{"id":"42494773-559b-41d9-812e-0879cc04c51e","arxiv_id":"2506.22973","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Learned per-splat Beta-distributed confidence scores enable test-time pruning of 3D Gaussian splats with minor quality loss, and average confidence is proposed as a scene quality metric.","lead":"This paper trains 3D Gaussian Splatting scenes to assign each splat a confidence score, modeled as the mean of a Beta distribution, and uses these scores to prune low-confidence splats at test time. The authors report substantial scene size reduction with little quality loss, and propose average confidence as a scene quality metric.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No baseline against opacity-based pruning; the claimed 'better signal' for identifying redundant splats is untested, so the central compression-fidelity claim is not established.","rationale":"The reader's weakest_assumption focuses on the saliency ranking loss (Eq. 7) and its use of image-space gradients as a proxy for splat importance. That is a reasonable internal concern, but the more load-bearing issue is external and easily testable: the paper never establishes that the learned confidence outperforms simple opacity-based pruning. The method modulates opacity with confidence (Eq. 8), and the sparsity loss drives confidence down, so a plausible degenerate outcome is that confidence becomes a monotonic transformation of effective opacity. If so, thresholding by confidence is essentially thresholding by opacity, and the claimed novelty provides no benefit over a well-known heuristic. The missing baseline is therefore the single most important check on the central claim. The paper has strengths: the method is architecture-agnostic, it is plug-and-play on MCMC-GS, and the reported quality-retention curves are plausible. The SQR and ACS metrics are secondary contributions and, while underspecified, do not undermine the compression claim as directly. Our concern is not an internal contradiction; it is an unsubstantiated comparative claim. The proposed concrete test would settle it cleanly. If confidence-based pruning outperforms opacity-based pruning, the paper's contribution is validated; if not, the central claim fails. Since the reader already requested baselines and made the verdict CONDITIONAL, our concern does not change the verdict—it sharpens the specific baseline needed.","tokens_in":163,"tokens_out":2875,"duration_ms":46214,"concrete_test":"Using the released or re-trained 'Ours @Base3DGS @100%' model and the baseline 3DGS model for each scene in Table 4, perform post-training pruning using four ranking criteria: (a) learned confidence ci, (b) effective opacity σ(oi)·ci, (c) original opacity σ(oi) before confidence modulation, and (d) opacity of the baseline 3DGS model. Sweep thresholds so that the retained-splat counts match (e.g., 90%, 75%, 50%, 30% of the original count). Plot PSNR, SSIM, and LPIPS versus retained splat count for each criterion. If criterion (a) does not dominate criterion (b) or (c) at equal retained counts, the claim that confidence provides a better pruning signal is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 5.2) is that confidence scores provide 'a better signal' for pruning than simple alternatives, and that pruning to roughly half the splats leaves PSNR essentially unchanged. For this claim to hold, the learned confidence must outrank trivial heuristics such as opacity-based pruning. The paper never tests this. In Table 4, the only comparisons are RadSplat and Mini-Splatting, both of which are training-time pruning methods with different pipelines; no comparison is made to pruning the same trained model by opacity. This is especially important because Eq. 8 modulates opacity by confidence (oeffective_i = σ(ooriginal_i)·ci), so confidence could simply become a learned re-scaling of opacity. If pruning by confidence is no better than pruning by opacity on the same model, then the method's contribution reduces to the extra training regularizers (L_sparse, L_ent, L_sal) plus opacity pruning, and the claimed 'better signal' evaporates. The saliency ranking loss (Eq. 7) is the mechanism intended to make confidence more meaningful than opacity, but its effect can only be verified by direct comparison. Without this baseline, the headline result is consistent with a much simpler explanation: any reasonable pruning heuristic that removes low-opacity, low-impact splats would achieve similar trade-offs.