{"id":"b09d3a78-c6b2-4ba0-ab69-31ee438fc5ba","arxiv_id":"2411.14514","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"NexusSplats replaces per-Gaussian appearance codes with kernel-level shared codes and 3D uncertainty propagation, achieving comparable rendering quality with 65.4% fewer parameters and 2.7x faster training.","lead":"The paper presents NexusSplats, a 3D scene reconstruction method that groups millions of Gaussian primitives under shared 'kernels' so lighting and moving-object effects are learned with far fewer parameters. On the Photo Tourism outdoor benchmark it reports rendering quality on par with the best existing method while cutting training time by roughly 2.7x and parameters by 65%.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art rendering quality is the load-bearing claim, but Table 1 contradicts it: NexusSplats wins PSNR while losing LPIPS on two of three Photo Tourism scenes, and Appendix B concedes it trails WildGauss on NeRF On-the-go. The efficiency gain is solid; the quality claim is not.","rationale":"The paper's central claim has two components: (1) state-of-the-art rendering quality and (2) large efficiency gains. The efficiency component is internally consistent and is the paper's strongest contribution: Table 2 shows a 79.3% reduction in appearance embeddings, 65.4% fewer total parameters than WildGauss, and 2.72x less training time, all from the centralized kernel-level design. Parameter counts are plausible given the Scaffold-GS hierarchy, and the ablations in Table 1 show each removed module degrades PSNR. The quality component, however, is the load-bearing weak point. In Table 1, NexusSplats is best on PSNR on all three Photo Tourism scenes, but WildGauss has higher SSIM on Brandenburg Gate, equal on Sacre Coeur, and lower LPIPS on Brandenburg Gate and Trevi Fountain; averaged, SSIM and LPIPS favor WildGauss. Appendix B explicitly says the method lags behind WildGauss on NeRF On-the-go. These are the paper's own measurements, not a challenge from an outside consensus, so this is an internal evidence gap rather than a 'controversial but possibly correct' claim. The reader's identified weakest assumption (DINO-based boundary refinement, Eqs. 8 and 10) is a real mechanism risk, but it is conditional on DINO misbehaving. The metric contradiction is unconditional: even with perfect DINO features, the reported numbers do not establish 'state-of-the-art rendering quality.' Thus the most load-bearing issue is not a speculative failure mode but a direct mismatch between the abstract's claim and the paper's own comparisons. A concrete fix is to re-evaluate with per-image metric distributions and, if LPIPS/SSIM are not systematically better, to scope the claim to PSNR and efficiency. This does not change the overall CONDITIONAL verdict, but it sharpens the condition: the quality claim in the abstract and conclusion should be revised or supported by additional evidence.","tokens_in":14070,"tokens_out":5775,"duration_ms":51601,"concrete_test":"Run an independent evaluation using the released code and NerfBaselines on the Photo Tourism test splits and report per-image PSNR/SSIM/LPIPS distributions for NexusSplats and WildGauss, not just scene means. If WildGauss is statistically better or tied on LPIPS/SSIM in more scenes, the phrase 'state-of-the-art rendering quality' in the abstract and conclusion should be revised to 'competitive rendering quality with higher PSNR and substantially lower parameter and training cost,' and the central claim should be scoped accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that hierarchical light decoupling plus structure-aware occlusion handling jointly deliver 'state-of-the-art rendering quality' and a 65.4% parameter reduction with 2.7x faster reconstruction. The efficiency half is well supported by Table 2 and the arithmetic is consistent. The quality half is not supported by the paper's own numbers. Averaged over the three Photo Tourism scenes, NexusSplats improves PSNR over WildGauss (24.95 vs 24.44) but has slightly worse mean SSIM (0.849 vs 0.850) and worse mean LPIPS (0.184 vs 0.180), and it loses LPIPS on Brandenburg Gate and Trevi Fountain. Appendix B explicitly states that on NeRF On-the-go 'quantitative results ... lag behind specialized baselines like WildGauss.' So the only metric on which NexusSplats is consistently first is PSNR, which is not the same as 'state-of-the-art rendering quality.' The DINO-based boundary assumption in Eqs. 8 and 10 is a legitimate mechanism-level risk, but the claim fails even under the optimistic assumption that DINO features are perfect: the reported metrics already do not show a perceptual-quality advantage. Therefore the load-bearing weakness is an internal evidence gap: the headline quality claim is broader than the measured outcomes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"NexusSplats proposes a 3D Gaussian Splatting