{"id":"df16efeb-f4db-4735-b29b-3be015aef890","arxiv_id":"2411.12440","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"3DLS replaces Gaussian kernels with bounded linear kernels plus distribution alignment and gradient scaling, yielding slightly better fidelity and faster rendering than 3DGS on some scenes.","lead":"A 3D rendering method replaces the soft Gaussian blobs used in 3D Gaussian Splatting with cone-shaped linear kernels to capture sharper edges and fine details. It reports modest quality gains on some benchmark scenes and about a 30% speedup, though the improvements are uneven and rely on several tuned parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (10) asserts, rather than derives, the 2D projection of the linear kernel; the rendered footprint is not the projection of a 3D linear ellipsoid, so the paper's theoretical grounding for the quality claim is unsupported.","rationale":"The paper presents a plausible engineering contribution: a bounded 2D linear splat parameterized by 3D covariances, with two training stabilizers and a custom CUDA rasterizer. The reported quality and speed gains are real if the implementation and measurements are as described. What is not supported is the stronger theoretical narrative that 3DLS renders actual projections of 3D linear kernels. Eq. (10) is asserted without derivation, and for non-Gaussian radial kernels the standard covariance-projection argument does not apply. The line-integral counterexample shows the rendered footprint is a different function, so the claim to have 'replaced Gaussian kernels with linear kernels' in 3D is, at best, a useful approximation or a re-definition of the 2D splat. This does not automatically invalidate the empirical comparisons, because the method is well-defined even if the projection identity fails; however, it means the paper's explanation of why the method improves high-frequency fidelity is not grounded in the stated 3D kernel geometry. The attribution is further weakened by the ablation design: the baseline and LK-only rows in Table 2 differ in loss weights and densification thresholds, so the kernel contribution is confounded. Neither issue alone disproves the central empirical claim, but together they justify a conditional acceptance pending code release and a controlled ablation. The reader's verdict of CONDITIONAL with moderate confidence remains appropriate; no verdict change is needed.","tokens_in":17369,"tokens_out":9219,"duration_ms":93965,"concrete_test":"Render a single splat with Sigma=I and camera along the z-axis; compare the alpha map from Eq. (10), L'(x')=max(0,1-|x'|), with the numerically ray-marched projection P(r)=2*integral_0^{sqrt(1-r^2)} (1-sqrt(r^2+z^2)) dz, normalized to the same peak. If the two profiles differ by more than a few percent at r=0.5, Eq. (10) is not the projection of the 3D linear kernel. Repeat at an oblique viewing angle using the full covariance transform to quantify the footprint mismatch. This single check settles whether the claimed 3D-to-2D projection identity is valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that 3DLS 'replaces Gaussian kernels with linear kernels' depends on Eq. (9), L(x)=max(0,1-D_M), and Eq. (10), L'(x')=max(0,1-D'_M). For Gaussian kernels, the 2D projection is exact because the marginal of a Gaussian is Gaussian and the covariance transforms as Sigma_2D = JW Sigma W^T J^T. For the linear kernel this is not exact: the line integral along the viewing direction of max(0,1-sqrt(r^2+z^2)) is not proportional to max(0,1-r). With Sigma=I and a camera along z, the projected profile is P(r)=2*integral_0^{sqrt(1-r^2)} (1-sqrt(r^2+z^2)) dz, which is visibly smoother than a linear ramp in r. Thus Eq. (10) is not a derived marginal or line integral; it is a new 2D kernel chosen by analogy. The bounded-support intuition in Fig. 1 still applies to this 2D kernel, so the empirical comparisons are not vacuous, but the paper's '3D linear ellipsoid' interpretation and the causal story that the 3D kernel's bounded support reduces artifacts rest on an unproven projection identity. Section B's generalization (Eqs. 19-25) lists attenuation functions but does not repair this gap. This is compounded by the ablation in Table 2: the baseline row and the LK row differ in loss weights and densification thresholds, so even the empirical attribution of gains to the kernel alone is not isolated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 3D Linear Splatting (3DLS), a modification of 3D Gaussian