{"id":"5bb9ccc2-937c-45e5-b74e-6bcebff8d01c","arxiv_id":"2411.14974","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"3D Convex Splatting replaces Gaussian splats with smooth convex primitives, achieving higher PSNR and LPIPS than 3DGS on Tanks and Temples and Deep Blending while using fewer primitives.","lead":"A team from Liège, KAUST, and Oxford introduces 3D Convex Splatting, a way to render 3D scenes from photos using flexible convex shapes instead of the fuzzy Gaussian blobs used by current state-of-the-art methods. On several standard benchmarks it reports sharper images with less memory, especially for indoor scenes with hard edges and flat surfaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The '3D' convex is rendered as a 2D convex-hull mask (Eqs. 7-8), not as a volume rendering of the 3D shape; the central 'geometrically-meaningful 3D primitive' claim rests on an unverified equivalence.","rationale":"The reader's weakest assumption correctly identifies the unverified link between the 3D smooth convex and the 2D rendering. I agree that this is the most load-bearing concern: the paper's novelty, title, and claims of geometric meaningfulness depend on the primitives being 3D convex volumes, yet the renderer never evaluates the 3D shape. The comparison to 3DGS is framed as a comparison of 3D primitives; if the convex splat is actually a 2D mask, the comparison is like 2DGS with a different kernel, and the reported advantages (Table 1, Fig. 10) are kernel effects, not geometric ones. The lack of a derivation akin to 3DGS's projected covariance is a real gap. I also note the abstract's 'superior performance' is overstated on Mip-NeRF360 SSIM (0.802 vs 0.815), but that is a claim-strength issue, not a correctness issue; the deeper problem is the mischaracterization of the primitive. A ray-marching check would settle whether the 3D description is faithful; if not, the paper should be reframed as a 2D convex-splatting method and the 3D volume claims removed. Since the empirical results may still hold, the conditional verdict is appropriate, but the required revision is more than cosmetic.","tokens_in":16728,"tokens_out":14183,"duration_ms":132956,"concrete_test":"Ray-march the 3D smooth-convex density I(p) (Eqs. 2-3) along camera rays through a single optimized convex, and compare to the paper's 2D-hull renderer (Eqs. 7-8) for the same primitive across viewpoints, especially grazing angles. Optimize δ and σ to match at one view, then evaluate at rotated views; if the error grows substantially or cannot be matched, the 2D proxy is view-inconsistent and the '3D convex' interpretation is false. Report the per-pixel relative error and the silhouette error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. 2.2, the authors define a 3D convex by a point set and a 3D signed-distance function (Eqs. 1-3), but the actual rendering never uses the 3D function. Instead, the 3D points are projected to 2D (Eq. 4), the 2D convex hull is computed, and a 2D indicator is constructed from line distances (Eqs. 7-8 in the supplement). Unlike 3DGS, where the 2D Gaussian is the exact ray integral of the 3D Gaussian, no derivation shows that the 2D LogSumExp/sigmoid over line distances equals the integral of the 3D smooth-convex density along a ray. The depth-scaling of δ and σ (Table 4), given without a derivation, is an ad hoc correction for this missing physical model. At oblique views, the 2D mask's interior profile is a function of 2D signed distance to the polygon, not of chord length through the 3D volume; a near-degenerate projection (e.g., viewing a thin convex edge-on) produces a 2D hull with very different soft-edge behavior than the 3D shape. Consequently, the primitive is effectively a 2D splat with a convex kernel, and the paper's central claims that 3DCS uses '3D smooth convexes' and yields 'geometrically-meaningful' or 'physically meaningful' 3D representations are unsupported. The reported quality gains over 3DGS could derive from the change in kernel shape, not from any 3D volumetric advantage, undercutting the novelty and the interpretation of Table 1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 3D Convex Splatting (3DCS), a primitive-based radiance field renderer that replaces 3D Gaussians with smooth convex shapes. The method defines a 3D convex by a point set, projects the points to the camera plane, computes a 2D convex hull, and evaluates a smooth 2D indicator function that is composited with alpha blending. The authors report