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes 'Confident Splatting', a post-training compression scheme for 3D Gaussian Splatting. For each splat, two learnable parameters define a Beta distribution, and the distribution mean is used as a per-splat confidence score. The confidence parameters are optimized jointly with the reconstruction loss through three regularizers: a sparsity loss, a negative-entropy loss, and a pairwise saliency ranking loss. During rasterization, the effective opacity is multiplied by the confidence score, and after training a user can prune splats below a chosen confidence threshold. The authors also propose the average confidence score (ACS) as a scene-quality metric and introduce a custom YouTube-derived dataset. Experiments are reported on MipNeRF-360, Tanks&Temples, and an Eiffel Tower scene, comparing against RadSplat and Mini-Splatting and demonstrating transferability to MCMC-GS.","tokens_in":12981,"tokens_out":6996,"duration_ms":74851,"significance":"If the central claims are correct, the method would be a simple, pipeline-agnostic test-time pruning mechanism with a user-controlled compression knob, and the released code and data would support further work on confidence-guided editing and uncertainty-aware rendering. The idea of modeling confidence with a Beta distribution and tying it to opacity is reasonable and worth investigating. However, the current evidence does not establish that the learned confidence ordering is better than plain opacity-based pruning, and several of the reported numbers weaken the headline claim that half the splats can be removed with negligible PSNR loss. The work is a plausible candidate for publication after a substantial comparative evaluation.","major_comments":[{"comment":"The central claim that the learned confidence is 'a better signal for identifying redundant or low-impact splats' (Section 5.2) is never tested against the simplest baselines. No comparison is made between pruning by confidence and pruning by opacity, by gradient magnitude, or by random order on the same trained model at identical retained-splat counts. This matters because Eq. (8) multiplies the original opacity by c_i, so confidence can be viewed as a learned rescaling of opacity; without such a baseline, the observed pruning success may be entirely consistent with opacity-based redundancy. I request a same-model ablation: rank splats by confidence, opacity, gradient magnitude, and random scores, prune each to the same retained counts, and report PSNR, SSIM, and LPIPS. In addition, ablating L_sparse, L_ent, and L_sal separately is necessary to show that the ranking loss in Eq. (7) is what improves the ordering; as written, L_sparse directly minimizes average confidence and already tilts the learned scores toward 'low-confidence means prunable' by construction.","section":"§4.2–4.3, §5.4, Table 4"},{"comment":"The headline statement that 'in most scenes, number of splats drops to half ... while PSNR does not change or changes with a really small tolerance' is contradicted by the paper's own numbers. In Table 4, Truck Ours@Base3DGS@~33% retains about one third of the splats but PSNR drops from 25.208 to 24.174 (−1.03 dB); Train @~39% drops from 21.797 to 20.277 (−1.52 dB); Treehill @~47% drops from 22.534 to 21.556 (−0.98 dB). At 50% retention, RadSplat achieves 25.419 PSNR on Truck and 22.488 on Treehill, so the proposed method is worse than the training-time baseline on these scenes at comparable or better sizes. The compression-fidelity claim should be restated per scene at a consistent retention level, with repeated-run variance, rather than as a global 'most scenes' assertion.","section":"§5.2, Table 4"},{"comment":"The transferability experiment to MCMC-GS only reports pruning 5% and 10% of the splats (PSNR@95% and PSNR@90%), not the roughly 50% compression advertised elsewhere. Showing near-identical PSNR when removing 5% of splats does not demonstrate that the method transfers at practically useful compression ratios. The authors should run the same threshold sweep on MCMC-GS, or explicitly limit the transferability conclusion to small pruning rates.","section":"§5.5, Table 2"},{"comment":"The proposed scene-quality metric ACS is asserted to correlate with quality, but the confidence values are trained with L_sparse, which actively minimizes their mean, and the comparison across 3DGS and MCMC-GS conflates pipeline differences in splat count and initialization. No correlation coefficient, error bars, or significance test is reported for the eight scenes. As a stated contribution, this claim