Splatting that replaces the Gaussian attenuation function with a linear (tent) kernel L(x)=max(0,1-D_M), together with a Distribution Alignment (DA) scale factor lambda and an Adaptive Gradient Scaling (AGS) heuristic. The claimed contributions are improved high-frequency detail and a roughly 30% rendering speedup. Experiments are reported on Mip-NeRF360, Tanks&Temples, and Deep Blending, using a custom CUDA implementation built on gsplat. The paper also includes an extended discussion of general kernel families and a supplementary with per-scene results and additional qualitative comparisons.","tokens_in":17681,"tokens_out":9591,"duration_ms":81091,"significance":"If the empirical results hold, the paper demonstrates a simple and computationally cheap change to the splatting kernel that yields competitive or better rendering quality on several benchmarks and a substantial speedup, which is practically useful. The measured FPS improvement is a concrete, falsifiable claim, and the custom CUDA implementation is a tangible contribution. However, the theoretical framing of the method as '3D linear ellipsoids' is unsupported because the 2D projection formula used in rendering is not the true projection of the 3D kernel, and the ablation does not isolate the kernel effect from threshold changes. These issues materially affect the paper's central claims and require revision.","major_comments":[{"comment":"Equation (10) asserts that the 2D projection of the linear kernel is L'(x') = max(0, 1 - D'_M) with the same covariance transform used for Gaussians. This is not a derived ray-marginal: for the 3D linear kernel with identity covariance, the line integral along the viewing direction is P(r) = 2 * integral_0^{sqrt(1-r^2)} (1 - sqrt(r^2 + z^2)) dz, which is not proportional to max(0, 1 - r). The covariance projection in Eq. (3) is exact for Gaussian kernels but does not generalize to arbitrary attenuation functions. Consequently, the rendered footprint is an ad hoc 2D kernel rather than the projection of a 3D linear ellipsoid, and the intuitive explanation in Figure 1 that the bounded support of the 3D kernel reduces artifacts is not supported. The authors should derive the correct projection or explicitly state that Eq. (10) defines a new screen-space kernel, and adjust the claims accordingly.","section":"Section 3.2, Eq. (10)"},{"comment":"The abstract's statement that 3DLS 'demonstrates state-of-the-art fidelity and accuracy' is contradicted by Table 1. On Deep Blending, 3DLS has lower SSIM (0.902 vs 0.905) and PSNR (29.44 vs 29.49) than the baseline 3DGS. On Mip-NeRF360, 3DLS attains an LPIPS of 0.196, worse than AbsGS's 0.186. The SOTA claim should be qualified to specific datasets and metrics, or removed.","section":"Abstract and Table 1"},{"comment":"The ablation study is confounded by the densification thresholds. Section 4.1 reports that 3DLS uses a growth threshold of 0.006, a 3D prune threshold of 0.4, and an opacity prune threshold of 0.025, whereas baseline 3DGS uses 0.01, 0.1, and 0.005. In Table 2, the baseline row uses the original thresholds while the LK row uses the modified thresholds, so the observed reduction in primitive count and the performance differences cannot be attributed to the linear kernel alone. The authors should either repeat the ablation with identical thresholds across rows or add a row that changes thresholds while retaining the Gaussian kernel.","section":"Section 4.1 and Table 2"},{"comment":"The distribution alignment factor lambda is not derived from a defined matching criterion. The paper says (Supplementary B) that 'empirical tests show that lambda = 2.5 aligns Cosine and Linear kernels to the Gaussian distribution,' but it never specifies what quantity is matched (variance, full-width at half-maximum, or some other measure). Since lambda directly determines the kernel's effective spread, the choice is a free parameter fitted to the baseline. The authors should define the alignment criterion and provide a sensitivity analysis over lambda to show that the reported improvements are not an artifact of this specific value.","section":"Section 3.3.1 and Supplementary B"}],"minor_comments":[{"comment":"The ablation does not include a row with LK + AGS (without DA), so the individual contribution of AGS is not measured; please add such a row or discuss the interaction between AGS and DA.","section":"Section 3.3.2 and Table 2"},{"comment":"The kernel comparison in Figure 5 is qualitative only; please report quantitative