{"id":"bf4e3392-163a-4fdb-9472-280c3a0ff9ab","arxiv_id":"2411.18311","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural-surface-guided Gaussian splatting pipeline propagates mesh edits through a triangle soup proxy, enabling mesh-based editing of reconstructed scenes.","lead":"This paper combines a learned 3D surface model with a fast scene-rendering technique so that edits to the surface, such as bending or moving parts, automatically change the rendered scene. It matters because it makes captured real-world scenes editable for content production, animation, and interactive applications.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 6's T=U'U^{-1} is a rigid rotation built from normalized bases, so mesh scaling and shearing are not propagated to proxy triangles; Section V's claim that scaling and shearing are 'well supported' is contradicted by the method.","rationale":"The strongest claim is that mesh-guided edits propagate through the triangle soup proxy to the recovered appearance. I focused on the mathematics of the propagation step. The reader's weakest_assumption already flags that Eq. 6 is a rotation and therefore scaling and shearing are 'not exactly represented'; my analysis shows the stronger statement that scaling and shearing are not propagated at all, because U and U' are built from normalized vectors and are insensitive to edge lengths and in-plane shears. This makes the Conclusions' explicit claim that scaling and shearing are 'well supported' internally inconsistent with the method, rather than merely approximate. The paper has genuine strengths: the two-stage surface-prior pipeline is coherent, the numerical view-synthesis results are plausible, and code is promised, though I did not verify the repository. The editing evaluation is qualitative, so this mathematical gap is not caught by the experiments. The appropriate outcome is to keep the reader's CONDITIONAL verdict: the condition should require either replacing the rigid T in Eq. 6 with an affine map that captures the full deformation of each mesh face (including scale and shear), or removing scaling and shearing from the list of supported edits and revising the abstract and conclusions accordingly. I marked agreement as 'partial' because the reader framed the issue as an approximation, whereas the derivation shows a complete omission for scale and shear components.","tokens_in":10111,"tokens_out":5010,"duration_ms":47372,"concrete_test":"Run a controlled edit on one scene (e.g., Lego): select a small patch of mesh faces, uniformly scale that patch by a factor of 2 about its centroid without changing face count or topology, then propagate via Eq. 6 and render the edited scene. Recover the Gaussian scales s1 and s2 from the proxy triangles associated with those faces (Eq. 4). If s1 and s2 do not double and the proxy triangle side lengths remain unchanged, Eq. 6 is not propagating scaling. Repeat with an in-plane shear of w2; if U' equals U and V' equals V up to translation, shearing is not transmitted. This directly checks the Section V claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C defines the propagation transform as T = U'U^{-1} (Eq. 6), where U and U' are orthonormal bases constructed from normalized mesh-face edges and the face normal (Eq. 5). Because the columns of both bases are unit vectors, U and U' are rotation matrices, so T is a rotation followed by the translation w'0. As a result, any component of a mesh edit that changes face shape but not face orientation is invisible to the proxy: uniformly scaling W by a factor of 2 leaves U'=U (the columns are normalized), an in-plane shear of w2 leaves the normalized first edge and the face normal unchanged, and the only transmitted information is rotation of the face's local frame plus its new anchor position. The proxy triangle V' therefore is not scaled or sheared, and the Gaussian scale parameters recovered via Eq. 4 remain at their pre-edit values. This is not a minor approximation: the Conclusions (Section V) state that 'transformations such as scaling, shearing, translation and rotation are well supported', and the Abstract advertises 'a wider range of modifications'. The paper's editing experiments (Fig. 7) are qualitative and show no separate scaling/shear propagation check, so the claimed support for exactly these transforms is unsupported by the derivation. The pipeline may correctly propagate rigid rotations and translations of mesh faces, but the central claim of flexible mesh-guided editing overreaches as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a mesh-guided editing pipeline for 3D Gaussian Splatting. First, PermutoSDF recovers a neural SDF and an extracted mesh; then flat Gaussian kernels are trained with opacity conditioned on SDF distance and with initial positions sampled from the mesh. For editing, the Gaussians are converted into a triangle soup proxy, the user edits the extracted mesh, and for each proxy triangle a per-face transformation is computed from orthonormal bases attached to the original and edited mesh faces. The transformation is applied to the proxy triangle, and updated Gaussian parameters are recovered. Experiments report novel-view synthesis PSNR against several baselines, a mesh-resolution study, an initialization ablation, and qualitative editing and physics demonstrations.","tokens_in":10452,"tokens_out":7364,"duration_ms":68771,"significance":"If