{"id":"b9cfe40f-7275-42c6-ad5e-dd200961c819","arxiv_id":"2412.16619","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 3D Gaussian Splatting variant that uses persistent homology to guide point-cloud densification and to regularize training, improving rendering quality on standard NVS benchmarks.","lead":"Topology-GS adds persistent homology to 3D Gaussian Splatting in two places: a Voronoi interpolation step that densifies sparse point clouds, and a training loss based on persistence barcodes that reduces perceptual distance to ground truth. The paper reports consistent, modest gains in PSNR, SSIM, and LPIPS on three standard novel-view synthesis benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's own numbers contradict the abstract's LPIPS claim: Pixel-GS has lower LPIPS on Mip-NeRF360 and a lower three-dataset average, so the strongest claim that Topology-GS 'outperforms existing methods in terms of ... LPIPS' is not supported.","rationale":"The reader's conditional verdict is appropriate; this concern does not overturn the PSNR/SSIM results or the overall conditional acceptance, but it sharpens the reader's noted overclaim. The reader mentioned that Pixel-GS is better on Mip-NeRF360 but did not compute the three-dataset average, which is the exact form of the strongest_claim. The PersLoss approximation issue raised by the reader is also real, but it affects the theoretical framing, not the empirical headline. The most load-bearing concern about the central claim is that the claim as stated in the abstract and strongest_claim is contradicted by Table 1 on LPIPS. This is concrete, verifiable, and directly testable from the published table. The recommended verdict remains CONDITIONAL, with a required correction to the LPIPS overclaim and release of per-scene numbers.","tokens_in":25902,"tokens_out":6817,"duration_ms":62625,"concrete_test":"Recompute the aggregate LPIPS from the full per-scene results used for Table 1 using the paper's aggregation rule (and also under unweighted dataset-mean aggregation). If Pixel-GS's aggregate LPIPS is below Topology-GS's, the abstract and the strongest claim must be revised to say Topology-GS achieves comparable LPIPS, or to claim LPIPS improvement only relative to Scaffold-GS. The authors should also release per-scene LPIPS values so this check can be reproduced independently.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is the headline result in Table 1: Topology-GS outperforms existing 3D-GS methods on average PSNR, SSIM, and LPIPS. The table itself undermines the LPIPS part. Pixel-GS reports LPIPS 0.165 on Mip-NeRF360, 0.161 on Tanks & Temples, and 0.251 on Deep Blending; Topology-GS reports 0.179, 0.160, and 0.245. Under the usual unweighted average of the three dataset means, Pixel-GS has (0.165+0.161+0.251)/3 = 0.1923, while Topology-GS has (0.179+0.160+0.245)/3 = 0.1947, so Pixel-GS is better by about 0.002. Under a per-scene aggregate weighted by dataset size (7 Mip-NeRF360 scenes, 2 Tanks & Temples, 2 Deep Blending), Pixel-GS also wins: (7*0.165+2*0.161+2*0.251)/11 ≈ 0.1799 vs (7*0.179+2*0.160+2*0.245)/11 ≈ 0.1875. Thus the abstract's unqualified claim of LPIPS superiority is false as written, and the strongest_claim's 'average LPIPS' phrasing is not supported by the paper's own data. This matters because it is the headline contribution of PersLoss: the paper claims PersLoss 'leads to a substantial reduction in LPIPS across all datasets' (Experiments, Overall Rendering Quality), yet the best competing method is still better on average LPIPS. The PSNR and SSIM claims may survive, but the LPIPS claim must be removed or qualified. This is not a subtle numerical issue; it is a direct contradiction between the stated conclusion and the reported evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Topology-GS, a 3D Gaussian Splatting variant with two novel components: LPVI, a persistent-homology-guided interpolation step that densifies sparse SfM point clouds while adaptively switching between 3D and 2D Voronoi interpolation based on a topological-difference threshold, and PersLoss, a regularization term that compares truncated persistence barcodes of rendered and ground-truth images. The method is evaluated on Mip-NeRF360, Tanks & Temples, and Deep Blending, with ablations, hyperparameter sensitivity studies, and extended experiments on NeRF Synthetic, BungeeNeRF, and IMW2020. The paper claims state-of-the-art PSNR, SSIM, and LPIPS, and provides an appendix with a convergence theorem for the topology-aware total loss, adapted from Zhang et al. (2022).","tokens_in":26391,"tokens_out":5466,"duration_ms":44557,"significance":"If the reported improvements hold, the paper would make a useful contribution: applying persistent homology to Gaussian Splatting for both initialization densification and training-time regularization is a plausible and relatively unexplored direction. The