{"id":"623518aa-ad9e-4baa-96a4-6e611cdb946a","arxiv_id":"2412.00333","paper_version":3,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A 4D Gaussian splatting method that adds a Kalman-like state consistency filter and Wasserstein log/exp geometry reports quality gains on dynamic scene benchmarks, but internal equation errors and unverified training-time claims weaken the result.","lead":"Dynamic 3D scene renderers are given smoother motion by fusing predicted and observed Gaussian positions with a Wasserstein-geometry update rule. The proposed 4D Gaussian splatting variant reports higher PSNR on D-NeRF and Plenoptic benchmarks, though derivation errors and missing code currently undercut the claims.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6) fuses Gaussian means and covariances with one asymmetric 3x3 gain, producing ill-defined and possibly non-SPD covariances; the core filter is undefined.","rationale":"The paper's central contribution is a state-space filter that merges a predicted Gaussian with an observed Gaussian and then renders the result. Every timestep therefore depends on Eq. (6). The equation is not a valid operation on the Gaussian state: the same matrix K multiplies both the mean and covariance while the state is a heterogeneous object (a 3-vector plus an SPD matrix); the covariance update is not guaranteed to remain SPD; and no repair is described. This is not a disagreement with current consensus or a stylistic choice; it is an internal algebraic inconsistency. The reader's weakest assumption identified this same spot. Additional inconsistencies, such as the conflict between Eq. (12) and Algorithm 2, the unexplained 'specified metric offsets' caption in Table 3, and the implausible training-time reductions, strengthen the reject conclusion but are not needed for the core objection. The proposed test is a deterministic numerical check on random SPD pairs; if the update fails even occasionally, the method as written cannot be the source of the reported gains. No code is provided, so this check is the minimal path to verification. The verdict should remain REJECT; the rejection is based on the method being undefined, not on disagreement with the community or on the authors' intent.","tokens_in":16959,"tokens_out":6759,"duration_ms":59767,"concrete_test":"Implement Eq. (6) as written on random pairs of SPD matrices: draw Sigma^Ob, Sigma^P from the set used in 3DGS (e.g., R S S^T R^T with positive scales), compute K and hat Sigma = Sigma^Ob + K(Sigma^P - Sigma^Ob), then test symmetry (norm(hat Sigma - hat Sigma^T)) and positive-definiteness (minimum eigenvalue). Repeat over 10^4 pairs with non-commuting covariances. If any hat Sigma fails the SPD test, Eq. (6) is invalid as a fusion rule. Additionally, compare the mean update with the classical BLUE covariance ((Sigma^Ob)^{-1} + (Sigma^P)^{-1})^{-1}; if the rendered quality depends on hat Sigma, the method must specify the correct update. The deterministic test requires no training data and settles the algebraic premise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the State Consistency Filter to produce valid 3D Gaussians at every timestep. Eq. (6) merges the observed and predicted Gaussians with K = Sigma^Ob_t (Sigma^Ob_t + Sigma^P_t)^{-1}, and feeds the result directly to the rasterizer. This is not a valid fusion rule. First, N contains a 3-D mean and a 3x3 covariance, while K is a single 3x3 matrix; the expression K(N^P_t - N^Ob_t) has no well-defined meaning unless K is block-diagonal with distinct blocks for mu and Sigma, which is never specified. Second, applying the same K to the covariance gives hat Sigma_t = Sigma^Ob_t + K(Sigma^P_t - Sigma^Ob_t); this linear combination is not guaranteed symmetric or positive-definite when Sigma^Ob_t and Sigma^P_t do not commute, and no symmetrization or SPD projection is given. Since 3DGS requires Sigma = R S S^T R^T for splatting, an invalid hat Sigma makes rendering undefined. The correct Kalman-style merge of two Gaussian estimates of a common mean would update the mean with this K but set the covariance to ((Sigma^Ob_t)^{-1} + (Sigma^P_t)^{-1})^{-1}, not a weighted sum. The paper's Eq. (6) conflates state value and state uncertainty, so the filter is undefined even before considering the additional inconsistency between Eq. (12) and Algorithm 2 (Delta Sigma_t = 2Sigma_{t-1} - P - P^T versus P + P^T - 2Sigma_t). The reported 34.45 dB PSNR and 2.0 dB ablation gains cannot be attributed to a component whose update rule is not a valid operation on SPD matrices.