angle, radius ratio, aspect ratio, and sliver percentage, but Figure 9 reports only angle distributions; the remaining metrics should be reported in a table or stated to be in the supplement.","section":"Sec. 4.1"},{"comment":"The claim that SuGar and 2DGS perform poorly on OmniObject3D 'due to their suboptimal handling of datasets with a single object and masked backgrounds' is an unsupported attribution to the baselines' implementation; please substantiate this with evidence or soften the wording.","section":"Sec. 4.2"},{"comment":"The Marching Tetrahedra reference is cited as Lorensen and Cline 1998, which is the Marching Cubes paper; a proper citation for Marching Tetrahedra should be used.","section":"Sec. 2.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears formatted for ACM TOG and would benefit from an editor's attention on reproducibility standards: the missing supplement and single-run comparisons are currently the main barriers beyond the technical concern about the outer-loop gradient signal. The central idea is interesting and the constraint design is well motivated, but the paper should not be accepted without quantitative evidence that the bi-level geometry signal is actually driving the reconstruction improvements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper reports the best mean Chamfer distance on OmniObject3D and NeRF Synthetic across a broad set of baselines, but the evidence for its main mechanism—the bi-level gradient isolation—is thin. It is a serious engineering effort, not a field reorganization.\n\nThe new part is the specific combination: FlexiCubes-style differentiable extraction, surface-anchored Gaussians with fixed opacity, and a detached inner loop. The constraints (opacity, scale, distribution) each pull their weight in the ablation, and the octree subdivision is a practical win for memory. If the numbers are reproducible, it is a direct improvement for asset creation and robotics.\n\nThe softest spot is the central claim that the outer-loop photometric gradient moves the surface. The stress-test note is right: the inner loop can adapt colors and barycentric weights and absorb much of the rendering error. Pinning opacity near 1 removes one degree of freedom, but not all of them. The only support for gradient isolation is Fig. 7(b), which is qualitative. There is no scalar ablation of that component. That needs to be addressed.\n\nAlso missing: code, supplementary material, and error bars. Table 1’s mean silently drops the reported “—” failures, which is an undisclosed rule, and a few hyperparameters (K, λ values, δ_target) are deferred. These are fixable in revision but currently undercut the paper. The claim of per-scene dominance is overstated since Ours is not best on several scenes.\n\nThis is a paper for people working on Gaussian-based reconstruction; they’ll want to see the numbers. It deserves a serious referee, but the referee should demand the missing material. I would send it to peer review with major-revision expectations, not desk reject.","headline":"Solid engineering contribution with best mean CD on two benchmarks, but the key gradient-isolation mechanism is only qualitatively supported and the paper lacks code, error bars, and a supplement.","tokens_in":15249,"tokens_out":2113,"would_cite":true,"duration_ms":25659,"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":"By anchoring Gaussians to an evolving signed distance field, Gaussian Sculpting reconstructs cleaner and more complete object meshes than prior NeRF- and Gaussian-based methods.","keywords":["surface reconstruction","3D Gaussian splatting","signed distance field","differentiable surface extraction","bi-level optimization","neural implicit representation","octree subdivision","mesh quality"],"falsifier":"Freeze the SDF on a specular or textureless object, run the inner appearance loop to convergence, and measure the residual photometric loss against the distance from the extracted mesh to the ground-truth surface. If the inner loop can drive the loss to near zero while the mesh remains far from the true geometry, then rendering error after appearance adaptation is not a faithful geometry signal and the central claim fails.","tokens_in":14234,"feed_emoji":"🧊","tokens_out":9042,"duration_ms":76119,"temperature":0.7,"pith_summary":"This paper claims that surface reconstruction from multi-view images can be done end-to-end by letting Gaussians sculpt a signed distance field, instead of extracting a mesh only as a post-processing step. The method anchors Gaussians to the surface mesh extracted from an evolving SDF, so the Gaussians act as geometry-aware rendering proxies whose photometric errors push the field. A bi-level optimization keeps the two roles separate: the inner loop fits Gaussian appearance to the fixed surface, and the outer loop moves the surface using the residual rendering error. On OmniObject3D the method reports a mean Chamfer distance of $9.09\\times10^{-3}$, the best among the listed NeRF-based and Gaussian-based baselines, and on NeRF Synthetic it reports $1.16\\times10^{-2}$ at 128 resolution, also best. If this holds, editable, watertight object meshes could be produced from limited views without manual cleanup.","feed_headline":"Gaussian Sculpting recovers cleaner meshes from limited views","feed_subtitle":"Bi-level SDF-and-Gaussian training beats NeRF and Gaussian baselines on object-level Chamfer distance.","key_machinery":"The central machinery is a surface-anchored Gaussian layer coupled to a neural signed distance field through a bi-level loop. Each triangular face of the extracted mesh hosts $K$ Gaussians whose centers are learnable barycentric combinations of the face vertices; each covariance is built from the face normal and edge directions with a small normal scale, so the Gaussians are flat disks lying on the surface. Three constraints keep the proxy honest: opacity is fixed near 1, the tangential scale is bounded by the face-aligned minimum enclosing ellipse via a spectral condition $\\lambda_{\\max}(\\tilde{\\Sigma}_k)\\le 1$, and distribution losses penalize boundary