improved PSNR/LPIPS over 3DGS on Tanks and Temples and Deep Blending, comparable results on Mip-NeRF360, reduced primitive counts, and real-time rendering via a custom CUDA rasterizer. The paper also includes synthetic shape experiments, ablations of the number of points per convex, densification strategies, and perspective scaling.","tokens_in":17153,"tokens_out":3805,"duration_ms":37531,"significance":"If the central claims held, 3DCS would be a meaningful step toward primitive-based radiance fields with more geometrically expressive primitives than Gaussians. The empirical results on Tanks and Temples and Deep Blending are promising, and the paper provides useful engineering contributions: a differentiable CUDA rasterizer, an adaptive split-into-K densification scheme, and a parameter-efficiency analysis. The ablations for K, densification, and perspective scaling are valuable. However, the core conceptual claim that the method renders 3D smooth convexes is not supported by the actual rendering equations, which operate on a 2D convex hull of projected points. As a result, the claimed geometric interpretability and physical meaningfulness of the representation are not established. The paper would be significantly stronger if it were reframed as a 2D convex-kernel splatting method or if a genuine equivalence to 3D volume rendering were derived.","major_comments":[{"comment":"The rendering pipeline never evaluates the 3D signed-distance function of Eqs. (2)-(3); after projecting the K points to 2D, it computes a 2D convex hull and evaluates a 2D LogSumExp/sigmoid over 2D line distances. No derivation shows that this 2D indicator equals the integral of the 3D smooth-convex density along a camera ray, in contrast to 3DGS, where the 2D Gaussian is the exact ray integral of the 3D Gaussian. The distance scaling of δ and σ in Eqs. (7)-(8), whose sensitivity is shown in Table 4, is a heuristic correction rather than a consequence of perspective projection. The abstract's claim of modeling '3D smooth convexes' and the 'geometrically-meaningful 3D representation' language in Sec. 3.4 are therefore not supported by the method as defined; the primitive is, as implemented, a 2D convex splat. The authors should either derive the relation to 3D volume rendering or re-state the contribution as 2D convex-kernel splatting and soften the 3D-interpretability claims.","section":"Sec. 2.2 and Supplementary Eqs. (7)-(8)"},{"comment":"The best-performing 3DCS model uses hyperparameters fine-tuned separately for indoor and outdoor scenes, including different densification thresholds, different split scaling factors (0.7 indoor vs. 0.6 outdoor), and different treatment of σ after splitting. This makes the headline comparison against 3DGS and other baselines a comparison that includes per-scene-type model selection, rather than a single method with fixed hyperparameters. Moreover, no error bars or multiple-run statistics are reported. Given these issues, the claim in Table 1 of consistent superiority over 3DGS on Mip-NeRF360 is not robust: the SSIM is lower (0.802 vs. 0.815) and the PSNR gain is only 0.08 dB. The paper should present the unified-hyperparameter light model as the primary comparison and report variance or at least acknowledge the model-selection issue explicitly.","section":"Sec. 3.2 and Supplementary Sec. 5"},{"comment":"The per-scene results for Mip-NeRF360 show that 3DCS underperforms 3DGS on several scenes, including Bicycle (PSNR 24.72 vs. 25.24, SSIM 0.737 vs. 0.771) and Treehill (PSNR 21.77 vs. 22.49, SSIM 0.595 vs. 0.638), in addition to the aggregate SSIM deficit. The statement in Sec. 3.3 that 3DCS 'consistently matches or surpasses existing methods' is therefore too strong. The scene-level variability should be discussed, especially since the indoor/outdoor split in Table 2 is exactly where the method's advantage concentrates.","section":"Tables 8-10 (Supplementary)"},{"comment":"The claim that 3DCS yields 'physically meaningful 3D representations' and decomposes objects into meaningful convex shapes is supported only by qualitative examples with a small number of primitives. No evidence is provided that the inferred convexes are consistent across views, that they correspond to actual scene surfaces, or that the decomposition is semantically meaningful. Since the rendered quantity is a 2D convex-hull indicator, the view-dependent masks do not by themselves establish a 3D volumetric decomposition. This section should be substantially toned down or