{"id":"61ff8564-b886-461d-966f-41f776417227","arxiv_id":"2608.07760","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"XClipGS derives an exact closed-form half-space clipping factor for Gaussian splatting and uses clipped reference views to supervise the exposed interior, reducing cut-face leakage by about 40x.","lead":"XClipGS finds a formula that lets a real-time 3D scan viewer cut open a CT or MRI volume, rendered as Gaussian splatting, without jagged or leaking edges. It also trains the hidden interior with clipped reference views, so anatomy revealed by the cut is more accurate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 is correct under affine EWA, but the operative concern is the unquantified gap between affine rays and true perspective rays; the paper scopes this honestly, so the conditional verdict stands.","rationale":"The reader's weakest_assumption is exactly the affine projection model, and my stress-test converges on the same point. I independently re-derived the supplement's proof: the factorization I_clip = I_full * P(omega <= tau | u=p) is correct, the Schur complement gives s^2 as in Eq. 4, and the CDF argument is affine in delta, so Eq. 5 follows. There is no hidden algebraic error. The per-primitive normalization concern (J vs. learned opacity) is benign because the clip factor is a ratio and opacity can absorb the constant. I considered other candidate concerns: (a) single-seed experiments, (b) best-effort ClipGS reimplementation, (c) accuracy of Leak/CDE metrics, (d) center-depth compositing artifacts. These affect the empirical conclusions but not the central mathematical claim, and the reader already flagged (a) and (b) in the rationale. The affine-vs-perspective gap is the only concern that touches the central claim directly: the proposition is conditional on a model that the actual renderer only approximates, and the paper gives no error analysis for that approximation. Because the paper explicitly and correctly scopes its exactness statement, and because a concrete test can settle the practical impact, the appropriate verdict remains CONDITIONAL as the reader concluded; no change is warranted.","tokens_in":24031,"tokens_out":18943,"duration_ms":180540,"concrete_test":"On the same trained checkpoints and the same voxel-axis and arbitrary-normal cut-eval sets, rerender each scene with a perspective-accurate clipped Gaussian renderer that computes the exact ray integral of each half-space-restricted 3D Gaussian along the true perspective ray (e.g., by integrating in ray parameter space or adapting 3DGEER's model). Compare the perspective-accurate clipped images against both XClipGS and the reference renders; report band SSIM, CDE, and Leak. If the metric shift between XClipGS and the perspective-accurate clipper is below, say, 0.01 SSIM, the affine approximation is benign in this regime; if it is larger, the 'exact' operator is not exact for the actual geometry and the paper's claims would need rescoping. To amplify possible failure, repeat with 3x primitive scales and a 100-degree FOV.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem (Proposition 1, Eq. 5) is mathematically sound within the stated local affine model: the supplement's derivation properly conditions the jointly Gaussian screen/clip coordinates and yields the Gaussian-CDF factor. The load-bearing risk is not internal validity but the applicability of the exactness claim to the actual rendering geometry. Equation 1 linearizes the camera projection about each primitive center, so the 'viewing ray' used in the proof is the affine-approximation ray, not the true perspective ray through the pixel. For large primitives, near-field cameras, or wide field-of-view settings, these rays diverge, and Proposition 1 provides no bound on this divergence. The paper acknowledges that a perspective-ray formulation would require re-derived coefficients (Section 3.2, citing 3DGEER), but it never quantifies how much the affine approximation affects the reported cut-face metrics. The numerical validation in Supplement S2 checks the CUDA path against a float64 autograd implementation of the same affine formulas, so it cannot catch errors introduced by the affine assumption. Consequently, the headline 'Exact Half-Space Clipping' could be read as exactness in the rendered image, whereas the exactness is conditional on a linearization whose practical cost is unmeasured. This is the weakest link between the theoretical result and the paper's practical claims, and it is the assumption the reader's weakest_assumption identifies.