authors ship code, report abundant quantitative results, include ablations for each component, and test across multiple benchmark families. These are concrete strengths. However, the headline LPIPS claim is contradicted by the paper's own Table 1, and the theoretical convergence guarantee rests on assumptions about the relationship between PersLoss and the Wasserstein distance that are asserted rather than established. The PSNR and SSIM claims appear supportable, but the central 'outperforms existing methods in terms of PSNR, SSIM, and LPIPS' statement needs to be corrected or substantially qualified before the paper can be recommended for publication.","major_comments":[{"comment":"The abstract and the 'Overall Rendering Quality' subsection claim that Topology-GS outperforms existing methods in LPIPS and that PersLoss leads to a substantial LPIPS reduction across all datasets. Table 1 contradicts this. Pixel-GS reports LPIPS 0.165 on Mip-NeRF360, 0.161 on Tanks & Temples, and 0.251 on Deep Blending; Topology-GS reports 0.179, 0.160, and 0.245. The unweighted three-dataset averages are 0.1923 for Pixel-GS versus 0.1947 for Topology-GS, and a dataset-size-weighted average (7:2:2) gives 0.1799 versus 0.1875. Thus the unqualified LPIPS-superiority claim is false as written. The PSNR and SSIM claims may survive, but the LPIPS claim must be removed or qualified (e.g., 'lower LPIPS than Scaffold-GS on all three datasets').","section":"Abstract; Experiments, Overall Rendering Quality; Table 1"},{"comment":"The convergence theorem in Appendix C rests on Assumption A1 (Lreg = 0) and Assumption A2 (PersLoss ≤ d2(Dgm(f), Dgm(g))). Both are asserted rather than derived. A2 is particularly hard to justify: PersLoss matches truncated top-k barcodes by sorted lifespan, which is not generally an optimal Wasserstein matching, and truncation can omit points that the optimal matching would send to the diagonal. Since the proof of Theorem 1 replaces PersLoss with the optimal-matching loss Ltopo (Definition 1 in B4), this inequality is load-bearing for the claimed convergence guarantee. Please validate A2 empirically on representative rendered/ground-truth image pairs by comparing PersLoss with the exact 2-Wasserstein distance, or state the theorem as conditional on a verified inequality.","section":"Appendix B3-B4; Assumptions A1-A2"},{"comment":"Equation (18) defines the matching as γt = arg max Σ (Bγ(p)^2 + Dγ(p)^2), which maximizes the loss, while Definition 1 in Appendix B4 minimizes the same sum. This sign inconsistency appears in the core derivation and should be corrected; as written it invalidates the derivation of Lemma 3, which is imported from Zhang et al. (2022).","section":"Appendix B5, Eq. (18)"},{"comment":"The quantity B is defined in the main text as B = k0 + k1 + k2, the number of retained barcodes, but in Appendix C1 it is defined as B = card(Dgm*), the cardinality of the full ground-truth diagram excluding the diagonal. These are not the same, and the step-size bound in Theorem 1 depends on B. Please disambiguate the notation and state whether the O(1/ϵ) bound uses the number of truncated barcodes or the full diagram cardinality.","section":"Main text, Theorem 1; Appendix C1"}],"minor_comments":[{"comment":"The notation y ⊂ R3×H×W should use set membership (∈) rather than the subset symbol (⊂).","section":"Notations"},{"comment":"The PersLoss weighting uses Betti numbers βi, but after truncating to the top-k barcodes the Betti numbers of the remaining features are not clearly defined; please clarify how βi is computed for the truncated barcode.","section":"Equation (8)"},{"comment":"The claim that this work 'pioneers' or is the 'first successful integration' of topology with 3D-GS is stronger than the related-work discussion supports, since Jignasu et al. (2024) and Nishikawa et al. (2024) are cited as using persistent homology for point clouds in a 3D or 3D-GS context; please soften or document the novelty more precisely.","section":"Introduction; Related Works"},{"comment":"The subsection title promises comparisons of depth distortion, but Table 4 reports training time, FPS, PSNR, and LPIPS; no depth distortion metric such as Chamfer distance or depth error is given.","section":"Appendix D4, Table 4"},{"comment":"The text says k0, k1, k2 represent '1D, 2D, and 3D homology'; these are H0, H1, and H2, so the indices are off by one and should be corrected.","section":"Appendix D2"},{"comment":"The caption refers to red and blue curves but does not state which color corresponds to which k setting; please annotate the figure or caption with the k values for each curve.","section":"Figure 8 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an AAAI camera-ready version (copyright notice present). If the target venue is a journal, the editor should consider whether the formatting and the very large appendix are appropriate, and