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 4D Gaussian Splatting framework with three components: a State Consistency Filter that merges predicted and observed Gaussian states via a Kalman-like update, a Wasserstein distance regularization term, and Wasserstein log/exp maps for dynamic prediction. The method is evaluated on the D-NeRF and Plenoptic Video datasets, reporting state-of-the-art PSNR/SSIM/LPIPS figures (34.45 dB on D-NeRF and 32.79 dB on Plenoptic) together with training-time reductions. The central claim is that combining state-space modeling with Wasserstein geometry yields more accurate, temporally coherent, and physically plausible dynamic scene rendering.","tokens_in":17216,"tokens_out":9494,"duration_ms":77202,"significance":"If the proposed components were correctly defined and the empirical results were reproducible, the paper would be a useful contribution to dynamic scene rendering, since it integrates distributional geometry with 4D Gaussian Splatting and provides extensive comparisons and ablations on two benchmarks. The paper deserves credit for its broad experimental coverage, per-scene tables, and explicit attempts to ablate each component. However, the mathematical core of the method contains load-bearing errors: the State Consistency Filter is not a well-defined operation on Gaussian distributions, and the Wasserstein logarithmic map is stated inconsistently between the text and Algorithm 2. These errors mean that the reported quality gains cannot be attributed to the method as written. The per-scene analysis also contains numerical claims that do not match the reported tables. The significance of the contribution is therefore currently not established.","major_comments":[{"comment":"The proposed merging rule is not a well-defined operation on Gaussian distributions. The Kalman gain K = Σ_t^Ob(Σ_t^Ob + Σ_t^P)^{-1} is a 3×3 matrix, while N_t^P and N_t^Ob are tuples (μ, Σ); the expression N_t^Ob + K(N_t^P - N_t^Ob) is undefined unless a block-diagonal application to μ and Σ is specified, which is not given. If the formula is applied elementwise to the covariance, the update \\(\\hat \\Sigma_t = \\Sigma_t^{Ob} + K(\\Sigma_t^P - \\Sigma_t^{Ob})\\) is not guaranteed to be symmetric or positive definite when Σ_t^Ob and Σ_t^P do not commute, and no symmetrization or SPD projection is provided. Because 3D Gaussian Splatting represents Σ as R S S^T R^T, an invalid \\(\\hat \\Sigma_t\\) makes rasterization undefined. In a standard Kalman filter the covariance update is not a convex combination of the prior and observation covariances; the 'no noise distribution' simplification in Sec. 3.1.3 removes the mechanism that makes the Kalman gain meaningful. The State Consistency Filter, a central claimed contribution, is therefore mathematically undefined as written.","section":"Sec. 3.1.3, Eq. (6)"},{"comment":"The logarithmic map formula is internally contradictory and is not used consistently. The text states \\(\\log_{\\Sigma_t} \\Sigma_{t-1} = (\\Sigma_t \\Sigma_{t-1})^{1/2} + (\\Sigma_{t-1} \\Sigma_t)^{1/2} - 2\\Sigma_{t-1}\\) and then identifies this with \\(P + P^T - 2\\Sigma_t\\), where \\(P = \\Sigma_{t-1}^{1/2}(\\Sigma_{t-1}^{1/2} \\Sigma_t \\Sigma_{t-1}^{1/2})^{1/2} \\Sigma_t^{-1/2}\\). These two expressions differ in both sign of the P terms and in the subscript of the last term. Algorithm 2 sets \\(\\Delta\\Sigma_t = 2\\Sigma_{t-1} - P - P^T\\), which is neither of them. Since \\(\\Delta\\Sigma_t\\) is the input to the exponential map in Eq. (14), the Wasserstein prediction step is not well-defined. The authors should state the correct formula and ensure the algorithm matches it.","section":"Sec. 3.3.1, Eq. (12) and Algorithm 2"},{"comment":"The per-scene analysis reports numbers that do not appear in the tables. The text states that in Hell Warrior 'Def-3D-Gauss achieves the highest PSNR of 41.54, while our method follows closely with 39.06'; Table 3 lists Hell Warrior Def-3D-Gauss as 32.37 and Ours as 34.38, and Table 4 lists 38.55 and 38.77. Similarly, the text claims 'in the Lego scene, our method achieves a PSNR of 34.74, surpassing 4D-Rotor-Gauss by approximately 9.5 dB,' while Table 3 Lego Ours is 30.14 and 4D-Rotor-Gauss is 29.48, and Table 4 lists 24.74 and 24.93. The text also gives Bouncing Balls PSNR 42.79, not found in either table. The per-scene discussion is therefore not supported by the reported tables, and the experimental claims cannot be independently checked.","section":"Sec. 4.3.1, Tables 3 and 4"},{"comment":"The paper claims that 'no previous work gave explicit and differentiable algorithms of Wasserstein geometry on Gaussians and involved such powerful tools in 3D/4D Gaussian Splatting,' yet the formulas in Eqs. (12)–(15) are taken directly from reference [37] without derivation. This overstates the novelty and places the mathematical backbone on a self-citation: [37] is co-authored by a co-author of this paper. For the reader to verify the method, the formulas should either be re-derived or the dependence on [37] should be clearly scoped; moreover, the correctness of the formulas should be checked independently, especially given the inconsistency