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes XClipGS, a clipping operator for Gaussian-splatting volume rendering. Under the local affine projection model of EWA splatting (Eq. 1), the authors derive a closed-form per-pixel factor for a plane-truncated Gaussian: the clipped contribution equals the unclipped EWA footprint times Phi(k^T delta + h), where k and h are computed from the primitive covariance, plane normal/offset, and projection Jacobian (Eq. 5). Proposition 1 (proved in Supplement S1) states that this factor equals the viewing-ray integral of the half-space-restricted Gaussian density, up to the per-splat normalization shared with the unclipped footprint. The operator has no learned parameters and is differentiable. The paper couples this with clip-aware interior supervision using multi-distance views with varied plane axes and offsets (Eq. 6), and introduces cut-face metrics CDE, Leak, and a geometric diagnostic CErr3D. On eight CT and MRI volumes with held-out plane offsets and arbitrary-normal planes, the method reports the best average on all reported metrics, over 650 FPS, and reduced leakage relative to a ClipGS reimplementation, moment-matched (MM) and hard-cull (HC) surrogates, and RaRa.","tokens_in":24313,"tokens_out":16709,"duration_ms":166191,"significance":"If the claims hold, the paper provides the first closed-form, parameter-free per-pixel clipping rule for Gaussian splatting: an exact factorization of the half-space-restricted Gaussian ray integral within the affine EWA model, cleanly separating the render-time operator from interior supervision. The manuscript has notable strengths: the derivation is first-principles and the proof in S1 is careful; the CUDA forward and backward passes are validated against a float64 autograd implementation to roughly 1e-6 relative error (S2); the experimental protocol includes controlled operator swaps on fixed interiors, a supervision ablation, numerically held-out offsets, arbitrary-normal transfer without retraining, and metric-parameter sensitivity sweeps; and the zero-by-construction CErr3D diagnostic is honestly distinguished from image-space metrics. The reported margins are large (about 1.2 dB PSNR over the ClipGS reimplementation, 0.05 band SSIM, roughly 40x leakage reduction on the voxel-axis set), which supports practical relevance if the numbers hold.","major_comments":[{"comment":"The exactness claim is established and validated only within the affine EWA model of Eq. (1). Supplement S2's float64 validation compares the CUDA rasterizer against an autograd implementation of the same affine formulas, so it bounds implementation error but cannot bound the divergence between the affine-approximation rays and the true perspective rays of the reference renderer. Because the title and the practical reading of \"exact\" concern the rendered image, the manuscript should quantify this gap in the reported configuration: for example, measure the per-primitive screen-space residual of Eq. (1) over the projected Gaussian support at the training and test camera distances (the checkpoints are available), or compare the analytic operator's output against a perspective-ray integration of the same restricted Gaussian on a subset of cut-eval views, and report the effect on Leak and Sb. If the gap is non-negligible at the close-up distances used in Table S5 and at the 60-degree FOV, the abstract's \"exact\" framing should be tempered; if it is small, the numbers should be stated so readers can see the boundary of the claim.","section":"§3.1–§3.2, Prop. 1; Supp. S2"},{"comment":"All empirical claims rest on one training run per scan–method pair at seed 0 (Supplement S3, \"Optimization\"). Densification and plane sampling are stochastic, so run-to-run variance is unmeasured. The headline margins are large, but the per-volume PSNR leads (Table 1) and the cut-face gaps (Table 2) would be more convincing with at least one additional seed reported as mean and range for the main comparisons, or with an explicit argument for why the fixed-seed pipeline is representative.","section":"Supp. S3, Optimization"},{"comment":"The headline system-level comparison (33.56 versus 32.34 dB, \"highest PSNR on every volume\") is measured against a best-effort reimplementation of ClipGS from its published description, since no public code was available (Supp. S4). The controlled MM/HC ladder does not depend on this baseline, but the abstract's headline contrast does. I ask the authors to release the reimplementation code, or to state clearly which claims would survive a comparison with the authors' original implementation, and to keep the controlled same-backbone ablations as the primary