{"id":"5dc2fd48-d0a1-48c1-84fc-5ebd46b70f3f","arxiv_id":"2411.15723","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GSurf learns a signed distance field supervised by Gaussian splat centers and renders via splatting, yielding compact meshes faster than previous Gaussian-SDF hybrids.","lead":"GSurf trains a signed distance field directly from the centers of 3D Gaussian splats, then renders with splatting, avoiding the slow volume rendering used by earlier hybrid methods. It targets high-quality 3D reconstruction from photos at a fraction of the training time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SDF is fit to Gaussian centers that are not guaranteed to lie on the true surface; the opacity entropy loss (Eq. 3) only binarizes opacity and never constrains center positions, so the supplementary limitation on sparse/reflective scenes directly undermines the central claim's generality.","rationale":"The reader's weakest_assumption exactly identifies the load-bearing premise: Gaussian centers must lie close to the true surface for Lpos to yield correct geometry. My stress-test sharpens this by pointing out that no loss term in the total objective actually enforces that premise. The opacity entropy loss only affects the scalar opacity, not the 3D position of a center; an opaque off-surface disk can still render correctly. Thus the SDF is entirely downstream of center quality, and the paper's own supplementary limitation acknowledges that centers can be sparse or inaccurate under sparse views and reflective surfaces. This does not refute the method's efficiency or its comparative performance on the tested benchmarks; it limits the scope of the central claim to settings where the Gaussian centers happen to align with the true surface. The reader's CONDITIONAL verdict already reflects this risk, so I recommend no change. The concrete test I propose directly measures whether the premise holds on a benchmark scan, which would settle whether the concern is real or merely theoretical.","tokens_in":15059,"tokens_out":5639,"duration_ms":54024,"concrete_test":"On DTU scan 97 (reported CD 0.79), compute the mean distance from the optimized Gaussian centers (the points that supervise Lpos) to the ground-truth point cloud, using the same cropping and evaluation routine that produced the reported CD. If the mean center-to-GT distance is comparable to the reported mesh CD (e.g., greater than half the CD value), then Lpos is fitting the SDF to off-surface anchors, and the claimed high-quality reconstruction is not supported by the central mechanism. A complementary check: rerun with only 10% of the training views; if GSurf's CD degrades substantially more than VolSDF's on the same views, the sparse-view limitation identified in the supplement is a structural weakness of the method, not a corner case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"GSurf's central premise is that Gaussian centers pk can serve as surface keypoints, so the position loss Lpos = Σ|fsdf(pk)| (Eq. 2) anchors the SDF zero level set to the true geometry. This premise is not enforced by any term in the total loss. The opacity entropy loss Lent (Eq. 3) pushes each opacity toward 1, but an opaque Gaussian can still be centered off the true surface: a small opaque disk placed slightly above a surface, with its normal tilted, reproduces nearly the same radiance while shifting its center. Neither Lgs (color, depth distortion, depth-normal consistency) nor the SDF losses (Lpos, Leik, Loff, Lori, Lnor) include a term that moves centers onto the surface; Lpos moves the SDF to the centers, not vice versa. The paper's Sec. 3.4 asserts the entropy loss 'ensures that all Gaussian centroids are correctly positioned on the surface,' but Eq. 3 depends only on opacity values, not on 3D location. Consequently, the reconstruction quality is upper-bounded by the accuracy of the centers as surface proxies. The supplement concedes this: 'Under sparse-view conditions or in challenging scenarios, such as reflective surfaces, the Gaussians may be generated sparsely or inaccurately, compromising the quality of the reconstructed geometry.' That is precisely the setting where Lpos transfers centroid error into the SDF. The DTU comparisons show GSurf is not better than 2DGS/GOF, consistent with the SDF inheriting center noise; the efficiency claim (40 min–1.3 h) is independent of this and appears solid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GSurf, an end-to-end framework that trains a neural signed distance field (SDF) directly under the supervision of 2D Gaussian primitive positions, without a volume-rendering branch. The SDF is fit to Gaussian centroids via a position loss (Lpos, Eq. 2), regularized by Eikonal, off-surface, orientation, and