metrics (e.g., PSNR/SSIM) for the 2D pattern fitting to support the claim that the linear kernel 'excels' in high-frequency regions.","section":"Section 5.1 and Figure 5"},{"comment":"Several reference entries appear to have stray page numbers or formatting artifacts (e.g., [1] ends with '1' and [19] ends with '9'); please clean up the bibliography.","section":"References"},{"comment":"The reported lambda values for the quadratic (6) and Laplacian (1) kernels are given without explanation; please clarify how these values were obtained.","section":"Supplementary B"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper presents a simple kernel replacement with competitive results, but the projection formula in Eq. (10) is mathematically incorrect as a projection, which undermines the paper's theoretical framing. The SOTA claim and the ablation need to be corrected. The issues are fixable in revision, so I recommend major revision rather than rejection. The authors should also be encouraged to compare with more recent splatting variants in any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what the paper actually does. It replaces the Gaussian kernel in 3DGS with L(x)=max(0,1-D_M), adds Distribution Alignment (a scaling of the Mahalanobis distance) and Adaptive Gradient Scaling, and reports a 30% FPS gain with small PSNR/SSIM improvements on Mip-NeRF360 and Tanks&Temples. The supplementary's unified kernel formulation (Gaussian, Laplacian, cosine, quadratic, linear) is a useful framing, even if only the linear kernel is evaluated in the main paper.\n\nThe paper does several things well. The integration is clean, built on gsplat with custom CUDA. The per-scene tables are complete. The speedup is plausible and consistent. The qualitative examples, especially the Train scene ballast and background forests, are persuasive. The authors also acknowledge in the limitations that the densification thresholds are empirical.\n\nThe soft spots are real, though. The stress-test note is right: Eq. (10) is not the projection of the 3D linear kernel. The line integral of max(0,1-sqrt(r^2+z^2)) along the viewing axis is smoother than max(0,1-r). So the 'bounded support in 3D' story is not what the rendered 2D kernel actually implements. That is a significant gap in the theoretical framing, but it is fixable: the 2D linear kernel is still a valid splatting kernel, so the empirical results are not vacuous—just reinterpreted.\n\nThe ablation also undercuts attribution. The baseline row and the LK-only row differ in loss weights (L2 added, 6:2:2) and densification thresholds. So the 'efficiency gains' and quality from LK alone cannot be separated from those changes. DA's lambda=2.5 is explicitly fitted to match Gaussian spread, which is fine as a practical choice but weak as a principled component. And the abstract's 'state-of-the-art fidelity' is contradicted by their own Table 1: on Deep Blending 3DGS has higher PSNR, and AbsGS has lower LPIPS on Mip-NeRF360. They hedge in the body, but the abstract overstates.\n\nOverall the central empirical claim—a linear kernel can match or slightly beat 3DGS with a speedup—probably holds. The derivation gap and the ablation weakness are addressable in revision. The paper deserves a serious referee, though I'd expect a major revision request. Would I cite it in my own work? Probably not in the next year. Worth a reading group slot if someone in the group is tracking 3DGS variants.","headline":"Useful incremental 3DGS kernel swap with a real projection-derivation gap and an overclaimed 'SOTA' label; worth reviewing, not rejecting.","tokens_in":18303,"tokens_out":3340,"would_cite":false,"duration_ms":31608,"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":"Replacing Gaussian kernels in 3D splatting with bounded linear (tent-shaped) kernels yields sharper high-frequency detail and about 30% faster rendering than the baseline.","keywords":["3D Gaussian Splatting","linear kernels","novel view synthesis","high-frequency detail","kernel design","real-time rendering","radiance fields","3D Linear Splatting"],"falsifier":"Render one isolated 3D linear ellipsoid with known position, orientation, and covariance onto a camera, and compute the pixel intensities two ways: analytically with Eq. (10) and numerically by integrating the 3D tent kernel along each pixel ray. If the two images differ beyond numerical tolerance, the paper's projection step is not an exact description of the 3D kernel.","tokens_in":17100,"feed_emoji":"📐","tokens_out":8123,"duration_ms":70420,"temperature":0.7,"pith_summary":"3D Gaussian Splatting renders novel views by blending many soft ellipsoids, but the Gaussian's unbounded tail smears high-frequency detail and produces floating artifacts. 