the propagation claim were corrected, the pipeline would be a plausible practical contribution: it combines a neural surface prior with flat Gaussians and a triangle soup proxy, and the code is released. The two-stage surface-guided training is clearly described, and the initialization ablation is useful. I do not see a circularity problem: the neural SDF is trained self-supervised on the same images, and the edit propagation is a geometric transform rather than a fitted target. However, the current formulation transmits only rigid rotations and translations of the local frame, not scaling or shear, and the editing evaluation is qualitative, so the advertised 'wider range of modifications' is not established.","major_comments":[{"comment":"The propagation transform T = U'U^{-1} is a rigid rotation, not a general deformation. Both U and U' are orthonormal bases built from normalized edge vectors and face normals, so T preserves lengths and angles. Consequently, a mesh face that is uniformly scaled or sheared in-plane leaves the basis U' equal to U (up to rotation of the first edge direction), and the recovered Gaussian scales s1 and s2 from Eq. (4) remain unchanged. This directly contradicts Section V's statement that 'transformations such as scaling, shearing, translation and rotation are well supported' and the Abstract's claim of 'a wider range of modifications'. The sentence that T 'transforms W into W'' is also mathematically false when W' differs from W by scaling or shear. Please replace the basis-to-basis map with an affine map between corresponding triangles (or otherwise justify how scaling and shear are transmitted), and add a dedicated quantitative scaling/shear experiment.","section":"III-C, Eqs. (5)-(6)"},{"comment":"The editing results are entirely qualitative. There are no metrics for edit fidelity, no comparison with GaMeS, SuGaR, Mani-GS, or GSDeformer, and no user study; Tables I-III measure novel-view synthesis and initialization only. Since the paper's central contribution is mesh-guided editing, the evaluation should include at least one quantitative protocol, such as comparing an edited render produced by this method against a mesh-bound baseline and reporting PSNR/LPIPS, or reporting a geometric deformation error on a controlled edit.","section":"IV-B, Fig. 7"},{"comment":"The Introduction states that edits can be applied 'whether directly applied or after optional remeshing,' but Section III-C assumes the edited mesh has the same number of faces as the extracted mesh and relies on a nearest-face association to maintain correspondence. Remeshing generally changes the face count and breaks this correspondence, so the 'optional remeshing' claim is unsupported. Section V correctly lists fixed topology as a limitation and should replace the Introduction's broader promise.","section":"I vs. III-C"},{"comment":"The paper claims to 'fully leverage' mesh topology, but the propagation step uses only per-face nearest-neighbor association and never reads mesh adjacency or connectivity. The method therefore does not demonstrate topology-aware operations such as rigging, region-based deformation, or connectivity-preserving propagation. Please either soften this claim or demonstrate an operation that actually exploits mesh connectivity.","section":"V and Fig. 2"}],"minor_comments":[{"comment":"The Conclusions duplicate the first limitation sentence: 'First, the underlying mesh must provide a structurally sound approximation of the scene geometry. First, the underlying mesh must provide a sufficiently consistent approximation...' Please keep only one formulation.","section":"V"},{"comment":"The phrase 'lightning-dependent effects' should be 'lighting-dependent effects'.","section":"V"},{"comment":"The indexing of the rotation-matrix columns is inconsistent: Section III-B says the normal is the first column of the rotation matrix, while Eq. (4) orders the recovered columns as the normal, the first edge direction, and the second in-plane direction. The triangle vertex definitions in Section III-C should state explicitly which columns correspond to the in-plane scaling axes.","section":"III-C, Eq. (4)"},{"comment":"The opacity function Phi_beta is described as a bell-shaped function, but the text does not define what happens when f_theta(x) is negative; a short sentence clarifying the range of the argument would improve readability.","section":"III-B, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The central derivation issue in Eq. (6) is fixable by using a per-triangle affine map, but it is load-bearing because the claimed support for scaling and shearing appears in the abstract and conclusions. The paper's experimental comparison with prior editing methods is thin: Table I only benchmarks novel-view synthesis, while the editing advantage over GaMeS is asserted without quantitative support. The novelty over GaMeS plus PermutoSDF is incremental and should be stated more carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper does one genuinely new thing—per-face propagation of mesh edits to a triangle-soup proxy for 3DGS—and the rest of the pipeline is a competent combination of PermutoSDF, 3DGSR-style opacity conditioning, and GaMeS. Code is released and the derivations are explicit. But the central claim about flexible editing overreaches: the propagation transform in Eq. 6, T=U'U^{-1}, with U and U' built from normalized edge vectors and the face normal, is a rotation. Scaling or shearing the mesh leaves U'=U, so the proxy triangles keep their original size and shape; only the rigid orientation and anchor translation are transmitted. Section V's statement that 'scaling, shearing, translation and rotation are well supported' is therefore wrong as written. The fix is easy—use the full affine map between the two mesh triangles rather than their orthonormal bases—but as submitted the method supports only rigid per-face motion. That is a load-bearing overclaim relative to the abstract's promise of 'a wider range of modifications.'