framework for in-the-wild scene reconstruction. The method introduces hierarchical light decoupling, in which spatially grouped 'nexus kernels' share appearance embeddings and coordinate color mapping through an MLP (Eq. 5), and a structure-aware occlusion-handling pipeline that predicts per-Gaussian 3D uncertainties, propagates them to 2D via alpha-blended rasterization (Eq. 7), and applies a DINO-derived boundary-aware refinement mask (Eq. 8). The training objective combines a color loss weighted by the predicted uncertainty mask (Eq. 9) and an uncertainty loss using DINO feature cosine similarity (Eq. 10). The paper reports quantitative results on the Photo Tourism dataset (Table 1), efficiency comparisons (Table 2), ablations, and qualitative results on NeRF On-the-go. The central claims are that NexusSplats achieves state-of-the-art rendering quality, reduces total parameters by 65.4% relative to WildGauss, and trains 2.7× faster.","tokens_in":14434,"tokens_out":3207,"duration_ms":30897,"significance":"The efficiency contribution is solid and well quantified: Table 2 shows a parameter reduction from 77.1M to 26.7M (64.5% fewer, matching the stated 65.4% when including the exact numbers) and a training-time reduction from 18.54 to 6.81 GPU hours (2.72× faster, consistent with the stated 2.7×). The ablations in Table 1 demonstrate that each proposed component contributes positively to PSNR, and the supplementary material includes code, pre-trained models, and detailed hyperparameters, which strengthens reproducibility. The conceptual idea of centralizing appearance embeddings at the kernel level rather than per-Gaussian is a reasonable design direction and may reduce overfitting and memory usage. However, the headline claim of 'state-of-the-art rendering quality' is not established by the paper's own numbers, and this overstatement currently weakens the contribution. The work has clear value for efficient in-the-wild Gaussian splatting if the quality claims are revised to match evidence.","major_comments":[{"comment":"The claim that NexusSplats 'achieves state-of-the-art rendering quality' is not supported by the reported metrics. Averaged over the three Photo Tourism scenes, NexusSplats improves PSNR over WildGauss (24.95 vs 24.44) but has a slightly worse mean SSIM (0.849 vs 0.850) and a worse mean LPIPS (0.184 vs 0.180). Per scene, NexusSplats loses LPIPS on Brandenburg Gate (0.141 vs 0.135) and Trevi Fountain (0.237 vs 0.228), and it does not win SSIM on any scene except Trevi Fountain by 0.001. Thus the only metric on which NexusSplats is consistently first is PSNR. The abstract and Section 1 should be revised to claim 'state-of-the-art PSNR' or 'comparable rendering quality with substantially improved efficiency', rather than 'state-of-the-art rendering quality.'","section":"Abstract, §4.2, Table 1"},{"comment":"The supplementary material contains a direct contradiction of the main-text quality claim. In Appendix B, the authors state that on the NeRF On-the-go dataset 'quantitative results ... lag behind specialized baselines like WildGauss', and Table 4 confirms this: NexusSplats is worse than WildGauss on nearly every metric across the six scenes, often by large margins (e.g., Patio PSNR 22.64 vs 24.02; Mountain LPIPS 0.287 vs 0.233). Appendix D then says the method achieves 'a critical balance' without 'compromising the core advantages of 3DGS — real-time rendering and photorealistic quality — in its target domain.' This internal inconsistency, combined with the Photo Tourism LPIPS results, means the paper's stated contribution of 'state-of-the-art rendering quality' is broader than the evidence. The discussion should either restrict the quality claim to PSNR on Photo Tourism or add an analysis of why LPIPS/SSIM do not improve, with supporting ablations.","section":"Appendix B, Table 4, Appendix D"},{"comment":"The boundary-aware refinement models an anisotropic 2D Gaussian mask whose parameters are 'estimated via DINO features', but the paper does not describe how the DINO features are converted into a foreground-scope estimate, nor does it evaluate the sensitivity of the final quality to this estimation. Since Eq. (8) is the mechanism that prevents clear scene boundaries from being suppressed, a mis-estimated foreground scope could directly degrade the quality gains claimed in the paper. An additional experiment varying the DINO feature layer or adding a quantitative comparison of the boundary mask against manual annotations would make this load-bearing component more credible.","section":"§3.3.2, Eq. (8)"}],"minor_comments":[{"comment":"The symbol Σ is used for the covariance matrix of a 3D Gaussian in Eq. (1) and again for the covariance of the 2D boundary Gaussian in Eq. (8); this overloads the notation within the same paper and should be clarified (e.g., use Σ_b or a subscript).","section":"§3.1, §3.3.2"},{"comment":"The caption says 'NerRF