{"id":"eaf9abd2-183f-42f3-9fd8-b106de60ce39","arxiv_id":"2411.09156","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DyGASR reconstructs 3D meshes faster and with less memory by replacing Gaussians with generalized exponential splats, adding SuGaR-style surface alignment, and training at progressively higher resolutions.","lead":"This paper combines generalized exponential splatting, surface-aligned regularization, and a coarse-to-fine resolution schedule to speed up 3D mesh reconstruction from images. It reports lower training time and memory than prior splatting-based methods while roughly matching or slightly improving rendered-image quality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GSR regularizer's SDF conversion in Eq. 11 is only valid for epsilon=2; for other epsilon values f becomes s_min^(1-eps/2)|t|^(eps/2), not a signed distance, so the surface-alignment and mesh-quality claims rest on an unverified assumption.","rationale":"The reader identified the same load-bearing assumption, and I agree. The speed and memory improvements are comparatively robust because they derive from the GES representation and the DRT schedule, and the ablations in Tables 2-3 and the loss curves in Fig. 5 are internally consistent. However, the jointly claimed advantage in mesh reconstruction quality over SuGaR is not supported by geometric mesh metrics, and the SDF conversion is the weakest point in the mathematical argument: the generalized exponential density does not invert to a signed distance for epsilon != 2, and the paper supplies neither a derivation nor a validation experiment for the epsilon values actually used. Missing lambda_1/lambda_2 values and the absence of released code are reproducibility concerns but do not by themselves undermine the central claim; the SDF issue does. I would keep the reader's conditional verdict, conditional on providing the missing derivation and geometric evaluation.","tokens_in":9539,"tokens_out":8943,"duration_ms":108154,"concrete_test":"Independently re-derive Eq. 11 from Eq. 3 for one flattened generalized exponential splat with epsilon != 2: compute f(x)=s_min*sqrt(-2*log d(x)) along the normal direction and check whether level sets {f=c} are equally spaced planes and whether ||grad f|| is constant. If the resulting f is not a Euclidean signed distance, the GSR losses do not enforce the surface alignment claimed, and the mesh-quality comparison in Table 1 requires geometric mesh metrics (e.g., Chamfer distance or F-score) to support the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DyGASR reconstructs meshes faster and with better quality than SuGaR depends on GSR actually aligning generalized exponential splats to the surface. That alignment is enforced through L_sdf and L_nor (Eqs. 12-13), which treat f(x)=s_min*sqrt(-2*log d(x)) (Eq. 11) as a signed distance. For a single flattened splat with shape parameter epsilon, the density in Eq. 3 is exp(-0.5*M^(eps/2)) with M ~ t^2/s_min^2, where t=<x-x_g*, n_g*> is the true signed distance to the tangent plane. Substituting into Eq. 11 gives f ~ s_min^(1-eps/2)*|t|^(eps/2), not |t|. For eps=2 this reduces to an SDF; for eps != 2 the gradient magnitude is not 1, and for eps>2 the expression diverges as s_min -> 0. As a result, L_sdf does not penalize Euclidean distance, and the isosurface offsets used for Poisson meshing become eps- and s_min-dependent, so splats with different learned eps/s_min produce inconsistent surface offsets. No derivation or experiment validates that this conversion is a true SDF for the eps values actually used in training, and no geometric mesh metrics (Chamfer distance, F-score) are reported. The rendering metrics in Table 1 do not directly measure mesh geometry, so the mesh-reconstruction-quality part of the central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DyGASR, a method for accelerated 3D mesh reconstruction that replaces 3D Gaussian splats with generalized exponential splats (GES), adds a surface-alignment regularization term (GSR) inspired by SuGaR, and introduces a cosine-scheduled dynamic resolution training (DRT) strategy. The authors claim that DyGASR reduces training time and memory usage while improving or maintaining rendering quality relative to prior 3DGS-based mesh reconstruction methods, and they support this with experiments on Mip-NeRF360 and Deep Blending datasets. The central novelty is the combination of GES primitives with a density-derived SDF regularizer and a coarse-to-fine resolution schedule.","tokens_in":9932,"tokens_out":9314,"duration_ms":194323,"significance":"If the technical concerns are resolved, the paper would offer a practical contribution: using generalized exponential primitives to reduce the number of splats and a dynamic resolution schedule to shorten training and cut memory use are both plausible and potentially transferable ideas. The paper includes ablations and reports speed/memory gains that could be useful to the community. However, the current evidence for the headline mesh-reconstruction claim is incomplete because the SDF conversion underlying the surface