3DLS replaces the Gaussian falloff with a bounded linear (tent) kernel defined on the same Mahalanobis distance, so each splat has compact support and interferes less with its neighbors. The paper claims this simple kernel swap improves fidelity on three standard view-synthesis benchmarks, is particularly beneficial for sharp edges and fine texture, and raises rendering speed by about 30% over the baseline. If correct, it shows kernel shape, not just scene structure or density control, is a first-order design choice in splatting rendering.","feed_headline":"Swap Gaussian splats for tent shapes: sharper 3D views, 30% faster","feed_subtitle":"A bounded linear kernel cuts blur and floating artifacts in 3D Gaussian Splatting while speeding up rendering.","key_machinery":"The central object is the linear kernel $L(x)=\\max(0,1-D_M)$, a tent-shaped attenuation defined on the same Mahalanobis distance $D_M=\\sqrt{(x-\\mu)^\\top\\Sigma^{-1}(x-\\mu)}$ that 3DGS uses for its Gaussian. Its bounded support (the kernel is exactly zero outside the ellipsoid) is what reduces inter-splat blending and preserves sharp transitions. Distribution Alignment re-scales the distance as $D_M/\\lambda$ so the linear kernel's footprint covers as much of the scene as the Gaussian it replaces, and Adaptive Gradient Scaling $\\omega(D'_M)=\\exp(-D'_M/2)$ reweights gradients so far-from-center pixels do not cause unstable updates. Together they let the linear kernel plug into the standard splatting pipeline as a drop-in replacement for the Gaussian while keeping training stable.","core_discovery":"The paper's central claim is that the soft boundary of the Gaussian ellipsoid is itself a source of the blurring, floating primitives, and over-reconstruction seen in 3DGS, and that replacing $G(x)=\\exp(-\\tfrac12 D_M^2)$ with the linear kernel $L(x)=\\max(0,1-D_M)$, where $D_M$ is the Mahalanobis distance from the splat center, removes much of that artifact. The projected 2D kernel is taken to be $L'(x')=\\max(0,1-D'_M)$ using the projected covariance, so the existing $\\alpha$-blending rasterizer is unchanged. Two fixes make the new kernel trainable: Distribution Alignment scales $D_M$ by $\\lambda$ to match Gaussian coverage, and Adaptive Gradient Scaling multiplies parameter gradients by $\\omega(D'_M)=\\exp(-D'_M/2)$ to stop distant pixels from dominating updates. On three benchmark datasets the paper reports state-of-the-art or competitive SSIM, PSNR, and LPIPS, with forward and backward passes about 30% faster than baseline 3DGS and only modest memory overhead.","pith_inferences":["The 2D projection in Eq. (10) is assumed rather than derived; a natural test is to numerically integrate the 3D tent kernel along pixel rays for a single splat and compare the footprint with $L'(x')$. A mismatch would mean the rendered image is not the exact projection of the claimed 3D ellipsoid.","Because the linear kernel is bump-shaped but discontinuous in derivative at its boundary, its interaction with mip-style anti-aliasing or multi-scale filtering may differ from the Gaussian case; 3DLS+AA already hints these can be combined without losing the sharpness gain.","The kernel-comparison experiments suggest a hybrid scheme—linear kernels where edges and high-frequency texture dominate, cosine or quadratic where surfaces are smooth and uniform—could outperform any single kernel, but the paper leaves that as future work.","The speed advantage should be measured at matched primitive counts: the paper reports higher $N$ for 3DLS than baseline while still being faster, so the per-splat cost reduction is the real quantity to benchmark, not just end-to-end FPS."],"forward_implications":["A rendering system built for Gaussian splats can adopt linear kernels by changing only the attenuation function and the two training adjustments, leaving accumulation, blending, and rasterization order intact.","Scenes with fine texture, straight edges, and high-frequency content should show less blur and fewer floating primitives than the Gaussian baseline on the same pipeline.","Because each linear splat has compact support, per-pixel blending cost drops, which is the mechanism behind the reported ~30% forward and backward FPS gain.","The paper's generalized kernel framework treats