the authors should be asked to signpost the adaptation of the convergence proof from Zhang et al. (2022) more clearly. The central issue is the unsupported LPIPS-superiority claim; this is fixable in a revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the empirical package, not the theory. The contribution is a plausible combination: LPVI, a persistent-homology-guided Voronoi interpolation for the initial sparse point cloud, and PersLoss, a barcode-matching regularizer activated during the ADC stage. The ablations in Table 2 support both components, and Table 1 shows modest but consistent PSNR/SSIM gains over Scaffold-GS, plus a small LPIPS win on Tanks & Temples and Deep Blending. That is a legitimate result for the 3DGS subfield, and the TDA hook is new relative to prior PH-in-point-cloud work.\n\nNow the soft spots.\n\nFirst, the abstract overclaims LPIPS. The paper says it outperforms existing methods in PSNR, SSIM, and LPIPS, but Table 1 contradicts that on Mip-NeRF360: Pixel-GS reports 0.165, Topology-GS 0.179. On the three-dataset average, Pixel-GS also wins. That is not subtle. The experiments section's claim of a 'substantial reduction in LPIPS across all datasets' is also not supported; the reduction is relative to their baseline, not to all methods. They need to fix the claim or report per-scene numbers.\n\nSecond, the convergence theorem is imported from Zhang et al. 2022 and rests on assumptions A1 and A2 that the paper asserts rather than defends. A1 sets Lreg = 0, justified by truncation; A2 says PersLoss is bounded above by the Wasserstein distance, justified by 'negligible' error between length-based matching and optimal matching. The paper does not test that empirically, and the truncation is the very thing being optimized, so the chain from relaxed loss to the actual rendered-image loss is weak. I would treat Theorem 1 as a borrowed guarantee for a neighboring problem, not a guarantee for this pipeline. That is more framing than fatal, since the empirical result does not depend on the theorem, but the authors should not present it as their own convergence proof.\n\nThird, reproducibility: the code link is there but no commit hash, and 'each scene run once on each of eight GPUs and averaged' is not the same as per-scene statistics; no error bars are given. Minor.\n\nOverall, the central argument that the combination helps holds up at the level of Table 1 and the ablations. The paper is for researchers working on 3DGS densification and TDA-based losses in rendering, and it deserves a serious referee—the method is clearly described and the ablations are honest enough to fix. I would send it to review, asking for a corrected LPIPS claim, per-scene numbers, and a softened theory section.","headline":"A solid empirical TDA-for-3DGS paper whose abstract overclaims LPIPS, with a convergence theorem that is more imported than proven; send it to review and make them fix the claim.","tokens_in":26935,"tokens_out":1537,"would_cite":true,"duration_ms":12797,"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":"Topology-GS inserts persistent homology in two places—densifying the initial point cloud and regularizing training—and reports better PSNR, SSIM, and LPIPS than earlier splatting methods.","keywords":["3D Gaussian Splatting","persistent homology","novel-view synthesis","point cloud densification","Voronoi interpolation","topological loss","perceptual similarity","convergence guarantee"],"falsifier":"Render a test scene with dense true topology, such as a chain-link fence whose many loops have nearly equal lifespans, and train Topology-GS while increasing the truncation counts $k_0$, $k_1$, $k_2$: the paper's own convergence bound tightens as $1/B^2$ for $B = k_0+k_1+k_2$, so a fixed learning rate should eventually diverge, and if instead quality keeps improving, the assumption that sorted-lifespan matching approximates the exact Wasserstein distance (the paper's Assumptions A1–A2) is doing more work than the analysis accounts for.","tokens_in":25654,"feed_emoji":"🖼️","tokens_out":21409,"duration_ms":163285,"temperature":0.7,"pith_summary":"The paper sets out to show that persistent homology—the study of when loops, holes, and connected components are born and die as a shape grows—can fix two known weaknesses of 3D Gaussian Splatting (3D-GS), the point-based rendering method for novel-view synthesis. Sparse Structure-from-Motion point clouds under-cover low-curvature surfaces, so the paper introduces Local Persistent Voronoi Interpolation (LPVI): before training it densifies the cloud locally, using the Wasserstein distance between persistence diagrams as a topological-difference check that decides whether safe 3D interpolation is possible or the algorithm must fall back to interpolation on a 2D plane. During optimization, PersLoss—a differentiable regularizer comparing the longest-lived topological features (truncated persistence barcodes) of the rendered and ground-truth images, seen as point clouds in RGB space—is switched on only during the adaptive density-control stage. On the Mip-NeRF360, Tanks & Temples, and Deep Blending benchmarks, Topology-GS reports higher PSNR and SSIM than every compared method on all three datasets and the lowest LPIPS on two of them, with minor memory cost and no inference-time cost. A reader should care because the paper demonstrates a structural prior that needs no extra neural network, comes with a stated convergence guarantee for the combined loss, and targets exactly the perceptual gap that pixel-level metrics miss.","feed_headline":"Topology-aware Gaussians beat splatting baselines on 3 benchmarks","feed_subtitle":"A homology-gated densifier plus a training-time barcode loss improves all three key metrics.","key_machinery":"The machinery is persistent homology over $\\alpha$ complexes of point sets, packaged as persistence barcodes: as a filtration radius $\\alpha$ grows, each connected component, loop, or void records an interval from birth to death, and longer intervals mark features the authors treat as structurally significant. LPVI uses the Wasserstein distance between two local persistence diagrams as its TopoDiff gate, deciding when adding 3D Voronoi vertices is topologically safe versus when to switch to 2D-plane interpolation. PersLoss uses the truncated top-$k$ barcodes of the rendered and ground-truth images reshaped into RGB point clouds, matched by sorted lifespan, and turns the birth–death coordinate differences into a differentiable loss. The convergence theorem binds the two uses: with $B = k_0 + k_1 + k_2$ kept small, the sorted matching with no shrinking-cost regularizer is assumed to approximate the Wasserstein-optimal topological loss, and the step-size bound $\\eta \\le \\min\\left\\{\\frac{1}{2\\ell_2 + 10\\lambda_{\\mathrm{topo}}B},\\ \\frac{\\epsilon}{4096\\lambda_{\\mathrm{topo}}^2 B^2}\\right\\}$ yields termination in $O(1/\\epsilon)$ iterations.","core_discovery":"The central claim is that one mathematical tool—persistent homology—can be used at two moments of the 3D-GS pipeline to restore structural integrity at two levels. At initialization, LPVI looks at each point of the sparse SfM cloud with its $K$ nearest neighbors, computes the persistence diagrams of that local set before and after adding Voronoi-tessellation vertices, and accepts the 3D interpolation only when the Wasserstein distance between the two diagrams stays below a threshold $\\tau$; otherwise it interpolates on a PCA-fitted plane and maps the new points back, a choice the authors justify by the manifold hypothesis for low-curvature regions. During optimization, PersLoss reshapes the rendered and ground-truth images into $HW \\times 3$ RGB point clouds, computes their $\\alpha$-complex persistence barcodes, keeps the $k_0$, $k_1$, and $k_2$ longest-lived features in homology dimensions 0, 1, and 2, matches them by sorted lifespan rather than by optimal assignment, and penalizes the squared birth and death differences weighted by Betti numbers. The paper proves that this topology-aware total loss—PersLoss plus the usual L1 and SSIM supervision—terminates in $O(1/\\epsilon)$ iterations under regularity assumptions on the filter function and the supervision loss, provided the step size obeys the bound in Theorem 1, and that the small truncation count $B = k_0 + k_1 + k_2$ is what keeps that bound practical. The empirical statement is that Topology-GS achieves state-of-the-art PSNR, SSIM, and LPIPS on the three benchmarks, which the authors read as the first successful integration of topology into 3D-GS.","pith_inferences":["Editorial inference: the 'outperforms earlier work' summary is an average statement; the paper's own Table 1 shows Pixel-GS still reports the lower LPIPS on Mip-NeRF360 (0.165 versus 0.179), so the safest public reading is best on average across benchmarks rather than best in every cell.","Extension: the TopoDiff gate is not specific to Voronoi interpolation—the same local persistence-diagram distance could guard any densification, pruning, or smoothing step in a point-cloud pipeline, making persistent homology a generic topology-safety check.","Testable variant: computing PersLoss on other filter functions (grayscale intensity, depth, or the feature maps of a pretrained network) instead of raw RGB coordinates would reveal whether the reported gains come from color-distribution topology or spatial-structure topology, and would show how closely the loss tracks what LPIPS actually measures.","Design implication: the truncation counts $k_0$, $k_1$, $k_2$ are the method's real tuning dials—too few retained features and PersLoss misses structure, too many and the step-size bound collapses as $1/B^2$ while noise bars dominate, exactly the trade-off the paper's Figure 8 and its appendix remark