normal-map constraints. An opacity entropy loss (Eq. 3) is introduced to binarize opacities, and an appearance MLP consumes geometric features and gradients from the SDF to model view-dependent color. Experiments on DTU, OmniObjects3D, OO3D-SL, and α-NeuS claim reconstruction quality comparable to VolSDF and NeuS with training times of 40 minutes to 1.3 hours and significantly fewer Gaussian primitives than competing GS-based methods.","tokens_in":15372,"tokens_out":6790,"duration_ms":59459,"significance":"If the central premise holds, GSurf offers a conceptually simple and computationally efficient way to combine the smoothness of SDFs with the speed of Gaussian splatting, potentially useful for practical 3D reconstruction pipelines. The paper is clearly written, the training pipeline is specified in enough detail to reproduce, and the ablations in Tables 2 and 3 isolate the contributions of opacity regularization and geometry-guided appearance modeling. The claim of reduced primitive count is supported by the reported counts in Fig. 4. However, the validity of the whole approach rests on the assumption that optimized Gaussian centroids lie close enough to the true surface to serve as anchors for the SDF zero level set. The paper's own supplementary limitation (§7) concedes that under sparse views or reflective surfaces the Gaussians may be generated inaccurately, which directly undercuts the generality of the headline claim. Moreover, the quantitative evidence is mixed: on DTU the method is not state-of-the-art (Table 1), and the paper's response is to argue that Chamfer distance is misleading without providing an alternative quantitative metric on that dataset.","major_comments":[{"comment":"The central training signal Lpos fits the SDF zero level set to the Gaussian centers pk, but no term in the total loss (Eq. 6) directly pulls pk onto the true surface; the opacity entropy loss (Eq. 3) depends only on opacity values, not on 3D positions. The assertion in Sec. 3.4 that this loss \"ensures that all Gaussian centroids are correctly positioned on the surface\" is not supported by the loss formulation. The supplementary limitation in Sec. 7 explicitly concedes that under sparse-view or reflective conditions the Gaussians may be generated sparsely or inaccurately, which compromises the reconstructed geometry. Because every downstream geometry claim depends on the accuracy of the centers as surface proxies, the paper should either add an explicit surface-pulling mechanism (e.g., a gradient-based pull term as in GS-Pull) or provide empirical evidence that the learned centers lie close to the ground-truth surface in the evaluated scenes (e.g., measure the distance from centers to the true mesh). Without this, the core claim that GSurf learns a reliable SDF from Gaussian positions is not established.","section":"Sec. 3.3–3.4, Eq. (2)–(3)"},{"comment":"The DTU quantitative evaluation shows GSurf's mean Chamfer distance (0.84) is worse than 2DGS (0.80) and GOF (0.74), and the paper's rebuttal that CD is misleading because of cropping and ground-truth noise is introduced only in the text, not supported by an alternative quantitative metric on DTU. The claim of \"high-quality\" reconstruction should be backed by a quantitative measure that favors the method (e.g., F-score or normal consistency on DTU, or a user study), or the paper should temper the claim to \"comparable to VolSDF and NeuS\" without asserting overall superiority to GS-based baselines.","section":"Sec. 4.2, Table 1"},{"comment":"The paper claims robustness to strong lighting and success on semi-transparent surfaces, but no quantitative results are reported for OO3D-SL or α-NeuS; the only support is visual comparisons. This is particularly concerning because the supplementary limitation (Sec. 7) lists reflective surfaces as a failure mode, and Sec. 3.4 states the method is designed for opaque objects. The authors should provide numeric metrics on these datasets or explicitly scope the claims to opaque objects with benign lighting; otherwise the headline claim of generality is not supported.","section":"Sec. 4.2, OO3D-SL and α-NeuS"}],"minor_comments":[{"comment":"The entropy loss -Σ o ln(o) is minimized both as o→0 and o→1, so the text \"to converge to either 0 or 1\" is accurate, but the later statement \"aim for the Gaussians to be as opaque as possible\" is inconsistent with the loss; clarify that the threshold-based pruning (not the loss itself) enforces high opacity for the surviving Gaussians.","section":"Eq. (3) and Sec. 3.4"},{"comment":"The phrase \"and state-of-the-art GS-based reconstruction techniques such as 2DGS [19], GOF [65] and GaussianSurfels [11], and neural implicit surface techniques such as VolSDF [63] and Voxurf [56]\" is repeated verbatim in the caption; please remove the duplicate.","section":"Fig. 1 caption"},{"comment":"The statement \"the SDF is initialized as a sphere\" should be specified more precisely (e.g., the sphere radius and the MLP initialization scheme) to ensure reproducibility, since the paper otherwise does not describe the network initialization.","section":"Sec. 3.6"},{"comment":"The title in the submitted text (\"GSurf: 3D Reconstruction via Signed Distance Fields with Direct Gaussian Supervision\") differs from the title in the abstract (\"GSurf: Learning Signed Distance Fields from Splatting Opaque Gaussians for High-quality 3D Reconstruction\"); please ensure the final manuscript uses one consistent title.","section":"Title"},{"comment":"The per-scene result for Ornament 8 (Ours CD 18.04, NC 0.856 vs. GaussianSurfels 13.95, 0.886) is substantially worse than a leading baseline, but the ranking highlighting may obscure this; consider reporting error bars or discussing such outliers to give a balanced view.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's self-citations to the Ying He group (refs 28, 29, 59, 60, 68) are confined to related work and are not load-bearing for the proposed method, so I see no citation-integrity concern. The paper is honestly reporting that it is not SOTA on DTU, which is a point in its favor. The main issue is that the mechanism of learning an SDF from Gaussian positions is not theoretically justified and the paper's own limitation statement concedes the key failure mode; however, this is addressable with additional experiments or a scoped claim, so major revision rather than rejection is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: GSurf's core move is genuinely new — training an SDF by direct supervision from Gaussian centroid positions (Lpos = Σ|f(pk)|), without a volume-rendering branch. The efficiency claim is the strongest part: 40 min to 1.3 hours on an A800, versus 2 to 16 hours for existing GS+SDF hybrids. The ablations are honest and show that both the opacity entropy loss and the geometry-guided appearance module matter; without them, the numbers get worse. The authors also state their central limitation plainly in the supplementary material: under sparse views or reflective surfaces, Gaussians may be sparse or inaccurate, which compromises geometry.\n\nThe soft spot is load-bearing. The entropy loss (Eq. 3) only drives opacities toward 0 or 1; it has no term that moves a Gaussian center onto the surface. An opaque disk can sit slightly off-surface and, with a tilted normal, reproduce nearly the same radiance. Nothing in the total loss pulls centers to the surface; Lpos moves the SDF to the centers, not the other way around. So the claim in Sec. 3.4 that entropy 'ensures all Gaussian centroids are correctly positioned on the surface' is not supported by the equations. The DTU results are consistent with this concern: mean CD 0.84, worse than 2DGS at 0.80 and GOF at 0.74. The paper's post-hoc argument that CD is misleading on DTU may have some merit, but it is not backed by a robust analysis, and no error bars are given. The opacity pruning threshold is also unspecified, and code is not released, so independent verification is harder.\n\nThat said, the central argument holds for the regime the method targets: dense-view, opaque objects. The method yields compact geometry with smooth surfaces and fast training, and the ablation tables show the components earn their keep. The limitation is real and acknowledged; it bounds the claims rather than invalidating them.\n\nI would send this to peer review. Referees should ask for error bars, the pruning threshold, code availability, and either a softening of the 'ensures' language or evidence that centers stay close to the surface (e.g., a histogram of center-to-mesh distances). This paper is for researchers working on Gaussian-splatting reconstruction who want a faster SDF alternative; it deserves a serious referee, but not a pass on the current evidence.","headline":"A genuinely faster SDF-from-Gaussian-centers method with honest ablations, but the key premise that Gaussian centers lie on the true surface is not enforced and the DTU numbers lag strong baselines; still worth refereeing with targeted fixes.","tokens_in":15964,"tokens_out":3426,"would_cite":false,"duration_ms":31046,"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":"This paper claims that a signed distance field can be learned directly from the centers of Gaussian splatting primitives, supervised by $L_{\\mathrm{pos}} = \\sum_k |f_{\\mathrm{sdf}}(p_k)|$, so that the volume-rendering branch can be…","keywords":["3D reconstruction","signed distance fields","Gaussian splatting","surface