regularizer is not valid for general shape parameters, and because no geometric accuracy metric is reported. The significance of the contribution therefore depends on whether these two issues can be fixed.","major_comments":[{"comment":"The SDF conversion in Eq. (11) is only exact for epsilon=2, and Eq. (10) itself is inconsistent with the GES density in Eq. (9) for epsilon != 2. For the density of a flattened generalized exponential splat, the exponent in Eq. (9) gives exp(-0.5 |t|^epsilon / s_min^epsilon), so Eq. (10) should contain s_min^epsilon in the denominator, not s_min^2. Substituting either the paper's Eq. (10) or the correct density into Eq. (11) yields a function proportional to |t|^(epsilon/2) (with an additional s_min-dependent prefactor in the latter case), whose gradient norm is not 1 for epsilon != 2 and which can diverge as s_min approaches 0 for epsilon > 2. Since epsilon is optimized per splat and no constraint or validation keeps it near 2, L_sdf and L_nor in Eqs. (12)-(13) do not implement the Euclidean surface-distance regularization that the paper claims, and the isosurface offset used in Poisson meshing becomes epsilon- and s_min-dependent. The paper provides no derivation, experiment, or ablation showing that this surrogate behaves like a signed distance for the epsilon values actually encountered. This issue is load-bearing for the surface-alignment and mesh-quality claims, and the authors should either restrict epsilon to 2, derive a correct SDF surrogate for general epsilon, or supply empirical validation (e.g., the distribution of learned epsilon and geometric accuracy on ground-truth meshes).","section":"Section 3.3, Eqs. (10)-(13)"},{"comment":"The evaluation of mesh reconstruction quality relies exclusively on rendering metrics (PSNR, SSIM, LPIPS) computed from views rendered using the reconstructed mesh and its associated splats. No direct geometric accuracy metric (e.g., Chamfer distance, F-score, or Hausdorff distance) is reported for the extracted meshes. Rendering fidelity can be high even when the underlying mesh geometry deviates from the true surface, especially because the final mesh is co-optimized with splats and the rendering is performed with splatting rather than by rasterizing the mesh geometry alone. Consequently, the central claim that DyGASR produces more accurate or higher-quality mesh geometry than SuGaR is under-supported. The authors should report quantitative geometric comparisons on a benchmark with ground-truth scans (e.g., DTU) or otherwise provide direct evidence of geometric accuracy.","section":"Sections 4.1 and 4.3, Table 1"},{"comment":"The regularization weights lambda1 and lambda2 in the total loss of Eq. (14) are never specified, and the GSR description in Section 4.2 does not state their values. The DRT schedule, shape strength rho, and isosurface alpha are given, but without lambda1 and lambda2 the GSR experiments are not reproducible and the relative contribution of L_sdf versus L_nor cannot be assessed. The authors should report these hyperparameters and any sensitivity analysis in the final version.","section":"Section 4.2 and Eq. (14)"}],"minor_comments":[{"comment":"The reported improvements are inconsistent: the abstract and conclusion state a 25% speed increase and 30% memory reduction, while Section 4.3 claims an 85% speed improvement and 37% VRAM decrease versus \"prevalent and efficient 3DGS-based methods.\" The latter number is computed against an average that includes NeuSG, which is neither prevalent nor efficient in the same sense as SuGaR; please clarify the comparison baseline and make the percentages consistent across the paper.","section":"Abstract, Introduction, and Section 4.3"},{"comment":"The caption says \"qualitative analysis\" but the table reports quantitative metrics; this is a wording error.","section":"Table 1 caption"},{"comment":"As detailed in the first major comment, the denominator in Eq. (10) should depend on epsilon (s_min^epsilon) for consistency with Eq. (9); even if the SDF issue were resolved, this equation should be corrected.","section":"Section 3.3, Eq. (10)"},{"comment":"The terms \"shape reset interval\" and \"shape pruning threshold\" are not defined in the paper; please explain these GES-specific hyperparameters or cite the GES reference more precisely.","section":"Section 4.2"},{"comment":"The reference to SuGaR is informal (\"A.Guédon\"); use the full author list and venue information as in the bibliography.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sound high-level idea but is currently under-supported. The SDF conversion issue is not just a presentation problem: it affects the core regularizer and the mesh extraction. I would like to see a corrected derivation, the reported lambda values, and quantitative geometric evaluation before endorsing it. The writing contains many grammatical errors and inconsistent numbers that should be fixed during revision. The comparison with more recent mesh-reconstruction methods (e.g., 2DGS, GOF) would strengthen the paper, though it is not a blocking issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the speed and memory numbers are the real contribution; the surface-alignment math is not.