evidence for the operator's effect.","section":"Supp. S4; Table 1"}],"minor_comments":[{"comment":"The statement that band PSNR \"differs by only hundredths and is omitted\" is surprising given the 0.05 average gap in band SSIM between Ours and ClipGS in Table 2; a supporting sentence or a small table in Supplement S5 would help readers understand why SSIM moves while PSNR does not.","section":"§4.1"},{"comment":"The grazing-view metric uses the least-squares line proxy l = pinv(P)^T [n; -tau] because a 3D plane projects nonlinearly under perspective; please report the pixel error of this proxy against the true projected plane for the 1.5-degree grazing cameras, so that the CDEg values can be read as metric noise rather than operator error.","section":"Supp. S5, near-edge line proxy"},{"comment":"The three compatibility fixes to RaRa's released kernel change baseline behavior for fully visible primitives and for chords with both endpoints invisible; documenting the effect of each fix (for example, reporting Leak and CDE for the unfixed kernel) would make the 16x leakage comparison fully auditable.","section":"Supp. S4, RaRa integration"},{"comment":"The ClipGS system baseline uses vanilla 3DGS without the Mip-Splatting filter while Ours/MM/HC use the filter; since filtering affects boundary appearance, this additional confound should be acknowledged in §4.2 alongside the representation difference.","section":"Supp. S3, Backbones and warm start"},{"comment":"The vector d_n = Sigma n / sigma_n is introduced only in the prose around Eq. (3); defining it inside the equation block would improve readability.","section":"§3.2, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a strong candidate for this venue if the authors address the EWA-to-perspective gap and the verification concerns. The derivation and the experimental scaffolding are unusually careful, and I do not see an internal inconsistency in Proposition 1 under the stated affine model; the residual risk is overclaiming in the title and abstract relative to a linearized projection whose practical cost is unmeasured, plus the single-seed protocol and the reimplementation status of the main baseline. Self-citations appear mainly through the N-DGS backbone and initialization works, which is appropriate. A focused diagnostic experiment on the affine approximation error (as suggested in the first major comment) would resolve the main uncertainty without redoing the benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The core math is right: under the affine EWA projection, the half-space-restricted Gaussian's ray integral factorizes into the ordinary footprint times a Gaussian CDF whose argument is affine in the pixel offset. Proposition 1 is proven carefully in the supplement, and the CUDA implementation is checked against a float64 autograd version to roughly 1e-6. That's real evidence. The paper is also well argued about why this matters—binary keep/drop and moment-matched re-Gaussianization have obvious failure modes at the cut face, and the controlled comparisons (HC, MM, Ours on a shared backbone, plus operator swaps on fixed interiors) cleanly separate the operator from the learned interior. The new cut-face metrics (CDE, Leak) are a genuine improvement over global PSNR, and the parameter sensitivity sweep shows the main conclusions are stable.\n\nSoft spots, in order of size. First, \"exact\" is exact for the affine-EWA ray, not the true perspective ray. The paper says this in Section 3.2 and cites 3DGEER, but it never quantifies how much the affine approximation bends the ray for large primitives or wide FOV. The float64 validation only checks the implementation of the same affine formulas, so it can't catch this. For medical volumes at moderate FOV it's probably fine, but a referee should ask for a sensitivity test—for example, compare against a perspective-ray integral on a few cases or report error versus primitive size. Second, every number comes from a single fixed run (seed 0). The per-volume consistency helps, but there's no estimate of variance; the +1.22 dB over the ClipGS reimplementation could be partly noise. Third, no code or data yet; the ClipGS baseline is a best-effort reimplementation, and RaRa required three compatibility fixes to the released kernel. None of this is fatal, and the paper is unusually transparent about the reimplementation and the fixes.