reconstruction","neural implicit surfaces","opacity regularization","multi-view reconstruction"],"falsifier":"Measure the distance from the optimized Gaussian centers to the ground-truth mesh on a reflective or sparse-view object: if a substantial fraction of centers lie farther from the surface than the resolution of the mesh, then $L_{\\mathrm{pos}}$ is not actually pinning the SDF to the true surface and the geometry quality should degrade accordingly.","tokens_in":14811,"feed_emoji":"🧊","tokens_out":10648,"duration_ms":87455,"temperature":0.7,"pith_summary":"Signed distance fields give smooth, hole-free surfaces but are normally trained through slow volume rendering, while 3D Gaussian splatting trains fast but yields fragmented geometry. GSurf claims to get the best of both by training an SDF directly on the centers of Gaussian primitives, with the position loss $L_{\\mathrm{pos}} = \\sum_k |f_{\\mathrm{sdf}}(p_k)|$ pulling the zero-level set onto the splats. An entropy-based opacity penalty forces most Gaussians to be opaque, preventing the semi-transparent floaters that would otherwise make the SDF fit the wrong shape. Because splatting is the only renderer, training takes 40 minutes to 1.3 hours on a single A800 GPU, compared with 2–16 hours for hybrid Gaussian+SDF methods that keep a volume-rendering branch. On DTU and OmniObjects3D the paper reports surfaces comparable to VolSDF and NeuS, with smoother geometry and fewer primitives, including on strongly lit and semi-transparent objects.","feed_headline":"Gaussian splat centers teach a signed-distance field directly","feed_subtitle":"No volume-rendering branch: quality matches slow implicit-surface methods in 40 min to 1.3 h.","key_machinery":"The machinery is a joint training loop with three modules: 2D Gaussian disks from 2DGS as discrete primitives, a geometry MLP $f_{\\mathrm{sdf}}$ that outputs signed distance and a geometric feature per center, and an appearance MLP that colors each Gaussian from position, viewing direction, SDF normal, and geometric feature. The load-bearing term is the position constraint $L_{\\mathrm{pos}} = \\sum_k |f_{\\mathrm{sdf}}(p_k)|$: it is what lets the SDF be learned without volume rendering. The opacity entropy loss $L_{\\mathrm{ent}}$ is what makes that constraint valid by forcing each Gaussian's opacity toward 0 or 1, so low-opacity floaters that contribute to images but not to the true surface are pruned. Together these let the zero-level set of the SDF take over the role that TSDF fusion or Poisson reconstruction played in earlier Gaussian methods.","core_discovery":"The central claim is that a correct surface can be recovered by jointly optimizing Gaussians and an SDF MLP with the loss $L_{\\mathrm{sdf}} = \\lambda_1 L_{\\mathrm{pos}} + \\lambda_2 L_{\\mathrm{eik}} + \\lambda_3 L_{\\mathrm{off}} + \\lambda_4 L_{\\mathrm{ori}} + \\lambda_5 L_{\\mathrm{nor}}$, where $L_{\\mathrm{pos}} = \\sum_k |f_{\\mathrm{sdf}}(p_k)|$ supplies direct 3D supervision from Gaussian centers and the other terms enforce Eikonal behavior, off-surface separation, orientation alignment, and normal-map agreement. The authors argue this works because the opacity entropy loss $L_{\\mathrm{ent}} = -\\lambda_6 \\sum_k o_k \\ln(o_k)$ drives Gaussian opacities toward 0 or 1, pruning the transparent off-surface primitives that would pull the SDF away from the true surface, and because normals and geometric features from the SDF are fed into the appearance MLP so that color prediction reinforces geometry. With only Gaussian splatting for rendering, the method reaches Chamfer distances and normal consistency comparable to VolSDF and NeuS on the DTU and OmniObjects3D benchmarks while training in 40 minutes to 1.3 hours, and it reconstructs semi-transparent or strongly lit objects that depth-fusion approaches fail on.","pith_inferences":["A natural extension would be to weight $L_{\\mathrm{pos}}$ by opacity so that already-opaque, on-surface Gaussians dominate the SDF fit, which could stabilize training before the entropy loss has pruned floaters.","If the direct-supervision recipe is the real cause of the speedup, the same loss should work with non-Gaussian oriented points or surfels, which would separate the SDF-learning idea from the splatting renderer.","For reflective or sparse-view objects, where the authors concede Gaussian centers can be sparse or inaccurate, adding a pull or normal-consistency term on the centroids themselves would be a direct extension of the method's own objective.","The paper does not isolate how much of the gain comes from $L_{\\mathrm{pos}}$ versus the opacity entropy loss; removing each separately would reveal which ingredient is load-bearing for the reported geometry quality."],"forward_implications":["If