\n\nThe editing evaluation is qualitative only: Figure 7 shows Blender tools, but there is no quantitative comparison to GaMeS or any other editing baseline, and no metric on edited views. The NVS results are competitive and plausible, though they lack error bars across the eight NeRF Synthetic scenes. The ablation only covers initialization. There is also a contradiction between the Introduction ('whether directly applied or after optional remeshing') and Section V's fixed-topology assumption; the latter is the honest one.\n\nWhat is good: the per-face transformation is a distinct addition over GaMeS, and the idea of using a neural SDF prior to keep Gaussians aligned without binding them to the mesh is cleanly executed. The paper is clearly written and the math is easy to check. I think this deserves a serious referee: it's a useful increment for 3DGS editing, reproducible, and the main flaw is fixable with a corrected claim or a modified transform. For a CV venue I'd send it to review with the expectation of major revision: fix the scaling/shear claim (or implement the affine version), add a quantitative editing evaluation, and reconcile the topology statements. My own verdict would be conditional until then.","headline":"Useful per-face edit propagation for 3DGS via an orthonormal-basis transform, but the claimed support for scaling and shearing is not backed by the math.","tokens_in":11002,"tokens_out":3605,"would_cite":true,"duration_ms":34692,"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":"Neural surface priors let mesh edits rewrite Gaussian Splatting scenes","keywords":["3D Gaussian Splatting","neural signed distance field","triangle soup proxy","mesh-guided editing","novel view synthesis","surface reconstruction","scene editing"],"falsifier":"Take a mesh region and apply a non-uniform scale (stretch along one axis only) while preserving face count; if the rendered appearance does not stretch correspondingly, the orthonormal transfer $T = U'U^{-1}$ has failed to carry the edit, disproving the claim of faithful mesh-guided propagation.","tokens_in":9969,"feed_emoji":"🎨","tokens_out":7950,"duration_ms":62649,"temperature":0.7,"pith_summary":"This paper claims that a scene reconstructed by 3D Gaussian Splatting can be edited intuitively by manipulating a mesh, provided the Gaussians are first trained under a neural surface prior. The authors' pipeline first recovers a neural signed distance field with PermutoSDF, extracts a mesh, and then trains flat Gaussians whose opacity is conditioned on distance to that surface. Each Gaussian is encoded as a triangle in a proxy soup; when the user edits the mesh with the same number of faces, each proxy triangle is re-expressed in the local frame of its nearest face, and the per-face basis change propagates the edit to the appearance. The result is a way to apply standard mesh tools such as translation, rotation, bevel, or knife cuts, and even physics simulations, to 3DGS scenes while keeping render quality close to plain 3DGS. If correct, it closes the gap between 3DGS and mesh-based modeling pipelines without requiring retraining after an edit.","feed_headline":"Neural surface priors let mesh edits rewrite Gaussian Splatting scenes","feed_subtitle":"Rotate, bevel, or simulate physics on a mesh and the scene's appearance updates without retraining.","key_machinery":"The central mechanism is the reversible triangle-soup encoding of flat Gaussians together with a per-face orthonormal-basis transfer map. Each Gaussian with center $m$, rotation $R$, and scaling $s$ is represented as an ordered triangle $V = [v_0, v_1, v_2]$ with $v_0 = m$, $v_1 = s_1 r_1$, $v_2 = s_2 r_2$; this encoding is invertible, so new triangles yield new Gaussian parameters. To propagate an edit, the method finds the nearest face $W$ of the extracted mesh for each proxy triangle, builds an orthonormal basis $U$ from that face, builds the analogous basis $U'$ from the corresponding edited face, and applies the transform $T = U'U^{-1}$ via $V' = T(V - w_0) + w'_0$. The opacity conditioning $\\Phi_\\beta(f_\\theta(x))$ and a normal-alignment regularizer keep the Gaussians flat and surface-aligned, which is what makes the triangle encoding faithful.","core_discovery":"On the paper's own terms, the discovery is that a neural SDF can serve as a geometric prior that makes 3D Gaussian Splatting amenable to mesh-guided appearance editing. The authors show that by conditioning Gaussian opacity on the signed distance to a PermutoSDF surface, fixing one scaling axis so each Gaussian is flat, and sampling initial Gaussian centers from the extracted mesh, each Gaussian becomes a reversible triangle in a triangular