\n\nWho it's for: anyone working on Gaussian-splatting rendering, especially interactive volume visualization. The operator is reusable as a differentiable primitive. I'd bring it to a reading group and cite it if I were working on clipping or EWA extensions. It deserves a serious referee—send it out, but expect a request for multi-seed numbers, code/data, and a direct test of the affine assumption.","headline":"A clean analytic clipping operator for EWA splatting, honestly scoped; the unquantified affine-vs-perspective gap and single-seed numbers are the main things to push on in review.","tokens_in":24841,"tokens_out":2511,"would_cite":true,"duration_ms":23678,"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":"This paper shows that clipping a Gaussian-splatted medical volume at an arbitrary plane reduces to a closed-form per-pixel Gaussian-CDF factor, making interactive cut views exact under the affine EWA model and free of learned clipping…","keywords":["Gaussian splatting","volume rendering","medical imaging","clipping plane","half-space clipping","EWA splatting","selection-normal distribution","interactive rendering"],"falsifier":"Take one Gaussian primitive that straddles a known plane, render it under the affine EWA model, and compare Equation (5) pixel by pixel with a dense numerical integration of the truncated 3D density along the affine viewing rays. Agreement to floating-point tolerance confirms the identity; any systematic disagreement would falsify the factorization. Repeating under a true perspective camera with large primitives should show divergence, which would confirm the stated affine scope rather than the claim's failure.","tokens_in":23855,"feed_emoji":"🩻","tokens_out":7402,"duration_ms":65370,"temperature":0.7,"pith_summary":"The paper claims that when a clipping plane cuts through a medical volume represented by Gaussian splats, the ray integral of a half-space-restricted Gaussian factorizes exactly into the ordinary 2D footprint times a conditional Gaussian CDF whose argument is affine in pixel coordinates. This closed-form per-pixel clip operator is differentiable and adds no learned parameters, so a viewer can move a clip plane through a CT or MRI volume in real time while the exposed interior is learned from clipped reference views. The paper supports the claim with a derivation under the local affine projection model of EWA splatting, and with experiments on eight CT and MRI volumes showing the highest PSNR and the least culled-side leakage among the compared methods.","feed_headline":"Exact plane clipping for Gaussian splats tops 650 FPS","feed_subtitle":"A per-pixel CDF factor cuts Gaussians exactly, beating prior clipping on eight medical volumes without retraining.","key_machinery":"The core object is the conditional law of the clip coordinate $\\omega = n^\\top x$ given the pixel position under affine projection. Its conditional variance is the Schur complement $s^2 = \\sigma_n^2 - b^\\top A^{-1}b$, and its conditional mean produces the affine CDF argument $\\ell(\\delta) = k^\\top \\delta + h$. The per-pixel factor $\\alpha K_G(\\delta;A)\\Phi(\\ell(\\delta))$ is the selection-normal form for the truncated density, and it reduces clipping to three transient per-primitive coefficients plus one CDF evaluation per covered sample, while remaining differentiable with respect to both the primitive parameters and the clipping plane.","core_discovery":"The paper's central claim is that, under the local affine projection model of EWA splatting, the per-pixel contribution of a Gaussian primitive restricted to a half-space remains closed form: it is the ordinary unclipped footprint $\\alpha K_G(\\delta; A)$ times a conditional Gaussian CDF $\\Phi(k^\\top \\delta + h)$, with $k$ and $h$ affine in the pixel offset $\\delta$. Proposition 1 states that this expression exactly equals the viewing-ray integral of the unnormalized half-space-restricted Gaussian density. The paired screen position and clip coordinate are jointly Gaussian under the affine model, so conditioning on the pixel factorizes the truncated ray integral; the CDF then locates the cut inside each splat rather than attenuating the whole primitive uniformly.","pith_inferences":["We infer that the same Gaussian-CDF factor could apply to any half-space restriction of a Gaussian splat, such as editable segmentations or multi-plane dissection, since the derivation depends only on the affine projection and the plane equation.","We infer that the operator's behavior as the plane approaches the viewing-ray direction (the conditional-variance limit $s \\to 0$) gives it a natural advantage in grazing-angle cuts; a targeted study comparing numerical integration at grazing angles could verify this directly.","We infer that the paired clipped/unclipped evaluation with difference-referenced error could be reused as a