identified in Eq. (12).","section":"Sec. 3.3 and Sec. 2.2"}],"minor_comments":[{"comment":"Equation (4) assigns v_t^P twice; the intended velocity update should be stated once. Also, the notation N_t is used for both a probability distribution and a state tuple, which makes expressions such as \\(N_t^{Ob} + K(N_t^P - N_t^{Ob})\\) ambiguous even if a block-diagonal structure were supplied.","section":"Sec. 3.1.2, Eq. (4)"},{"comment":"The caption of Table 3 says results are 'after applying the specified metric offsets,' but no offsets are specified in the text or caption; this must be clarified before the table can be used for comparison.","section":"Table 3 caption"},{"comment":"The training schedule is under-specified: the Filter is said to be 'introduced after 6k iterations' in Sec. 4.1, while Algorithm 3 applies the filter from the first iteration; the gating rule based on \\(\\|\\mu_{\\text{pred}} - \\mu_{\\text{obs}}\\| < 0.1\\sigma\\) and the >3\\sigma reversion are not reflected in the algorithm listings.","section":"Sec. 4.1 and Algorithm 3"},{"comment":"References [65] and [67] are the same paper (4D Gaussian Splatting); this duplicate citation should be corrected.","section":"References"},{"comment":"In Figure 1, the label 'Diff Gaussian Rasterization' appears to be a typo for 'Differentiable Gaussian Rasterization.' Also, the per-scene tables use different background settings (white vs black) with no statement of which setting is used for the aggregate Tables 1 and 2; this should be stated for reproducibility.","section":"Figure 1"}],"recommendation":"reject","confidential_remarks":"The mathematical core of the paper relies on a co-authored reference [37], and the novelty claim in Sec. 3.3 overstates the contribution relative to that reference. The load-bearing errors in Eq. (6) and Eq. (12) are not presentation-level issues; they make the proposed filter and prediction step undefined. Since the headline empirical numbers cannot be attributed to a method whose main component is not a valid operation on SPD matrices, I recommend rejection. If the authors correct the mathematical definitions, re-derive or properly cite the Wasserstein formulas, and reconcile the per-scene text with the tables, a substantially revised submission could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: treating 3D Gaussians as distributions on a Wasserstein manifold and fusing predicted/observed states with a Kalman-style gain is a sensible research direction, and the paper reports large quality and speed gains on standard benchmarks. But the core mathematics as written is not coherent. Eq. (6) never defines how the matrix gain acts on the Gaussian tuple (μ, Σ). If the same 3×3 K is applied to the covariance, nothing guarantees symmetry or positive-definiteness; the Kalman gain formula only makes sense for the mean, and the correct variance update is a precision sum, not a weighted sum. Eq. (12) contradicts itself (subtracting 2Σ_{t−1} then 2Σ_t) and Algorithm 2 flips the sign, so an independent reader cannot implement the log map. These are load-bearing, not cosmetic. The training-time reductions in Tables 6–7 (adding a loss that reduces wall-clock time by 20–57%) are unexplained and implausible unless iterations were cut, which is not stated. The evaluation also has an odd 'metric offsets' caption in Table 3 and repeated identical baseline numbers across scenes, which undermines confidence in the tables. On the positive side, the experimental setup follows standard protocols (150k iterations, A800) and the reported gains over 4D-Rotor-Gauss are sizable if true. The ablation design is logical: adding each component gives incremental PSNR gains. And the related work is reasonably placed, including KFD-NeRF and OTDNeRF, so the novelty claim is honest as a combination. But the paper needs a corrected derivation of both the filter and the log/exp maps, a release of code or at least an explanation of the training-time behavior, and a cleaned-up evaluation. As written, I would not accept or conditionally accept it. It is, however, the kind of paper that deserves a real referee rather than a desk reject, because the idea is timely and the empirical claims, if reproducible, would be useful to the 4DGS community. I'd tell the authors: fix the math, release the code, and resubmit. For a reading group, it's a good case study in why equations must be checked before experiments.","headline":"High-level idea worth exploring, but the core equations don't hold together and the reported training-time gains are unexplained.","tokens_in":17900,"tokens_out":3548,"would_cite":false,"duration_ms":31071,"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":"A state-space filter and Wasserstein geometry make 4D Gaussian splatting more accurate on moving scenes.","keywords":["4D Gaussian splatting","dynamic scene