\n\nWhat's actually new: DyGASR is the first system I know of that puts GES into the SuGaR-style mesh-reconstruction pipeline, and the cosine resolution schedule is a sensible add-on. The ablation in Table 3 is honest: each component buys something, and the reported 25% speed-up and 6.3 GB memory saving over SuGaR are internally consistent with the training-time and VRAM columns in Table 1. If those numbers hold on other hardware, that's a useful engineering result for the 3DGS mesh crowd.\n\nThe soft spot is exactly where the stress-test note lands. Eq. (11) defines f(x) = ±s_min sqrt(-2 log d(x)) and then L_sdf and L_nor treat f as a signed distance. For a flattened generalized exponential splat with shape parameter epsilon, that formula gives f ~ s_min^(1-epsilon/2) |t|^(epsilon/2), not |t|, unless epsilon=2. Since epsilon is learned and the paper never reports what values it takes, the claim that GSR aligns splats to the surface is not supported by the derivation. The gradient of f won't have unit norm, and the isosurface offset used for Poisson meshing will vary with epsilon and s_min across splats. This is load-bearing: both regularization terms and the mesh extraction are built on f behaving like Euclidean distance. The authors need to either constrain epsilon to 2 during regularization, derive a proper SDF for general epsilon, or show empirically that the learned epsilon stays close to 2.\n\nAlso missing: lambda1 and lambda2 are never given, and no geometric mesh metrics (Chamfer distance, F-score) are reported. Table 1 only measures rendering quality through the reconstructed mesh and splats, so the \"reconstruction quality\" claim is really about novel-view synthesis, not mesh fidelity. Code would help, but the absent metrics matter more.\n\nIs this paper worth a serious referee? Yes, conditionally. The engineering combination is plausible and the efficiency gains are worth verifying, but the GSR derivation needs major repair and the geometry claims need real metrics before this should be published as a mesh-reconstruction method.","headline":"Useful engineering combination, but the surface-regularization math only works for epsilon=2 and the paper never addresses that.","tokens_in":10388,"tokens_out":4271,"would_cite":false,"duration_ms":48856,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DyGASR replaces 3D Gaussians with generalized exponential splats, aligns them to scene surfaces, and ramps resolution during training to reconstruct 3D meshes about 25% faster than prior splatting-based methods while cutting GPU memory by…","keywords":["3D Gaussian Splatting","3D mesh reconstruction","generalized exponential splatting","surface regularization","dynamic resolution training","novel view synthesis","surface alignment"],"falsifier":"Compute the gradient norm of $f(x)=s_{\\min}\\sqrt{-2\\log d(x)}$ for a flat generalized exponential splat with $\\epsilon=1$ and tiny $s_{\\min}$; if $\\|\\nabla f\\|$ deviates from 1 away from the zero level set, or if $f$ diverges at the splat center, the SDF assumption fails and the normal-loss term is not measuring a geometric normal.","tokens_in":9349,"feed_emoji":"🧊","tokens_out":4859,"duration_ms":77980,"temperature":0.7,"pith_summary":"This paper aims to make 3D mesh reconstruction from posed photographs faster and lighter by changing the primitive used in splatting-based radiance fields. Instead of many tiny 3D Gaussians, it uses generalized exponential splats, whose shape parameter can match sharper signals, so fewer primitives cover the same scene. A surface-regularization term flattens these splats onto the actual surface so that Poisson meshing works reliably, and a cosine-resolution training schedule starts low and rises to full resolution, cutting training time and memory. The authors report that the full system trains about 25% faster than the strong SuGaR baseline and uses about 6.3 GB less GPU memory while slightly improving rendering quality.","feed_headline":"Generalized exponential splats build meshes 25% faster","feed_subtitle":"Replacing Gaussians with shape-flexible splats plus resolution ramp-up cuts training time and GPU memory.","key_machinery":"The machinery is the generalized exponential splat (GES), a primitive with falloff $\\exp\\left(-\\left(\\frac{|x-\\delta|}{\\gamma}\\right)^\\epsilon\\right)$ for a learnable shape parameter $\\epsilon$; $\\epsilon=2$ recovers a Gaussian, while other values yield Laplace-like, triangular, or squarer profiles. Its density field is converted to a signed distance function by $f(x)=\\pm s_{\\min}\\sqrt{-2\\log d(x)}$, and two losses push each splat flat ($s_{\\min}\\to 0$) and its normal perpendicular to the surface, after which Poisson reconstruction extracts the mesh. A dynamic-resolution schedule multiplies the image scale by a cosine ramp from 0.26 to 1.0 over 75% of training, which stabilizes the loss landscape and accelerates convergence.","core_discovery":"The central claim is that combining generalized exponential