On-the-go dataset' — the typo 'NerRF' should be 'NeRF'.","section":"Figure 4 caption"},{"comment":"The sentence 'our future work will explore adaptive kernel scaling' starts with a lowercase letter after a period and should be capitalized as a new sentence.","section":"Appendix B"},{"comment":"The text immediately before Eq. (9) says 'weighted sum of DSSIM and L1 losses' but the equation displays a division by 2β²M; the relationship between the text '1 − λ' and the equation's notation should be made consistent (e.g., write \\(\\frac{1-\\lambda}{2\\hat{\\beta}^2 \\mathcal{M}}\\)).","section":"§3.4, Eq. (9)"},{"comment":"Several reference entries (e.g., [6], [13], [14], [20], [39], [47]) have trailing page numbers after the year, such as '2024. 2, 4, 5, 1', which appear to be citation-page markers rather than journal page ranges; these should be cleaned up to match standard reference formatting.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The efficiency story is strong and the code release is a positive signal. The main issue is that the 'state-of-the-art rendering quality' claim, repeated in the abstract, introduction, and conclusion, is directly contradicted by the paper's own quantitative results on LPIPS (Photo Tourism) and on nearly all metrics for NeRF On-the-go. This is a wording and evidence-calibration problem rather than a flaw in the method itself; I believe it is fixable by revising the claims and adding a more nuanced discussion of the metric-wise trade-offs. The DINO-based boundary refinement is an interesting idea but needs one validation experiment to be convincing. I recommend major revision, not rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing: ignore the headline, trust the efficiency numbers. The architectural idea — moving appearance embeddings from per-Gaussian to per-anchor 'nexus kernels' on top of Scaffold-GS — is a real and sensible compression of WildGauss-style appearance handling, and the 65.4% parameter reduction and 2.7x training speedup are arithmetically consistent with Table 2. The ablations show each module contributes, and the paper includes code and pretrained models, which is more than most in this area.\n\nWhat's genuinely new is the combination: kernel-level shared appearance embeddings plus 3D uncertainty propagation through Gaussian splatting, with a DINO-driven boundary refinement. The idea of grounding occlusion reasoning in 3D geometry rather than purely 2D uncertainty maps is worth taking seriously.\n\nThe soft spots are where the paper overreaches. The abstract and conclusion claim 'state-of-the-art rendering quality,' but the paper's own Table 1 does not support that across metrics: versus WildGauss, NexusSplats wins PSNR on all three Photo Tourism scenes but has slightly worse mean SSIM and worse LPIPS on two of three scenes. Appendix B then concedes that on NeRF On-the-go, quantitative results lag behind WildGauss. So 'state-of-the-art' only holds if you define quality as PSNR alone, and the paper never says that. The DINO-based boundary model (Eqs. 8 and 10) is a legitimate mechanism risk, but the claim fails even with perfect DINO features — the measured perceptual metrics already don't show an advantage.\n\nAlso worth noting: the paper excludes Wild-GS, SWAG, and WE-GS, citing lack of public implementations or hardware constraints. That exclusion might be legitimate, but the comparison set is still a bit thin for the 'state-of-the-art' label.\n\nBottom line: this is a solid efficiency-oriented contribution that deserves a serious referee. The central efficiency claim is reproducible in principle and the ablations are coherent. The quality claim needs to be restated or supported with more complete metrics. A revision that narrows the claim to 'state-of-the-art PSNR with substantially lower parameter count' would make the paper honest and still strong. I'd send it out.","headline":"Solid efficiency win, but the 'state-of-the-art rendering quality' claim only survives if you define quality as PSNR alone.","tokens_in":14949,"tokens_out":2870,"would_cite":true,"duration_ms":24291,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"NexusSplats claims that organizing 3D Gaussians into shared nexus kernels for appearance and uncertainty learning achieves state-of-the-art in-the-wild rendering with 65.4% fewer parameters and 2.7× faster reconstruction.","keywords":["NexusSplats","3D Gaussian Splatting","in-the-wild reconstruction","hierarchical light decoupling","transient occlusion handling","uncertainty propagation","nexus kernels","appearance embedding"],"falsifier":"Annotate the transient objects (pedestrians, vehicles) in a held-out subset of Photo Tourism test images and compare the predicted uncertainty masks against those annotations; if the masks miss occluders that resemble the static scene, or if removing boundary-aware refinement improves PSNR on scenes with strongly non-elliptical