rendering","Wasserstein distance","state-space modeling","Kalman filter","Gaussian deformation field","temporal consistency","novel view synthesis"],"falsifier":"Take two non-commuting symmetric positive-definite matrices, compute the merged covariance by Eq. (6), and check whether it is symmetric and positive definite; if it fails on generic inputs, then the core update is not well-defined and the reported gains would need a repair step. A cleaner experiment would rerun the paper's full ablation with the merge replaced by a symmetrized, SPD-projected version and compare PSNR and training time on the same scenes.","tokens_in":16597,"feed_emoji":"🎬","tokens_out":5505,"duration_ms":50259,"temperature":0.7,"pith_summary":"The paper is trying to establish that dynamic-scene rendering improves when each 3D Gaussian is treated not as a bag of parameters but as a full Gaussian distribution that travels through Wasserstein space under a state-space model. It adds a state consistency filter that merges a deformation network's observation with a constant-velocity prediction, a Wasserstein-distance regularization that penalizes abrupt frame-to-frame changes, and Wasserstein logarithmic and exponential maps that replace Euclidean velocity and prediction updates. On standard synthetic and real-world dynamic benchmarks the full combination reports 34.45 and 32.79 dB PSNR, with the ablations attributing about 2 dB of the gain and 51 to 57 percent training-time savings to the Wasserstein components. A sympathetic reader would summarize the contribution as a geometric, physically motivated way to keep Gaussians consistent over time, uniting position and shape in one update.","feed_headline":"Wasserstein geometry and a state filter smooth 4D Gaussians over time","feed_subtitle":"A state-consistency filter and Wasserstein regularization lift dynamic-scene quality and speed up training.","key_machinery":"The load-bearing object is the squared 2-Wasserstein distance between two Gaussians, $W_2^2(\\mathcal N_1,\\mathcal N_2)=\\|\\mu_1-\\mu_2\\|^2+\\mathrm{Tr}(\\Sigma_1+\\Sigma_2-2(\\Sigma_1\\Sigma_2)^{1/2})$, together with the Wasserstein logarithmic and exponential maps on the manifold of symmetric positive-definite covariance matrices. The logarithmic map computes the velocity as a tangent vector from the previous Gaussian, the exponential map predicts the next Gaussian by solving a Sylvester equation, and the Kalman-like update in Eq. (6) merges that prediction with the network observation using a gain built from the two covariances. These formulas carry the argument: the stated quality and speed gains depend on the merged covariance remaining a valid symmetric positive-definite matrix and on the Wasserstein distance being the right coupling between position and shape.","core_discovery":"The central claim is that the natural state space for time-varying 3D Gaussians is the metric space of Gaussian distributions under the 2-Wasserstein distance, so dynamics should be written with exp and log maps there rather than as independent Euclidean increments of means and covariances. The paper implements this claim by using the deformation network's output as an observation, a constant-velocity prior as the prediction, and a Kalman-like gain to merge the two; the merged distribution is then rendered. It further claims that penalizing the Wasserstein distance between consecutive merged Gaussians is what suppresses flicker, because the metric couples translation and shape changes in a way that the Frobenius norm or mean-only regularization cannot. The quantitative assertion is that this package outperforms prior dynamic-scene renderers on the two benchmarks tested while cutting training time roughly in half.","pith_inferences":["Beyond the paper, a strict test would replace the Wasserstein log and exp maps with SPD-projected Euclidean updates while keeping the same losses; if quality is unchanged, the manifold geometry is not the active ingredient and the gains come from the smoothing losses alone.","The Kalman-like merge as written is not symmetric when the observation and prediction covariances do not commute, so an obvious extension is a proper Riemannian or projected filter; the paper's own experiments suggest the numerical optimizer may be implicitly repairing invalid covariances.","The same Wasserstein regularization idea could transfer to NeRF-style dynamic fields or voxel grids, but only in latent or distributional form, since those representations lack explicit Gaussian covariances."],"forward_implications":["If the paper is right, dynamic-scene renderers can obtain temporally coherent output without per-frame retraining, because the filter plus Wasserstein regularization keep adjacent Gaussians close in distribution.","The same Wasserstein log and exp dynamics could replace Euclidean deformation prediction in other