Gaussian, Laplacian, raised-cosine, quadratic, and linear kernels as the same construction with different falloff functions, so kernel choice becomes a tunable design axis rather than a fixed assumption.","On smooth, low-frequency content the advantage narrows; the method is competitive rather than uniformly best, consistent with the paper's stated limitation."],"supporting_citations":[{"why":"Defines the baseline 3DGS pipeline and Gaussian kernel that 3DLS replaces and compares against.","marker":"[19]"},{"why":"Provides the Mip-NeRF360 benchmark dataset and a strong anti-aliased radiance-field baseline.","marker":"[1]"},{"why":"Supplies the 2D disk-splat geometric representation baseline that must be beaten on fidelity metrics.","marker":"[15]"},{"why":"Supplies the anti-aliasing technique combined with 3DLS to form 3DLS+AA and a comparison baseline.","marker":"[50]"},{"why":"Supplies the density-control baseline that is the strongest perceptual-metric competitor on two datasets.","marker":"[49]"},{"why":"Provides the Tanks and Temples benchmark with outdoor scenes that stress high-frequency reconstruction.","marker":"[20]"},{"why":"Provides the Deep Blending benchmark where the linear kernel's advantage is weakest but LPIPS is best.","marker":"[12]"}],"fun_headline_variants":["Linear kernels sharpen 3D splatting, 30% faster than Gaussians","Swap Gaussian splats for linear tents: sharper, 30% faster","Tent kernels beat Gaussian blur in 3D splatting","Linear splat kernels cut blur and floating artifacts in 3D views","3D splatting upgrade: linear kernels boost fidelity and FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on assuming that when a 3D tent-shaped ellipsoid is viewed by a camera, its 2D image is just the same tent shape measured by the projected ellipse distance; the paper states this rather than proving it, so if that projection is wrong the rendered image differs from the claimed 3D geometry.","fun_headline_variants_meta":{"raw":{"variants":["Linear kernels sharpen 3D splatting, 30% faster than Gaussians","Swap Gaussian splats for linear tents: sharper, 30% faster","Tent kernels beat Gaussian blur in 3D splatting","Linear splat kernels cut blur and floating artifacts in 3D views","3D splatting upgrade: linear kernels boost fidelity and FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000344,"raw_usage":{"total_tokens":1901,"prompt_tokens":966,"completion_tokens":935,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":837}},"tokens_in":582,"tokens_out":935,"duration_ms":7345,"temperature":1.0,"reasoning_tokens":837,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:30:56.615835+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render one isolated 3D linear ellipsoid with known position, orientation, and covariance onto a camera, and compute the pixel intensities two ways: analytically with Eq. (10) and numerically by integrating the 3D tent kernel along each pixel ray. If the two images differ beyond numerical tolerance, the paper's projection step is not an exact description of the 3D kernel.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Defines the baseline 3DGS pipeline and Gaussian kernel that 3DLS replaces and compares against."},{"cited_title":"Barron, Ben Mildenhall, Dor Verbin, Pratul P","cited_arxiv_id":null,"evidence_quote":"Provides the Mip-NeRF360 benchmark dataset and a strong anti-aliased radiance-field baseline."},{"cited_title":"2d gaussian splatting for geometrically ac- curate radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D disk-splat geometric representation baseline that must be beaten on fidelity metrics."},{"cited_title":"Mip-splatting: Alias-free 3d gaussian splat- ting","cited_arxiv_id":null,"evidence_quote":"Supplies the anti-aliasing technique combined with 3DLS to form 3DLS+AA and a comparison baseline."},{"cited_title":"Absgs: Recovering fine details in 3d gaussian splat- ting","cited_arxiv_id":null,"evidence_quote":"Supplies the density-control baseline that is the strongest perceptual-metric competitor on two datasets."},{"cited_title":"Tanks and temples: Benchmarking large-scale scene reconstruction","cited_arxiv_id":null,"evidence_quote":"Provides the Tanks and Temples benchmark with outdoor scenes that stress high-frequency reconstruction."},{"cited_title":"Deep blending for free-viewpoint image-based rendering","cited_arxiv_id":null,"evidence_quote":"Provides the Deep Blending benchmark where the linear kernel's advantage is weakest but LPIPS is best."}],"review_version":1}