backed by quantitative multi-view consistency or geometry metrics.","section":"Sec. 3.4 and Fig. 11"}],"minor_comments":[{"comment":"The opacity factor in Eq. (5) is not clearly defined in the text: the sentence refers to 'on the opacity' but does not state that the symbol o_n denotes opacity, nor is the product term typeset unambiguously.","section":"Eq. (5)"},{"comment":"The text says '3DCS light outperforms 3DGS and GES on the T&T and DP dataset'; 'DP' should be 'DB' (Deep Blending).","section":"Sec. 3.3"},{"comment":"The title of [14] contains a typo: 'algorith' should be 'algorithm'.","section":"Reference [14]"},{"comment":"The sentence 'by representing scenes with millions of 3D Gaussian' should read 'millions of 3D Gaussians'.","section":"Sec. 1"}],"recommendation":"major_revision","confidential_remarks":"The central concern is not that the method fails empirically, but that the paper's framing as a 3D primitive method is not supported by the rendering equations. This is fixable by reframing the contribution as 2D convex splatting, or by deriving a true ray-integral equivalence, and by making the hyperparameter tuning transparent. Given the promising results on T&T and DB and the well-engineered rasterizer, I do not think rejection is warranted, but the required changes affect the abstract, methodology, and interpretation, so major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you track novel-view synthesis. The paper replaces 3D Gaussians with smooth convexes defined as hulls of point clouds, and shows real quality gains on Tanks & Temples and Deep Blending, plus indoor Mip-NeRF360 scenes, with fewer primitives and lower memory. The CUDA rasterizer is a genuine engineering contribution, and the ablations on K and densification are honest. The toy experiments showing convexes fitting hard edges with few primitives are convincing.\n\nThe soft spot is the one the stress-test flagged, and it lands. The rendering never uses the 3D signed-distance function from Sec. 2.1. It projects the 3D points, computes a 2D convex hull, and evaluates a 2D LogSumExp/sigmoid over line distances. That gives the correct silhouette, but the interior softness and edge profile are a function of 2D distance to the polygon, not of any ray integral through a 3D volume. The depth-scaling of delta and sigma is an ad hoc patch, not a derivation. So the claims about '3D smooth convexes' and 'physically meaningful 3D representations' are overstated. What you actually have is a 2D convex-kernel splat whose parameters live in 3D. That is still a legitimate primitive, and the empirical gains may well come from the kernel shape, but the paper does not show that the 3D interpretation earns its keep.\n\nThe evaluation has other weaknesses: no error bars, the best model uses different hyperparameters for indoor and outdoor scenes, and on Mip-NeRF360 overall SSIM is worse than 3DGS. The abstract says 'superior performance' but on that benchmark the method only matches or slightly beats 3DGS on aggregate; the real wins are on structured scenes. No code release either.\n\nNone of this kills the paper. The central idea is plausible, the results on T&T and indoor scenes are meaningful, and the primitive is new. But the authors should either derive the 2D projection from the 3D volume rendering or drop the '3D volumetric' framing and present it as a convex-silhouette splat. I would send this to review, with a referee who can pressure them on the math and the hyperparameter story. Worth a reading-group slot too, since the gap between title and math is a good teaching case.","headline":"New primitive for splatting that works well empirically, but the '3D' in the title is doing more work than the math supports.","tokens_in":17700,"tokens_out":2334,"would_cite":true,"duration_ms":22656,"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":"Smooth convexes replace Gaussian blobs for sharper 3D views","keywords":["novel view synthesis","3D Gaussian Splatting","smooth convexes","differentiable rendering","radiance fields","convex hull rasterization","real-time rendering","3D scene reconstruction"],"falsifier":"Render a synthetic scene of a single tilted rectangular slab both with 3DCS and with a ray-marched ground truth of the same 3D smooth convex; if the predicted images diverge substantially at grazing camera angles, or if pulling the projected points nearly collinear changes the image far more than the true 3D shape would, the 2D proxy is