should be either supported by a scatter plot of ACS against PSNR, SSIM, and LPIPS across matched scenes, or explicitly softened to an anecdotal observation.","section":"§5.6, Table 3"}],"minor_comments":[{"comment":"Figure 1 contains a typo: 'proposded' should be 'proposed'.","section":"Figure 1 caption"},{"comment":"Equation (2) writes an integral from −∞ to ∞ over a Beta density whose support is [0,1]; the limits should be 0 and 1.","section":"§4.1, Eq. (2)"},{"comment":"Table 3 lists LPIPS@Orig% = 0.7221 for Eiffel Tower Ours@MCMC-GS, which is inconsistent with 0.072 reported in Table 4; this appears to be a decimal typo and should be corrected.","section":"Table 3"},{"comment":"The text says SQR 'should be worse if the quality is low (0 at its lowest)', but the formula approaches 1/(1+0)=1 as PSNR tends to 0, not 0; the sentence and the formula need to be aligned.","section":"§5.4, Eq. (9)"},{"comment":"The supplementary Gumbel-noise experiments in Eqs. (11) and (12) use α̂ and β̂ without defining them relative to the α and β parameters in the main text.","section":"Supplementary §10"},{"comment":"The hyperparameters λ1, λ2, λ3, P, the ranking margin, and the SQR scale are not reported; without these values the experiments are not reproducible.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript proposes a plausible compression add-on, but the evaluation currently does not establish the advertised advantage over trivial pruning heuristics. The decisive missing experiment is a same-model comparison between confidence-based pruning and opacity-based pruning (and gradient-magnitude-based pruning) at matched retained-splat counts. If that comparison shows no gain, the contribution reduces to the training regularizers plus opacity pruning. I recommend requiring this baseline and the per-scene accuracy fix in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a plausible idea and decent feasibility evidence, but it never runs the one baseline that would make the central claim meaningful — pruning the same trained model by opacity. Until that is added, \"better signal\" is not established.\n\nWhat is actually new: per-splat Beta-distributed confidence (alpha, beta) rather than a scalar, a ranking-based saliency loss that tries to order splats by image-space gradient contribution, and a genuinely useful test-time threshold knob that works on top of different 3DGS variants. The MCMC-GS transfer experiments are a good sign, and the results do show that on several scenes you can drop roughly half the splats with small PSNR/SSIM/LPIPS changes. That is real evidence of feasibility, and the code/data link is a plus, though the prose is inconsistent about whether the dataset is available now or only upon acceptance.\n\nThe problems, in order of size. First, no opacity-based pruning baseline. Since Eq. 8 multiplies opacity by confidence, the learned confidence could simply be a rescaling of opacity; if thresholding confidence is no better than thresholding opacity on the same model, the contribution collapses to extra regularizers plus opacity pruning. The saliency loss is meant to prevent that, but the paper has to demonstrate it. This is the load-bearing missing experiment. Second, the sparsity loss directly minimizes average confidence, so low-confidence splats are being pushed down by construction; pruning them is partly a consequence of the objective. Third, calling this compression oversells it: you report splat counts, not bytes, SH coefficients, or quantization. Fourth, the SQR metric is arbitrary and should be replaced with standard rate-distortion curves. Finally, several hyperparameters (lambda_1..3, P, margin, scale) are not given, so reproducibility is incomplete; Table 3 also has a suspicious LPIPS entry for Eiffel Tower that needs checking.\n\nThe paper is honest about many limitations and the core mechanism is not crazy. It just needs the opacity baseline and a bitrate-aware evaluation. I'd send it to review, not desk-reject, but the referee should require those additions before acceptance.","headline":"Neat idea with a useful test-time pruning knob, but the missing opacity-pruning baseline leaves the central 'better signal' claim untested.","tokens_in":13597,"tokens_out":2620,"would_cite":false,"duration_ms":27674,"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":"After training, users can cut roughly half of a scene's 3D Gaussian splats and keep PSNR nearly unchanged.","keywords":["3D Gaussian Splatting","lossy compression","confidence scores","Beta distribution","splat