about $B$ describe."],"forward_implications":["On the three reported benchmarks, Topology-GS tops every compared method in PSNR and SSIM and lowers LPIPS relative to the Scaffold-GS baseline on every dataset, so adopting it gives both sharper pixels and closer perceptual features.","PersLoss runs only during the adaptive density-control stage of training, so the perceptual gains come with no added inference cost and only a small training-time and memory overhead (under 20 MB per scene from LPVI).","The depth-map comparisons show LPVI smooths depth in low-curvature regions while PersLoss sharpens depth at structural boundaries, indicating that the topological constraints improve geometry rather than pixel statistics alone.","The ablations show the two contributions are nearly separable—LPVI carries the PSNR and SSIM gains while PersLoss carries the LPIPS gains—so each can be adopted independently depending on which deficiency matters.","Theorem 1 licenses using PersLoss together with L1 and SSIM supervision and the adaptive density control without a separate shrinking-cost regularizer, so the extra topological term does not require reworking the 3D-GS optimization loop."],"supporting_citations":[{"why":"Defines 3D Gaussian Splatting itself—the Gaussian parametrization, tile-based rasterization, and adaptive density control that Topology-GS extends, plus the L1 and SSIM supervision terms.","marker":"(Kerbl et al. 2023)"},{"why":"Supplies the convergence framework for persistent-homology losses—the matching decomposition into restoration and shrinking costs, the Lreg regularizer, and the bounded-increase lemma behind Theorem 1.","marker":"(Zhang et al. 2022)"},{"why":"Provides the differentiable topology layer that lets gradients flow through the persistence barcodes, making PersLoss trainable.","marker":"(Gabrielsson et al. 2020)"},{"why":"Provides the alpha-complex persistence computation used to build the persistence diagrams and barcodes in both LPVI and PersLoss.","marker":"(Maria et al. 2014)"},{"why":"Documents the low-curvature sparsity problem and supplies SAGS, the structure-aware densification baseline that LPVI is compared against in the ablations.","marker":"(Ververas et al. 2025)"},{"why":"Supplies Scaffold-GS, the baseline whose losses Topology-GS inherits and the main comparison point in the tables and ablations.","marker":"(Lu et al. 2024)"},{"why":"Defines LPIPS, the perceptual metric whose reduction PersLoss is designed to achieve and by which feature-level quality is measured.","marker":"(Zhang et al. 2018)"},{"why":"Supplies the Mip-NeRF360 benchmark and evaluation protocol used for the main quantitative and visual results.","marker":"(Barron et al. 2022)"},{"why":"The Structure-from-Motion pipeline whose sparse, uneven point clouds create the low-curvature coverage problem that LPVI densifies.","marker":"(Schönberger and Frahm 2016)"}],"fun_headline_variants":["Homology-spiked Gaussians outscore splatting on 3 tests","Topology-GS: persistent homology boosts 3D Gaussian splatting","Barcode loss fixes Gaussian splatting integrity, tops benchmarks","Persistent homology strengthens Gaussian splatting, beats baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the few longest-lived topological features of the rendered and ground-truth images, matched by rank instead of by true geometric correspondence, give a faithful enough measure of perceptual difference that shrinking that measure during training improves rendering the way the paper claims; the paper asserts this as Assumptions A1 and A2 in the appendix rather than deriving or testing it.","fun_headline_variants_meta":{"raw":{"variants":["Homology-spiked Gaussians outscore splatting on 3 tests","Topology-GS: persistent homology boosts 3D Gaussian splatting","Barcode loss fixes Gaussian splatting integrity, tops benchmarks","Persistent homology strengthens Gaussian splatting, beats baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000571,"raw_usage":{"total_tokens":2791,"prompt_tokens":1128,"completion_tokens":1663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":1586}},"tokens_in":744,"tokens_out":1663,"duration_ms":11100,"temperature":1.0,"reasoning_tokens":1586,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:24:07.271869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a test scene with dense true topology, such as a chain-link fence whose many loops have nearly equal lifespans, and train Topology-GS while increasing the truncation counts $k_0$, $k_1$, $k_2$: the paper's own convergence bound tightens as $1/B^2$ for $B = k_0+k_1+k_2$, so a fixed learning rate should eventually diverge, and if instead quality keeps improving, the assumption that sorted-lifespan matching approximates the exact Wasserstein distance (the paper's Assumptions A1–A2) is doing more work than the analysis accounts for.","supporting_citations":[],"review_version":1}