foregrounds, then the structure-aware occlusion claim fails in that regime.","tokens_in":13848,"feed_emoji":"📷","tokens_out":8603,"duration_ms":74394,"temperature":0.7,"pith_summary":"Photorealistic reconstruction from unstructured photo collections must cope with changing lighting and transient objects, and the paper argues that the standard fixes—per-Gaussian appearance embeddings and 2D-only uncertainty maps—are both at the wrong granularity. NexusSplats instead groups 3D Gaussians under shared nexus kernels that learn appearance and uncertainty centrally, then propagates the resulting 3D uncertainties into 2D masks through the same rasterizer used for color, with a DINO-based boundary refinement to stop clear scene edges from being misclassified as occluders. The paper reports that this combination reaches higher PSNR than per-Gaussian in-the-wild baselines on Photo Tourism while reducing total parameters by 65.4% and cutting reconstruction time from 18.54 to 6.81 GPU hours. If correct, the contribution is a practical efficiency result: in-the-wild quality and occlusion handling at a cost that no longer grows with every Gaussian.","feed_headline":"NexusSplats cuts scene parameters 65% and trains 2.7x faster","feed_subtitle":"Grouping Gaussians into shared nexus kernels improves in-the-wild rendering and occlusion removal while slashing reconstruction time.","key_machinery":"The load-bearing object is the nexus kernel, an anchor-based 3D primitive inherited from Scaffold-GS's hierarchical Gaussian management, which spawns h neural Gaussians and provides them with shared appearance and uncertainty embeddings. This kernel carries the argument because it converts appearance learning from scattered per-Gaussian parameters into centralized per-kernel learning, which is both cheaper and, the paper argues, better able to encode non-local lighting effects. On the occlusion side, the same tile rasterizer that composites colors composites per-Gaussian 3D uncertainties into 2D masks, and a boundary-aware refinement model $M(x') = e^{-\\frac{1}{2}(x'-\\mu')^\\top \\Sigma^{-1}(x'-\\mu')}$ estimated from DINO features modulates those masks near scene edges; the uncertainty loss $L_u = D(\\hat{C}, C)/(2 \\hat{\\beta}^2) + \\lambda_1 \\log \\hat{\\beta}$ uses DINO feature agreement between rendered and ground-truth images as the signal for what counts as an occlusion.","core_discovery":"The central claim is that in-the-wild scene reconstruction improves when appearance and uncertainty are modeled at the level of the scene's hierarchical structure rather than at the level of individual Gaussians. Nexus kernels—dynamic 3D primitives managing h neural Gaussians each—carry a shared appearance embedding that transforms all their Gaussians' colors through one lighting-sensitive MLP, and a shared uncertainty embedding that predicts per-Gaussian 3D uncertainties. Those uncertainties are alpha-blended onto the image plane exactly like color, giving 2D uncertainty masks that inherit 3D geometry, and the anisotropic 2D Gaussian boundary refinement, estimated from DINO features, downweights uncertainty near scene peripheries where rendering quality is naturally low. The paper's reported results are state-of-the-art average PSNR on Photo Tourism (24.95 versus WildGauss's 24.44), 79.3% fewer appearance embeddings, 65.4% fewer total parameters, and 2.7× faster training, with qualitative demonstrations of relighting to target conditions and removal of pedestrians and vehicles.","pith_inferences":["A direct extension the paper leaves implicit: the same kernel-level sharing argument applies to any other per-primitive latent, such as semantic labels, deformation codes, or per-object transient flags, so the parameter savings should generalize beyond appearance and uncertainty.","The boundary-aware refinement assumes the foreground roughly follows a single anisotropic 2D Gaussian in every image; scenes with disjoint foreground objects, wide panoramas, or off-center compositions would stress this prior and could be used as a discriminating test of the refinement's value.","The paper's own limitation section notes that masking occluded pixels leaves holes in the reconstruction; coupling the uncertainty masks with inpainting could turn occlusion filtering into occlusion completion, which is a different capability than the one measured here.","Average image metrics like PSNR can hide where occlusion errors occur; labeling transient objects in a few test images and computing mask IoU against the predicted uncertainty would give a sharper test of the structure-aware occlusion claim than the reported global metrics."],"forward_implications":["Grouping Gaussians into shared nexus kernels removes the need for per-Gaussian appearance embeddings: appearance embeddings drop by 79.3% relative to WildGauss, total parameters by 65.4%, and training time from 18.54 to 6.81 GPU hours on Photo Tourism scenes.","Rendering 