splatting, surface-alignment regularization, and dynamic resolution training yields faster, cheaper, and higher-quality mesh reconstruction than existing 3D Gaussian splatting-based approaches. On the Mip-NeRF360 and Deep Blending scenes, the method reports a 25% training-time reduction and a 30% memory reduction versus SuGaR, with PSNR rising from 27.28 dB to 27.57 dB on Mip-NeRF360 and from 27.88 dB to 29.05 dB on Deep Blending. The underlying discovery is that the shape parameter of generalized exponential splats lets a scene be represented by fewer primitives while retaining sharp details, and that surface alignment via SDF and normal regularization makes those primitives suitable for explicit mesh extraction.","pith_inferences":["Because the dynamic-resolution strategy is decoupled from the choice of primitive, it could likely be applied to vanilla 3D Gaussian Splatting or other radiance-field methods to obtain similar speed-ups, an extension the paper tests only within DyGASR.","The density-to-SDF formula $f(x)=\\pm s_{\\min}\\sqrt{-2\\log d(x)}$ is mathematically a signed distance only in the Gaussian limit $\\epsilon=2$; for other $\\epsilon$ it is a fractional power of distance and may misrepresent Euclidean distance near the surface, so the method's success may depend on the trained $\\epsilon$ staying close to 2.","The reported memory savings might translate to training on smaller GPUs or larger scenes than the nine evaluated, but the paper does not demonstrate this beyond its current benchmarks."],"forward_implications":["On the nine test scenes, the method reports higher PSNR and SSIM and lower LPIPS than SuGaR while training about 25% faster and using about 6.3 GB less VRAM.","Replacing Gaussians with generalized exponential splats reduces primitive count (e.g., from 2.78M to 1.76M on the bicycle scene) with negligible quality loss.","The GSR regularization drives SDF loss near zero, confirming that the splats flatten into thin, surface-aligned structures suitable for Poisson meshing.","The cosine resolution ramp stabilizes training losses and shortens training by roughly 18 minutes per scene in ablations, independent of the quality gains."],"supporting_citations":[{"why":"Supplies the generalized exponential splatting model and the adjustable shape parameter that replaces ordinary Gaussians.","marker":"[13]"},{"why":"Provides the surface-aligned regularization approach and the strong baseline that DyGASR extends and compares against.","marker":"[14]"},{"why":"Defines the 3D Gaussian Splatting framework and rendering pipeline that GES modifies.","marker":"[12]"},{"why":"Gives the Poisson surface reconstruction algorithm used to extract the final mesh from the flattened splats.","marker":"[1]"},{"why":"Provides the Mip-NeRF360 dataset and the benchmark scenes on which most comparisons and ablations are run.","marker":"[17]"}],"fun_headline_variants":["Exponential splats cut mesh reconstruction time by 25%","Generalized splats speed up 3D mesh building by 25%","Surface-aligned splats boost mesh speed 25%, cut memory 30%","DyGASR: faster mesh reconstruction with flexible splats","Shape-flexible splats align to surfaces for 25% faster meshes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The surface-regularization loss treats the density-to-SDF formula $f(x)=\\pm s_{\\min}\\sqrt{-2\\log d(x)}$ as a true signed distance to the surface, but for shape parameters $\\epsilon\\neq 2$ that formula is not a Euclidean distance and can blow up as $s_{\\min}$ approaches zero.","fun_headline_variants_meta":{"raw":{"variants":["Exponential splats cut mesh reconstruction time by 25%","Generalized splats speed up 3D mesh building by 25%","Surface-aligned splats boost mesh speed 25%, cut memory 30%","DyGASR: faster mesh reconstruction with flexible splats","Shape-flexible splats align to surfaces for 25% faster meshes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1340,"prompt_tokens":980,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":264}},"tokens_in":596,"tokens_out":360,"duration_ms":4257,"temperature":1.0,"reasoning_tokens":264,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:57:33.899943+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the gradient norm of $f(x)=s_{\\min}\\sqrt{-2\\log d(x)}$ for a flat generalized exponential splat with $\\epsilon=1$ and tiny $s_{\\min}$; if $\\|\\nabla f\\|$ deviates from 1 away from the zero level set, or if $f$ diverges at the splat center, the SDF assumption fails and the normal-loss term is not measuring a geometric normal.","supporting_citations":[{"cited_title":"19812–19822 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized exponential splatting model and the adjustable shape parameter that replaces ordinary Gaussians."},{"cited_title":"ACM Transactions on Graphics 42(4), 1–14 (2023)","cited_arxiv_id":null,"evidence_quote":"Defines the 3D Gaussian Splatting framework and rendering pipeline that GES modifies."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Provides the Mip-NeRF360 dataset and the benchmark scenes on which most comparisons and ablations are run."}],"review_version":1}