Gaussian-splatting variants, making motion estimates more physically plausible.","The reported training-time reductions, from 3.5 to 1.5 hours on the synthetic benchmark and 4.5 to 2.2 hours on the real-world one, suggest that geometric regularization is a cheaper way to guide optimization than unconstrained parameter tuning.","Because the filter produces a cleaner motion field, the approach yields a natural speed field that can be projected to optical flow and could serve as motion supervision for other tasks."],"supporting_citations":[{"why":"Defines the 3D Gaussian representation with rotation and scale decomposition of the covariance that the paper builds on.","marker":"[26]"},{"why":"Supplies the Kalman filtering update structure whose gain equation is adapted to merge prediction and observation.","marker":"[25]"},{"why":"Gives the closed-form 2-Wasserstein distance between Gaussian distributions used in the regularization losses.","marker":"[19]"},{"why":"Provides the Wasserstein log and exp map formulas and the Sylvester equation solution used for dynamic prediction.","marker":"[37]"},{"why":"Contributes the synthetic dynamic-scene benchmark and the temporal encoding concept used by the deformation network.","marker":"[45]"},{"why":"Contributes the real-world dynamic-scene benchmark and its ground-truth video for optical-flow evaluation.","marker":"[31]"},{"why":"Is the strongest prior 4D Gaussian splatting baseline that the paper compares against and surpasses on both benchmarks.","marker":"[15]"}],"fun_headline_variants":["Wasserstein geometry guides 4D Gaussians along their natural path","State filter plus Wasserstein metric gives 4D Gaussians fluid motion","4D Gaussians follow Wasserstein geometry for smoother dynamic scenes","4D splatting with Wasserstein-constrained state-space dynamics","State consistency filter and Wasserstein loss improve 4D Gaussian rendering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Eq. (6) is a valid way to merge two Gaussians: the gain $K=\\Sigma^{\\mathrm{Ob}}_t(\\Sigma^{\\mathrm{Ob}}_t+\\Sigma^{\\mathrm{P}}_t)^{-1}$ can make the merged covariance asymmetric or non-positive-definite when the two covariances do not commute, and the paper gives no symmetrization or SPD projection before rendering.","fun_headline_variants_meta":{"raw":{"variants":["Wasserstein geometry guides 4D Gaussians along their natural path","State filter plus Wasserstein metric gives 4D Gaussians fluid motion","4D Gaussians follow Wasserstein geometry for smoother dynamic scenes","4D splatting with Wasserstein-constrained state-space dynamics","State consistency filter and Wasserstein loss improve 4D Gaussian rendering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00144,"raw_usage":{"total_tokens":5785,"prompt_tokens":910,"completion_tokens":4875,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":4783}},"tokens_in":526,"tokens_out":4875,"duration_ms":32047,"temperature":1.0,"reasoning_tokens":4783,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:30:42.222386+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two non-commuting symmetric positive-definite matrices, compute the merged covariance by Eq. (6), and check whether it is symmetric and positive definite; if it fails on generic inputs, then the core update is not well-defined and the reported gains would need a repair step. A cleaner experiment would rerun the paper's full ablation with the merge replaced by a symmetrized, SPD-projected version and compare PSNR and training time on the same scenes.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Kalman filtering update structure whose gain equation is adapted to merge prediction and observation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the closed-form 2-Wasserstein distance between Gaussian distributions used in the regularization losses."},{"cited_title":"Ge- ometric characteristics of the wasserstein metric on spd (n) and its applications on data processing.Entropy, 23(9):1214,","cited_arxiv_id":null,"evidence_quote":"Provides the Wasserstein log and exp map formulas and the Sylvester equation solution used for dynamic prediction."},{"cited_title":"D-nerf: Neural radiance fields for dynamic scenes","cited_arxiv_id":null,"evidence_quote":"Contributes the synthetic dynamic-scene benchmark and the temporal encoding concept used by the deformation network."},{"cited_title":"Neural 3d video synthesis from multi-view video","cited_arxiv_id":null,"evidence_quote":"Contributes the real-world dynamic-scene benchmark and its ground-truth video for optical-flow evaluation."},{"cited_title":"4d-rotor gaussian splatting: 11 Towards efficient novel view synthesis for dynamic scenes","cited_arxiv_id":null,"evidence_quote":"Is the strongest prior 4D Gaussian splatting baseline that the paper compares against and surpasses on both benchmarks."}],"review_version":1}