{"id":"dfa8d460-2796-4bea-922d-b81d816f6041","arxiv_id":"2603.23297","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"WD-R, a lightly regularized Wasserstein Distortion loss, is preferred by humans 2.3× over the original 3DGS loss and 1.5× over Perceptual-GS while matching or reducing splat count and generalizing to other 3DGS frameworks and compression.","lead":"A regularized Wasserstein Distortion loss (WD-R) can replace the usual pixel losses in 3D Gaussian Splatting and produce sharper, more textured novel views that humans prefer, without needing more Gaussians. The result is a drop-in training change that also cuts bitrate roughly in half for compressed 3DGS scenes at matched perceptual quality.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified; the matched-budget human and metric gains for WD-R are robustly evidenced.","rationale":"The strongest claim is purely empirical and is supported by the large-scale human study (39 k ratings, Bayesian Elo), consistent SOTA perceptual metrics under matched or lower resource budgets, visual texture recovery, anisotropy analysis, and transfer to Mip-Splatting, Scaffold-GS and Comp-GS. The free parameters flagged by the reader are ablated (Apps. B–C) and do not introduce internal inconsistency or unsupported extrapolation; they are the usual price of fair comparison when densification depends on gradient magnitude. No hidden assumption about the 3-D geometry or the human visual system is required beyond the 2-D WD definition already validated in the image-compression literature. Consequently the ACCEPT verdict with high confidence stands; the result is useful and the evidence is adequate for the claim as stated.","tokens_in":26474,"tokens_out":563,"duration_ms":23872,"concrete_test":"Retrain WD-R on the five outdoor Mip-NeRF 360 scenes with a single fixed γ equal to the median of Tab. 4 (no per-dataset retuning) and recompute LPIPS/DISTS plus a 2 000-trial Elo subset versus the original loss; if the preference ratio remains ≥1.8× and metrics stay superior at the resulting (unmatched) #G, the drop-in claim is even stronger; a collapse below 1.2× would show that γ matching is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (fixed σ=4, per-dataset γ for densification matching, fixed β=1/0.09, 3–5k warm-up) is real but ordinary and does not undercut the central claim. Eq. (1) and Tab. 4 make the γ calibration explicit so that #G (or model size) stays comparable or lower; Fig. 12 confirms the WD term still supplies the dominant gradient (mean ratio ~1.6); pure-WD web artifacts are acknowledged (Fig. 6) and suppressed by the modest L_orig regularizer without capacity inflation. Gains hold across 21 scenes, four datasets, three frameworks and the Comp-GS RD curves (Tabs. 1–3, Figs. 4,7,8,11), all under the same protocol. These are standard controls for adaptive densification, not an untested regime that would reverse the 2.3×/1.5× preference or the ~50 % bitrate claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that 3D Gaussian Splatting can be made substantially more perceptually faithful by replacing the usual L1+SSIM training objective with a drop-in 2D distortion loss, without changing densification, pruning, or architecture. After comparing the original loss, a composite of L1/L2/MS-SSIM/LPIPS, Wasserstein Distortion (WD) in VGG feature space, and a regularized variant WD-R (WD plus a modest original-loss term), the authors identify WD-R as the strongest performer. Under matched or lower splat counts (or model sizes), WD-R improves LPIPS, DISTS, FID and CMMD on 21 scenes from four datasets, is preferred by human raters more than 2.3\times over the original loss and 1.5\times over Perceptual-GS (39,320 pairwise ratings, Bayesian Elo), transfers to Mip-Splatting and Scaffold-GS (1.8× and 3.6× preference), and yields roughly 50% bitrate savings at matched perceptual quality under Comp-GS rate–distortion optimization.","tokens_in":26735,"tokens_out":1081,"duration_ms":11767,"significance":"If the results hold, the work cleanly separates perceptual modeling from 3DGS algorithmic design and shows that a single, architecture-agnostic loss change can recover fine texture at equal or lower capacity. The large-scale, pre-screened human preference study is a first for 3DGS and substantially strengthens the claim beyond automatic metrics. Successful transfer to anti-aliased, structured, and compressed pipelines, together with public code and project page, makes the contribution