pruning","novel view synthesis","scene quality assessment","test-time compression"],"falsifier":"Take a trained scene, rank splats by the paper's confidence scores and also by measured per-splat PSNR impact (remove each splat individually and record the PSNR change on held-out views). If pruning by confidence loses more PSNR than pruning by measured impact at the same retention rate, or if the rank correlation between confidence and measured impact is near zero on benchmark scenes, the central claim is refuted.","tokens_in":12525,"feed_emoji":"✂️","tokens_out":7440,"duration_ms":72946,"temperature":0.7,"pith_summary":"This paper proposes a test-time compression method for 3D Gaussian Splatting scenes. Each splat is assigned a confidence score, modeled as the mean of a learnable Beta distribution, and this score is trained alongside the scene so that low-confidence splats are the ones that matter least for the rendered image. After training, a user picks a confidence threshold and simply deletes all splats below it, with no retraining; across most tested scenes the splat count drops by roughly half while PSNR changes only within a small tolerance. The paper also claims that the scene-wide average of these confidence scores correlates with reconstruction quality, so the same numbers serve as a quality metric.","feed_headline":"Prune half a scene's Gaussian splats, keep the view","feed_subtitle":"A per-splat confidence score lets users trim 3D scenes at test time with little quality loss.","key_machinery":"The load-bearing object is a per-splat Beta distribution. For each splat $i$ the optimizer learns shape parameters $\\alpha_i$ and $\\beta_i$, and the splat's confidence is the distribution mean $c_i = \\alpha_i/(\\alpha_i + \\beta_i)$. The Beta distribution matters because it carries a full belief shape, not just a point value: the sparsity loss pushes means down, the negative-entropy loss penalizes distributions peaked at 0 or 1, and the saliency loss uses hinge ranking on pairs of splats ordered by image-space gradient magnitude. Confidence is fed into rendering by scaling each splat's opacity, $o_{\\text{eff}} = \\sigma(o_{\\text{orig}})\\,c_i$, which couples the learned scores to the reconstruction loss. At test time the same $c_i$ values are thresholded to prune splats, which is what produces the reported compression.","core_discovery":"The central claim is that confidence can be made into a learned, interpretable, per-splat property rather than a post-hoc heuristic. For each Gaussian the method fits a Beta distribution with parameters $\\alpha_i$ and $\\beta_i$, takes its mean $c_i = \\alpha_i/(\\alpha_i + \\beta_i)$ as the splat's confidence, and optimizes those parameters with three reconstruction-aware regularizers: a sparsity loss that pushes average confidence down, a negative-entropy loss that discourages distributions peaked at 0 or 1, and a ranking-based saliency loss that forces splats with higher image-space gradient contributions to receive higher confidence. The confidence is multiplied into each splat's opacity during rasterization, so it participates in training. After training, thresholding on $c_i$ yields the compression knob. On the tested scenes the paper reports that halving the number of splats leaves PSNR essentially unchanged, and that applying the same procedure on top of the MCMC-GS variant gives comparable trade-offs; the average confidence over a scene is proposed as a quality metric.","pith_inferences":["A natural next step the paper does not run is ablating the three losses: the ranking saliency loss is the only component using image-space gradients, so measuring PSNR after pruning when $L_{sal}$ is removed would isolate how much of the compression quality actually comes from the ranking signal rather than from the sparsity and entropy regularizers.","The same confidence scores could be reused for tasks the paper only names as future work: main-object extraction and floater removal. Those applications are plausible because thresholding in the provided visualizations progressively strips background and floaters, but the paper does not quantify their accuracy.","Because confidence is a continuous value, the method could support adaptive streaming or level-of-detail rendering, where distant or low-contribution splats are dropped first; this is an extension, not something the paper tests.","The ACS quality metric could be validated against human perceptual judgments or against no-reference image quality metrics on a larger set of scenes; the paper only