falsified.","tokens_in":16548,"feed_emoji":"🧊","tokens_out":6677,"duration_ms":54877,"temperature":0.7,"pith_summary":"This paper argues that the standard primitive of real-time radiance field rendering, the 3D Gaussian, is the wrong shape for many scenes, and replaces it with a 3D smooth convex, a point-set convex hull with soft edges. The authors claim that smooth convexes capture hard edges, flat surfaces, and dense volumes more faithfully than Gaussians, so a scene needs fewer primitives and less memory for the same or better image quality. On the Mip-NeRF360, Tanks and Temples, and Deep Blending benchmarks, the method reports gains over 3D Gaussian Splatting of up to 0.81 dB PSNR and 0.026 LPIPS, while rendering in real time. The practical stake is that a primitive with an explicit geometric boundary could give sharp, editable, compact 3D reconstructions instead of a cloud of soft blobs.","feed_headline":"Smooth convexes replace Gaussian blobs for sharper 3D views","feed_subtitle":"Convex primitives beat Gaussian blobs: up to +0.81 PSNR, ~30% less memory","key_machinery":"The load-bearing object is the 3D smooth convex primitive: a convex hull of $K$ 3D points whose boundary is softened by a log-sum-exp signed-distance function with smoothness $\\delta$ and sharpness $\\sigma$. The rendering machinery skips an explicit 3D hull: it projects the $K$ points to 2D, builds the 2D convex hull with a standard planar hull algorithm, evaluates the same smooth indicator per pixel on the hull's delimiting lines, scales $\\delta$ and $\\sigma$ by the depth $d$, and composites the resulting opacities by depth-ordered $\\alpha$-blending. All steps, including point projection, hull construction, line-distance evaluation, and backpropagation, run in custom GPU kernels, which is what keeps rendering real-time.","core_discovery":"3D Convex Splatting (3DCS) treats a scene as a collection of 3D smooth convexes, each defined by a set of $K$ freely moving 3D points. To render one convex, the points are projected onto the camera plane, their 2D convex hull is computed, and the per-pixel opacity is set by a smooth indicator function $I(q)=\\mathrm{Sigmoid}(-d\\,\\sigma\\,\\phi(q))$, where $\\phi(q)=\\log\\sum_t \\exp(d\\,\\delta\\, L_t(q))$ accumulates signed distances $L_t(q)$ to the hull's delimiting lines, with smoothness $\\delta$, sharpness $\\sigma$, and depth scale $d$ inherited from the underlying 3D shape. The authors claim that this differentiable splatting, together with a densification step that splits each convex into $K$ smaller copies when its sharpness loss is high, lets convexes represent hard-edged and flat geometry more compactly than Gaussians. They report consistent quality gains over 3DGS on structured indoor and human-made scenes, with the largest margin on Tanks and Temples, and a roughly 30% memory reduction in the full model.","pith_inferences":["A natural extension the authors do not develop is surface extraction: because convexes have explicit hull boundaries, one could read off depths, normals, or meshes directly from the primitives, something Gaussian clouds make ill-posed.","The depth-scaling of $\\delta$ and $\\sigma$ is introduced empirically; a tighter derivation from perspective projection might predict when the 2D proxy breaks and suggest a corrective term for grazing views.","The split-into-$K$ densification rule is an unusual target for future work: coupling it to a learned importance signal could further cut primitive counts without retraining the rasterizer.","Since convexes are closed shapes, they invite semantic or object-level decomposition, potentially turning a renderer into a parseable scene representation."],"forward_implications":["On structured indoor scenes, 3DCS reports gains of 0.9 PSNR, 0.007 SSIM, and 0.023 LPIPS over 3DGS, indicating the primitive's advantage is largest where walls, edges, and furniture dominate.","The full model uses roughly 70% of 3DGS memory on the same benchmark while reporting better or equal quality; the lightweight variant uses under 15% of 3DGS memory and still beats it on Tanks and Temples and Deep Blending.","Because each primitive has a geometric boundary, the method produces visibly sharper renderings in qualitative comparisons, with the paper attributing PSNR's tendency to favor blur to pixel-level sensitivity.","Training remains practical at roughly one hour for the full model, versus 48 hours for the NeRF baseline, while rendering stays in real time."],"supporting_citations":[{"why":"Defines 3D Gaussian Splatting, the baseline