immediately usable and falsifiable. The free parameters (per-dataset γ, fixed β and σ, warm-up) are ordinary for adaptive densification and are documented; they do not reduce the central claim to a tautology.","major_comments":[{"comment":"§3.1–3.2, Eq. (1) and Appendix A.2 / Table 4: the global scale γ is tuned per dataset so that adaptive densification produces comparable #G (or model size). While this is the right control and is made explicit, the manuscript should state more clearly how sensitive the human-preference and metric rankings are to modest mis-calibration of γ (e.g., a small grid around the chosen values on one indoor and one outdoor scene). Without that, readers cannot judge how much of the 2.3×/1.5× preference is locked to the exact operating point.","section":null},{"comment":"§4.1 and Fig. 6: pure WD can produce web-like artifacts under tight splat budgets; WD-R suppresses them via a fixed β=1/0.09. Fig. 12 shows that the WD gradient still dominates on average (~1.6×), but the paper would be stronger if it quantified residual artifact frequency (or a simple automatic detector) across all 21 scenes rather than a single Bicycle example, so that the residual failure mode of the regularizer is bounded.","section":null}],"minor_comments":[{"comment":"Table 1 / Table 8: PSNR/SSIM drop under WD/WD-R is expected (perception–distortion trade-off) but should be flagged more explicitly in the main text so readers do not misread the tables as a pure win on every axis.","section":null},{"comment":"Appendix A.1: the ~4.5× training-time overhead of WD is acknowledged; a short note on whether feature caching or lower-resolution VGG maps can reduce it would help practitioners.","section":null},{"comment":"§3.1: the choice of constant σ=4 is justified by ablation (Table 6) and human ratings, but a one-sentence pointer to the adaptive-σ experiments (Appendix C.2) in the main text would avoid the impression that only a single σ was tried.","section":null},{"comment":"Fig. 1 caption and Elo figures: the “2.4× vs. original” style ratios are helpful; stating the exact Elo-to-odds conversion used (Bradley–Terry) once would make the numbers fully self-contained.","section":null},{"comment":"Minor notation: L_orig is defined with coefficients 0.8 L1 + 0.2 SSIM in Appendix A.2; the same coefficients should appear when the original loss is first introduced in §2.2 / §3.1 for consistency.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The central claim is well-supported by multi-dataset metrics, a large human study, and cross-framework transfer; the free parameters are ordinary and disclosed. I see no circularity that would reverse the preference or bitrate claims. Minor revision is appropriate mainly to tighten the γ-sensitivity and residual-artifact discussion; I would not require new experiments beyond a small grid or a short quantitative artifact count. Fit for a top CV venue is strong."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: replace the usual L1+SSIM with a regularized Wasserstein Distortion (WD-R) and you get clearly better human-preferred novel views at matched or lower splat/model budgets, plus a clean ~50% bitrate win under Comp-GS. That is immediately usable for the 3DGS crowd.\n\nWhat is new is not WD itself (Ballé et al. CVPR 2025, Qiu et al.), but the systematic application, the modest L_orig regularizer that kills the web-like artifacts pure WD produces under tight budgets, the first large-scale 3DGS preference study (39k pairwise ratings, Bayesian Elo, pre-screened), and the demonstration that the same drop-in improves Mip-Splatting (1.8×) and Scaffold-GS (3.6×) without touching densification or architecture. They keep #G or model size comparable by scaling γ per dataset (Table 4), show the WD gradient still dominates (mean ratio ~1.6, Fig. 12), and report LPIPS/DISTS/FID/CMMD wins across 21 scenes. The anisotropy analysis (erank histograms) gives a plausible geometric story for why textures improve without more primitives. Code and project page are promised; that helps.\n\nSoft spots are real but ordinary. γ is hand-tuned per dataset so densification lands in the right ballpark, β and σ=4 are fixed, and a 3–5k warm-up is required. Adaptive-σ experiments did not beat the constant map enough to justify the complexity. PSNR/SSIM drop, as expected when you stop optimizing them. None of this undercuts the central claim: under the stated matched-budget protocol the preference and metric gains hold. Circularity risk is low—evaluation LPIPS uses a different backbone, FID/CMMD use Inception/CLIP.