compares a handful of scenes."],"forward_implications":["A trained scene can be compressed at deployment time by choosing a threshold, and the same scene can be served at multiple quality/size operating points without retraining.","Because the confidence parameters are added on top of an existing pipeline, the method should port to any splatting-based renderer that exposes per-splat opacity, rotation, scale, and color.","Pruning half the splats in most benchmark scenes keeps PSNR within a small tolerance, so storage and rasterization cost can be cut roughly in half before visible degradation.","The average confidence score (ACS) tracks scene quality, giving a reference-free way to compare reconstructions and to detect scenes with more floaters or artifacts.","The method introduces Splats-to-Quality Ratio (SQR) as a single number for comparing quality-compression trade-offs across methods and pruning ratios."],"supporting_citations":[{"why":"Defines the 3D Gaussian Splatting representation, training losses, and adaptive density control that the method extends and evaluates on.","marker":"[10]"},{"why":"The MCMC-GS variant used to demonstrate that the confidence module transfers to another splatting pipeline.","marker":"[11]"},{"why":"RadSplat, a train-time importance-pruning baseline that the paper compares against; it requires retraining per threshold.","marker":"[18]"},{"why":"Mini-splatting, a constrained-Gaussian baseline that the method compares against in compression tables.","marker":"[8]"},{"why":"LP-3DGS, a Gumbel-Sigmoid train-time pruning baseline whose learned pruning ratio motivates the need for a test-time knob.","marker":"[21]"},{"why":"Mip-NeRF 360 dataset (garden scene) used for the main compression figures and quality evaluations.","marker":"[3]"},{"why":"Cited to support the observation that adaptive filtering can grow millions of splats, many of which are not crucial to final fidelity.","marker":"[14]"}],"fun_headline_variants":["Learned confidence scores prune half of 3D splats","Beta-distributed confidence guides splat compression","Trim 3D scenes to half size with little quality loss","A per-splat confidence score yields lossy compression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The saliency ranking loss assumes that a splat's contribution to image-space gradients of the reconstruction loss is a reliable measure of how much that splat matters for visual quality; if that proxy is inaccurate, the learned confidence ordering may be no better than simple opacity-based pruning.","fun_headline_variants_meta":{"raw":{"variants":["Learned confidence scores prune half of 3D splats","Beta-distributed confidence guides splat compression","Trim 3D scenes to half size with little quality loss","A per-splat confidence score yields lossy compression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1231,"prompt_tokens":900,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":266}},"tokens_in":516,"tokens_out":331,"duration_ms":4029,"temperature":1.0,"reasoning_tokens":266,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:53:36.348438+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained scene, rank splats by the paper's confidence scores and also by measured per-splat PSNR impact (remove each splat individually and record the PSNR change on held-out views). If pruning by confidence loses more PSNR than pruning by measured impact at the same retention rate, or if the rank correlation between confidence and measured impact is near zero on benchmark scenes, the central claim is refuted.","supporting_citations":[{"cited_title":"3d gaussian splat- ting as markov chain monte carlo","cited_arxiv_id":null,"evidence_quote":"The MCMC-GS variant used to demonstrate that the confidence module transfers to another splatting pipeline."},{"cited_title":"Mini-splatting: Repre- senting scenes with a constrained number of gaussians","cited_arxiv_id":null,"evidence_quote":"Mini-splatting, a constrained-Gaussian baseline that the method compares against in compression tables."},{"cited_title":"Mip-nerf 360: Unbounded anti-aliased neural radiance fields","cited_arxiv_id":null,"evidence_quote":"Mip-NeRF 360 dataset (garden scene) used for the main compression figures and quality evaluations."},{"cited_title":"Maskgaussian: Adaptive 3d gaussian representation from probabilistic masks","cited_arxiv_id":null,"evidence_quote":"Cited to support the observation that adaptive filtering can grow millions of splats, many of which are not crucial to final fidelity."}],"review_version":1}