proxy. A mesh edit is then transferred by associating each proxy triangle with its nearest original mesh face, computing the orthonormal frame of that face before and after the edit, and applying the rotation $T = U'U^{-1}$ to the proxy triangle. The updated triangle recovers new Gaussian position, orientation, and scale, so the entire appearance follows the user's mesh manipulation. Experiments on NeRF Synthetic and BlendedMVS show PSNR close to plain 3DGS and demonstrate translation, rotation, knife, bevel, randomize, and wind-driven reshaping edits.","pith_inferences":["Because the transfer map is a rotation, a natural testable extension is to replace $U'U^{-1}$ with an affine or per-axis scaling map; this would likely improve fidelity for stretching and shearing edits, but it is not what the paper claims.","The fixed-topology requirement suggests that future work could build a hierarchical or deformation-aware correspondence between the proxy and the mesh, which would let remeshing edits be handled; the paper explicitly leaves this open.","The dependence on a separate PermutoSDF prior implies that the editing fidelity is bounded by the quality of the neural surface reconstruction, so scenes with floating artifacts or disconnected components will show editing artifacts even when the 3DGS appearance is clean.","The admitted shadow limitation (a moved object keeps its original shadow) indicates that the method edits geometry placement but not lighting, so integrating relighting or shadow recomputation is an obvious next step."],"forward_implications":["A user can edit a 3DGS scene with standard mesh tools (translate, rotate, knife, bevel, randomize) and the appearance updates without any retraining or per-edit optimization.","Physics simulations computed on the mesh, such as wind moving a plant's branches, can be transferred frame by frame to the rendered appearance.","Mesh resolution can be lowered for editing speed without hurting render quality, since metrics stay stable across mesh resolutions from 200 to 1000.","The method supports topology-aware workflows such as rigging and finite-element simulation because the underlying mesh retains connectivity, unlike pure triangle-soup approaches."],"supporting_citations":[{"why":"Supplies the base 3D Gaussian Splatting appearance model that the method adapts for editing.","marker":"[15]"},{"why":"Provides the neural signed distance field and extracted mesh used as the geometric prior for Gaussian alignment and editing.","marker":"[26]"},{"why":"Introduces the flat-Gaussian triangle soup proxy encoding that the method relies on to transfer mesh edits to appearance.","marker":"[30]"},{"why":"Supplies the bell-shaped opacity function conditioning Gaussian opacity on SDF distance, used to keep Gaussians surface-aligned.","marker":"[18]"},{"why":"Algorithm used to extract the mesh from the neural SDF; the extracted mesh is what the user edits.","marker":"[17]"},{"why":"Provides the NeRF Synthetic dataset and is the primary comparison baseline for novel-view synthesis evaluation.","marker":"[19]"},{"why":"Additional dataset used to demonstrate editing on real-world scenes.","marker":"[35]"}],"fun_headline_variants":["Mesh edits rewrite Gaussian Splatting scenes via neural priors","Neural surface priors make Gaussian Splatting mesh-editable","Edit a mesh, and Gaussian Splatting scenes update without retraining","No retraining: mesh edits reshape Gaussian Splatting scenes","Sculpt a mesh, and Gaussian Splatting scenes reshape"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the user's mesh edit leaves the number of faces unchanged and that every proxy triangle can be matched to a unique nearest mesh face; any edit that remeshes, splits, or merges faces breaks the correspondence, and the transfer map is a rotation, so non-uniform scaling or shear of faces is not faithfully represented.","fun_headline_variants_meta":{"raw":{"variants":["Mesh edits rewrite Gaussian Splatting scenes via neural priors","Neural surface priors make Gaussian Splatting mesh-editable","Edit a mesh, and Gaussian Splatting scenes update without retraining","No retraining: mesh edits reshape Gaussian Splatting scenes","Sculpt a mesh, and Gaussian Splatting scenes reshape"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001473,"raw_usage":{"total_tokens":5910,"prompt_tokens":925,"completion_tokens":4985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":4895}},"tokens_in":541,"tokens_out":4985,"duration_ms":34117,"temperature":1.0,"reasoning_tokens":4895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:18:54.231857+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a mesh region and apply a non-uniform scale (stretch along one axis only) while preserving face count; if the rendered appearance does not stretch correspondingly, the orthonormal transfer $T = U'U^{-1}$ has failed to carry the edit, disproving the claim of faithful mesh-guided propagation.","supporting_citations":[{"cited_title":"BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks","cited_arxiv_id":null,"evidence_quote":"Additional dataset used to demonstrate editing on real-world scenes."},{"cited_title":"3D Gaussian Splatting for Real- Time Radiance Field Rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the base 3D Gaussian Splatting appearance model that the method adapts for editing."}],"review_version":1}