standard protocol for clip operators, separating removal accuracy from reconstruction error in any volume renderer."],"forward_implications":["A medical volume proxy trained once with clip-aware views can be cut at any plane axis or orientation at render time, with no plane-conditioned network and no retraining.","The analytic operator makes the exposed cut face differentiable, so interior supervision reduces to ordinary photometric training; removing the clipped training views costs 8.3 dB on held-out clipped views.","Because the clipped contribution is the unclipped footprint times a CDF factor, render cost stays near hard-cull speed: the paper measures over 650 FPS at 1600x1600 resolution.","Localized paired clipped/unclipped metrics (band SSIM, difference-referenced cut error, culled-side leakage) expose cut-face quality that global PSNR and SSIM dilute, and the method leads all four on both voxel-axis and arbitrary-normal planes."],"supporting_citations":[{"why":"Supplies the local affine EWA projection model, equation (1), whose joint Gaussian assumption is the basis for the half-space integral's factorization.","marker":"[Zwicker et al. 2001]"},{"why":"Provides the 3D Gaussian splatting representation and the per-pixel alpha-compositing framework into which the clip operator is inserted.","marker":"[Kerbl et al. 2023]"},{"why":"Establishes the selection-normal distribution class whose Gaussian-CDF product form Proposition 1 specializes to half-space ray integrals.","marker":"[Arellano-Valle, Branco, and Genton 2006]"},{"why":"Boundary reference for the exactness claim: an exact perspective-ray formulation would require re-derived conditional coefficients, so the paper scopes Proposition 1 to affine EWA.","marker":"[Huang et al. 2026]"},{"why":"Primary medical-splatting baseline whose trainable binary keep/drop clip operator is compared as the system-level alternative.","marker":"[Li et al. 2025a]"},{"why":"Supplies the ray-ellipsoid chord-ratio clip operator used in the fixed-interior operator swap comparison.","marker":"[Li et al. 2025b]"}],"fun_headline_variants":["Exact plane-clip for Gaussians: CDF factor, 650 FPS","No retraining, exact clipping: Gaussian CDF factor for medical splats","CDF factor clips Gaussians exactly, beats prior on all volumes","Half-space clipping made exact: per-pixel CDF, no extra network","Gaussian splatting clip solved: exact cut, 40x less leak"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that projection is locally affine, so screen position and clip coordinate are jointly Gaussian; with true perspective projection and large primitives this is an approximation, and Proposition 1 inherits that approximation.","fun_headline_variants_meta":{"raw":{"variants":["Exact plane-clip for Gaussians: CDF factor, 650 FPS","No retraining, exact clipping: Gaussian CDF factor for medical splats","CDF factor clips Gaussians exactly, beats prior on all volumes","Half-space clipping made exact: per-pixel CDF, no extra network","Gaussian splatting clip solved: exact cut, 40x less leak"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000829,"raw_usage":{"total_tokens":3669,"prompt_tokens":1038,"completion_tokens":2631,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":2529}},"tokens_in":654,"tokens_out":2631,"duration_ms":18314,"temperature":1.0,"reasoning_tokens":2529,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:19:10.533229+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one Gaussian primitive that straddles a known plane, render it under the affine EWA model, and compare Equation (5) pixel by pixel with a dense numerical integration of the truncated 3D density along the affine viewing rays. Agreement to floating-point tolerance confirms the identity; any systematic disagreement would falsify the factorization. Repeating under a true perspective camera with large primitives should show divergence, which would confirm the stated affine scope rather than the claim's failure.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the local affine EWA projection model, equation (1), whose joint Gaussian assumption is the basis for the half-space integral's factorization."},{"cited_title":"B.; Branco, M","cited_arxiv_id":null,"evidence_quote":"Establishes the selection-normal distribution class whose Gaussian-CDF product form Proposition 1 specializes to half-space ray integrals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Boundary reference for the exactness claim: an exact perspective-ray formulation would require re-derived conditional coefficients, so the paper scopes Proposition 1 to affine EWA."}],"review_version":1}