the central claim holds, volume-rendering branches in Gaussian+SDF hybrids are unnecessary, cutting training from 2–16 hours to under 1.3 hours.","SDF regularization replaces holes and noise from unreliable depth fusion with a continuous zero-level set, producing smoother and more complete meshes.","Opacity entropy pruning concentrates mass in opaque, surface-aligned primitives, so scenes can be represented with fewer Gaussians than prior GS-based reconstruction methods.","Feeding SDF normals and geometric features into the appearance MLP captures fine geometric detail that spherical-harmonic appearance alone misses.","Because geometry no longer depends on rendered depth maps, the method can also handle strong-lighting and semi-transparent objects that TSDF and point-fusion approaches get wrong."],"supporting_citations":[{"why":"Supplies the 2D Gaussian splatting renderer and the depth-distortion and depth-normal regularization losses that GSurf inherits as its rendering base.","marker":"[19]"},{"why":"Introduces the Gaussian primitive representation and CUDA rasterization that form the discrete backbone of the method.","marker":"[25]"},{"why":"Provides the SDF-with-volume-rendering baseline whose reconstruction quality GSurf claims to match, defining the quality bar.","marker":"[47]"},{"why":"Provides the other volume-rendered SDF baseline used for both quality and training-time comparisons.","marker":"[63]"},{"why":"Supplies the geometry-appearance MLP architecture and the idea of conditioning appearance on normal and geometry cues.","marker":"[62]"},{"why":"Provides the SDF-from-point-cloud network used in the 2DGS+NSH baseline, showing that treating Gaussian centers as a fixed point set loses detail without image-guided training.","marker":"[54]"}],"fun_headline_variants":["SDF meets splatting: fast, continuous 3D surfaces","Splat centers supervise SDF directly, cut rendering cost","One-hour splatting rivals slow implicit surface methods","Joint SDF-Gaussian optimization yields compact meshes","No volume rendering: splat-based SDF trains in minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that, after optimization, the positions of the rendering splats lie close to the true object surface, so that forcing the signed-distance field to be zero at those positions yields the correct shape.","fun_headline_variants_meta":{"raw":{"variants":["SDF meets splatting: fast, continuous 3D surfaces","Splat centers supervise SDF directly, cut rendering cost","One-hour splatting rivals slow implicit surface methods","Joint SDF-Gaussian optimization yields compact meshes","No volume rendering: splat-based SDF trains in minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1505,"prompt_tokens":999,"completion_tokens":506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":423}},"tokens_in":615,"tokens_out":506,"duration_ms":5206,"temperature":1.0,"reasoning_tokens":423,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:57:53.892363+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the distance from the optimized Gaussian centers to the ground-truth mesh on a reflective or sparse-view object: if a substantial fraction of centers lie farther from the surface than the resolution of the mesh, then $L_{\\mathrm{pos}}$ is not actually pinning the SDF to the true surface and the geometry quality should degrade accordingly.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically ac- curate radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D Gaussian splatting renderer and the depth-distortion and depth-normal regularization losses that GSurf inherits as its rendering base."},{"cited_title":"Neus: Learning neural im- plicit surfaces by volume rendering for multi-view recon- struction","cited_arxiv_id":null,"evidence_quote":"Provides the SDF-with-volume-rendering baseline whose reconstruction quality GSurf claims to match, defining the quality bar."},{"cited_title":"V ol- ume rendering of neural implicit surfaces","cited_arxiv_id":null,"evidence_quote":"Provides the other volume-rendered SDF baseline used for both quality and training-time comparisons."},{"cited_title":"Multiview neu- ral surface reconstruction by disentangling geometry and ap- pearance","cited_arxiv_id":null,"evidence_quote":"Supplies the geometry-appearance MLP architecture and the idea of conditioning appearance on normal and geometry cues."},{"cited_title":"Neural-singular-hessian: Implicit neural representation of unoriented point clouds by enforcing sin- gular hessian","cited_arxiv_id":null,"evidence_quote":"Provides the SDF-from-point-cloud network used in the 2DGS+NSH baseline, showing that treating Gaussian centers as a fixed point set loses detail without image-guided training."}],"review_version":1}