primitive method and rendering pipeline that 3DCS extends and compares against.","marker":"[25]"},{"why":"Supplies the smooth-convex signed-distance and indicator formulation that 3DCS reuses in 2D.","marker":"[9]"},{"why":"Provides the Mip-NeRF360 benchmark dataset and the NeRF baseline used for comparison.","marker":"[3]"},{"why":"Provides the convex hull algorithm used to build the 2D projection of each primitive.","marker":"[14]"},{"why":"Provides the Tanks and Temples dataset used to evaluate 3DCS.","marker":"[27]"},{"why":"Provides the Deep Blending dataset used to evaluate 3DCS.","marker":"[19]"},{"why":"Defines 2D Gaussian splatting, a planar-primitive baseline that 3DCS is compared with.","marker":"[21]"},{"why":"Defines Generalized Exponential Splatting, a primitive baseline that 3DCS compares against.","marker":"[18]"},{"why":"Supplies the regularization and mask loss used in 3DCS training.","marker":"[29]"}],"fun_headline_variants":["Convex splatting beats Gaussian splatting on sharp edges and flat surfaces","3D Convex Splatting: sharper rendering with fewer primitives","Convex primitives cut memory 30% and boost PSNR by 0.81","Smooth convexes render hard edges crisper than Gaussian blobs","3DCS: convex-based radiance fields outperform 3DGS in fidelity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that a 2D soft projection of a convex hull is a faithful stand-in for the true 3D volume's appearance; unlike a 3D Gaussian, whose 2D projection is the exact ray integral, no such equivalence is established for convexes, so the proxy could diverge at oblique angles or when projected points become nearly collinear.","fun_headline_variants_meta":{"raw":{"variants":["Convex splatting beats Gaussian splatting on sharp edges and flat surfaces","3D Convex Splatting: sharper rendering with fewer primitives","Convex primitives cut memory 30% and boost PSNR by 0.81","Smooth convexes render hard edges crisper than Gaussian blobs","3DCS: convex-based radiance fields outperform 3DGS in fidelity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000519,"raw_usage":{"total_tokens":2591,"prompt_tokens":1099,"completion_tokens":1492,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":715,"completion_tokens_details":{"reasoning_tokens":1402}},"tokens_in":715,"tokens_out":1492,"duration_ms":9943,"temperature":1.0,"reasoning_tokens":1402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:39:50.164210+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic scene of a single tilted rectangular slab both with 3DCS and with a ray-marched ground truth of the same 3D smooth convex; if the predicted images diverge substantially at grazing camera angles, or if pulling the projected points nearly collinear changes the image far more than the true 3D shape would, the 2D proxy is falsified.","supporting_citations":[{"cited_title":"3D Gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Defines 3D Gaussian Splatting, the baseline primitive method and rendering pipeline that 3DCS extends and compares against."},{"cited_title":"CvxNet: Learnable convex decomposition","cited_arxiv_id":null,"evidence_quote":"Supplies the smooth-convex signed-distance and indicator formulation that 3DCS reuses in 2D."},{"cited_title":"Barron, Ben Mildenhall, Dor Verbin, Pratul P","cited_arxiv_id":null,"evidence_quote":"Provides the Mip-NeRF360 benchmark dataset and the NeRF baseline used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the convex hull algorithm used to build the 2D projection of each primitive."},{"cited_title":"Tanks and temples: benchmarking large-scale scene reconstruction","cited_arxiv_id":null,"evidence_quote":"Provides the Tanks and Temples dataset used to evaluate 3DCS."},{"cited_title":"Deep blending for free-viewpoint image-based rendering","cited_arxiv_id":null,"evidence_quote":"Provides the Deep Blending dataset used to evaluate 3DCS."},{"cited_title":"2D Gaussian splatting for geometrically ac- curate radiance fields","cited_arxiv_id":null,"evidence_quote":"Defines 2D Gaussian splatting, a planar-primitive baseline that 3DCS is compared with."},{"cited_title":"GES: Generalized exponential splatting for ef- ficient radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Defines Generalized Exponential Splatting, a primitive baseline that 3DCS compares against."},{"cited_title":"Compact 3D Gaussian representation for radiance field","cited_arxiv_id":null,"evidence_quote":"Supplies the regularization and mask loss used in 3DCS training."}],"review_version":1}