collapse, excess clustering, and incomplete coverage of the barycentric domain. The geometry itself is an MLP-valued SDF; a Flexicubes-style differentiable extraction converts it to a mesh each iteration, the inner loop fits the Gaussians to that fixed mesh, and the outer loop backpropagates the remaining photometric error through the synchronized Gaussians into the SDF. An octree-like subdivision refines only surface-intersecting voxels, keeping memory low.","core_discovery":"On its own terms, the paper's central discovery is that constraining Gaussians to lie on an evolving differentiable surface converts photometric rendering error into a usable gradient for signed distance field optimization, and that this yields cleaner and more complete meshes than prior NeRF-based and Gaussian-based methods. The supporting numbers are the mean Chamfer distances: $9.09\\times10^{-3}$ on OmniObject3D and $1.16\\times10^{-2}$ on NeRF Synthetic at resolution 128, both best among the compared baselines. The method also removes redundant floating surfaces, recovers structures in regions with missing viewpoints, and produces meshes with fewer sliver triangles and more equilateral faces.","pith_inferences":["Editorial inference: the bi-level pattern is transferable: any differentiable renderer whose appearance parameters can be fit in an inner loop could be used to sculpt an SDF, not just Gaussian splatting.","Editorial inference: the outer-loop gradient magnitude could serve as a measure of geometric observability; regions where the inner loop absorbs all photometric error are exactly where additional views or priors are needed.","Editorial inference: scaling beyond object-level scenes would likely require retaining converged Gaussians instead of re-optimizing the inner loop from scratch, which the authors note as future work.","Editorial inference: a direct stress test for the method would be a scene with a large textureless region, where appearance can trivially explain the images and geometry has no photometric anchor."],"forward_implications":["Because the mesh is the direct optimization target, downstream editing and simulation receive a watertight, topologically regular surface rather than a noisy point cloud or an opacity field.","The strong result at resolution 128 indicates that the method weakens the usual resolution-versus-memory trade-off for object-level reconstruction.","Fixing opacity near 1 prevents erroneous geometry from being hidden by transparent Gaussians, which is why floating artifacts disappear.","Recovering structures in missing-view regions follows from optimizing a continuous field, since observed viewpoints can indirectly supervise unobserved surface areas."],"supporting_citations":[{"why":"It supplies the 3D Gaussian splatting renderer and the default inner-loop training settings used by the method.","marker":"[Kerbl et al. 2023]"},{"why":"It supplies the differentiable iso-surface extraction module that the paper extends to multi-resolution grids and SDF inputs.","marker":"[Shen et al. 2023]"},{"why":"It is the NeuS baseline whose floating artifacts motivate the surface-anchored design, and a key comparison in the experiments.","marker":"[Wang et al. 2021]"},{"why":"It is the 2DGS post-processing baseline that the paper compares against on both datasets.","marker":"[Huang et al. 2024]"},{"why":"It is the Gaussian Opacity Fields baseline whose evaluation scripts and metrics the paper adopts for Chamfer distance and mesh quality.","marker":"[Yu et al. 2024b]"},{"why":"It is the GSDF joint Gaussian-SDF baseline with decoupled training that the paper aims to improve upon.","marker":"[Yu et al. 2024a]"},{"why":"It is the 3DGSR joint baseline that achieves the second-best mean Chamfer distance on NeRF Synthetic.","marker":"[Lyu et al. 2024]"},{"why":"It provides the NeRF baseline and the NeRF Synthetic dataset used for the resolution-controlled experiments.","marker":"[Mildenhall et al. 2021]"},{"why":"It provides the OmniObject3D dataset of real scanned objects used for the completeness and accuracy evaluation.","marker":"[Wu et al. 2023]"},{"why":"It gives the ellipsoid-inclusion condition used to formulate the minimum-enclosing-ellipse scale constraint.","marker":"[Calbert et al. 2023]"}],"fun_headline_variants":["Gaussian Sculpting: cleaner meshes from limited views","Bi-level training sharpens Gaussian surface geometry","Gaussian anchors fix sparse-view surface holes","Field-based sculpting removes redundant surfaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the photometric error left after the inner appearance loop has done its best is a reliable signal for moving the surface; if appearance parameters can absorb that error without any geometry change, the outer-loop gradients will push the SDF in misleading directions.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian Sculpting: cleaner meshes from limited views","Bi-level training sharpens Gaussian surface geometry","Gaussian anchors fix sparse-view surface holes","Field-based sculpting removes redundant surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000385,"raw_usage":{"total_tokens":2015,"prompt_tokens":904,"completion_tokens":1111,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1053}},"tokens_in":520,"tokens_out":1111,"duration_ms":23096,"temperature":1.0,"reasoning_tokens":1053,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:05:06.703831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Freeze the SDF on a specular or textureless object, run the inner appearance loop to convergence, and measure the residual photometric loss against the distance from the extracted mesh to the ground-truth surface. If the inner loop can drive the loss to near zero while the mesh remains far from the true geometry, then rendering error after appearance adaptation is not a faithful geometry signal and the central claim fails.","supporting_citations":[{"cited_title":"Egidio and Raphaël M","cited_arxiv_id":null,"evidence_quote":"It gives the ellipsoid-inclusion condition used to formulate the minimum-enclosing-ellipse scale constraint."}],"review_version":1}