3D uncertainties through the same rasterizer as colors gives occlusion masks that are geometrically consistent with the scene, suppressing false positives in textureless regions and at clear scene boundaries.","Because appearance is a kernel-level learned embedding, a trained scene can be recolored to match arbitrary target lighting conditions, as demonstrated across six lighting conditions and three viewpoints per Photo Tourism scene.","The claimed gains are specific to large outdoor scenes: on the small-scale NeRF On-the-go indoor scenes, the paper reports competitive occlusion removal but lower PSNR than WildGauss, so the efficiency advantage does not transfer to confined indoor geometry.","The full pipeline—centralized appearance learning, 3D uncertainty propagation, and boundary-aware refinement—is needed for the reported quality; ablations in the paper show removing any one module lowers PSNR."],"supporting_citations":[{"why":"Supplies the hierarchical anchor-based Gaussian management that NexusSplats adapts into nexus kernels.","marker":"[20]"},{"why":"The per-Gaussian in-the-wild 3DGS baseline whose appearance embeddings, parameter count, training time, and DINO-based 2D uncertainty constitute the main comparison.","marker":"[14]"},{"why":"Introduces the appearance-and-transient decomposition and the in-the-wild evaluation protocol followed by the paper.","marker":"[21]"},{"why":"Provides the base 3D Gaussian splatting representation and differentiable tile rasterization used for color and uncertainty rendering.","marker":"[12]"},{"why":"Supplies DINO features used for the boundary-aware refinement mask and the semantic uncertainty loss.","marker":"[28]"},{"why":"Photo Tourism is the main benchmark for lighting decoupling and occlusion handling.","marker":"[36]"},{"why":"NeRF On-the-go provides the graded urban occlusion scenes used to test generalization.","marker":"[31]"},{"why":"Defines the reproducible evaluation setup and dataset splits used for retraining baselines.","marker":"[13]"}],"fun_headline_variants":["NexusSplats: 65% fewer parameters, 2.7x faster training","Hierarchical light decoupling makes Gaussian splatting 2.7x faster","Nexus kernels: 65% fewer params, 2.7x speedup, SOTA PSNR","In-the-wild 3D reconstruction: 65% leaner, 2.7x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that DINO features correctly tell clear scene boundaries apart from real transient occluders in every training image: if the anisotropic 2D Gaussian foreground model in Eq. (8) is wrong, or if a pedestrian or vehicle shares visual features with the static scene, then the boundary refinement and the uncertainty loss will either suppress valid geometry or keep the occluder.","fun_headline_variants_meta":{"raw":{"variants":["NexusSplats: 65% fewer parameters, 2.7x faster training","Hierarchical light decoupling makes Gaussian splatting 2.7x faster","Nexus kernels: 65% fewer params, 2.7x speedup, SOTA PSNR","In-the-wild 3D reconstruction: 65% leaner, 2.7x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000667,"raw_usage":{"total_tokens":3040,"prompt_tokens":940,"completion_tokens":2100,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1996}},"tokens_in":556,"tokens_out":2100,"duration_ms":13965,"temperature":1.0,"reasoning_tokens":1996,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:28:59.791647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Annotate the transient objects (pedestrians, vehicles) in a held-out subset of Photo Tourism test images and compare the predicted uncertainty masks against those annotations; if the masks miss occluders that resemble the static scene, or if removing boundary-aware refinement improves PSNR on scenes with strongly non-elliptical foregrounds, then the structure-aware occlusion claim fails in that regime.","supporting_citations":[{"cited_title":"Scaffold-gs: Structured 3d gaussians for view-adaptive rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the hierarchical anchor-based Gaussian management that NexusSplats adapts into nexus kernels."},{"cited_title":"Wildgaussians: 3d gaussian splatting in the wild","cited_arxiv_id":null,"evidence_quote":"The per-Gaussian in-the-wild 3DGS baseline whose appearance embeddings, parameter count, training time, and DINO-based 2D uncertainty constitute the main comparison."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Provides the base 3D Gaussian splatting representation and differentiable tile rasterization used for color and uncertainty rendering."},{"cited_title":"Photo tourism: exploring photo collections in 3d","cited_arxiv_id":null,"evidence_quote":"Photo Tourism is the main benchmark for lighting decoupling and occlusion handling."},{"cited_title":"Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild","cited_arxiv_id":null,"evidence_quote":"NeRF On-the-go provides the graded urban occlusion scenes used to test generalization."}],"review_version":1}