\n\nThis is for anyone training or compressing 3DGS who cares about perceived quality. It deserves a serious referee. I would accept it, cite the human-study numbers and the drop-in recipe, and bring it to reading group.","headline":"Solid drop-in perceptual loss for 3DGS with a real human study and clean transfer; free parameters are ordinary, not fatal.","tokens_in":27425,"tokens_out":533,"would_cite":true,"duration_ms":6797,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A drop-in regularized Wasserstein Distortion loss makes 3D Gaussian Splatting reconstructions preferred by humans more than twice as often as the standard loss, without needing more Gaussians.","keywords":["3D Gaussian Splatting","perceptual optimization","Wasserstein Distortion","human preference study","novel view synthesis","scene compression","drop-in loss"],"falsifier":"A head-to-head human preference study on held-out large-scale outdoor scenes in which WD-R, under identical splat budgets and without further hyper-parameter retuning, is no longer preferred over the original L1+SSIM loss or over Perceptual-GS.","tokens_in":27350,"feed_emoji":"👁️","tokens_out":729,"duration_ms":7092,"temperature":0.7,"pith_summary":"3D Gaussian Splatting produces novel views that people look at, yet most pipelines still train with ad-hoc pixel losses such as L1 plus SSIM and therefore tend to blur fine texture. The authors search over families of distortion losses and evaluate them with the first large-scale human preference study for the task (nearly 40 000 pairwise ratings). A regularized form of Wasserstein Distortion, called WD-R, wins: it recovers high-frequency texture while keeping or even lowering the Gaussian count, is chosen by raters more than 2.3 times as often as the original loss and 1.5 times as often as the previous best perceptual method, and improves standard perceptual metrics across indoor and outdoor scenes. The same loss can be dropped into anti-aliased and structured variants of the pipeline and into a rate-constrained compression setting, where it yields roughly 50 percent bitrate savings at matched perceptual quality. The result shows that a better 2-D perceptual loss alone is enough to raise visual quality without redesigning densification, pruning or the underlying representation.","feed_headline":"Drop-in loss doubles human preference for 3D Gaussian views","feed_subtitle":"Regularized Wasserstein Distortion recovers texture without extra Gaussians and cuts compression bitrate by half","key_machinery":"WD-R: Wasserstein Distortion (local mean and standard-deviation matching in VGG feature space with fixed pooling size σ=4) lightly regularized by a scaled copy of the original L1+SSIM term, used as a pure drop-in replacement for the 2-D distortion loss that drives both parameter updates and adaptive densification.","core_discovery":"Replacing the usual L1+SSIM training objective of 3D Gaussian Splatting with a regularized Wasserstein Distortion loss (WD-R) produces novel-view reconstructions that human raters prefer more than 2.3 times as often as the original loss and 1.5 times as often as the prior best perceptual method, while matching or reducing the number of Gaussians and improving LPIPS, DISTS and FID on multiple datasets.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["WD-R drop-in beats 3DGS loss 2.3x in human ratings","Regularized Wasserstein preferred 2.3x for sharper 3DGS","WD-R recovers textures in 3DGS without more Gaussians","Replace L1+SSIM with WD-R: 1.5x edge over Perceptual-GS","WD-R lifts 3DGS views 2.3x preferred, halves bitrate"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A single fixed pooling size, a hand-tuned global loss scale per dataset, a fixed regularization weight, and a short pixel-loss warm-up are enough to keep the Wasserstein term dominant and to suppress web-like artifacts on the scenes that were tested.","fun_headline_variants_meta":{"raw":{"variants":["WD-R drop-in beats 3DGS loss 2.3x in human ratings","Regularized Wasserstein preferred 2.3x for sharper 3DGS","WD-R recovers textures in 3DGS without more Gaussians","Replace L1+SSIM with WD-R: 1.5x edge over Perceptual-GS","WD-R lifts 3DGS views 2.3x preferred, halves bitrate"]},"model":"grok-4.5","effort":"low","cost_usd":0.004292,"raw_usage":{"total_tokens":1356,"prompt_tokens":865,"num_sources_used":0,"completion_tokens":119,"cost_in_usd_ticks":42920000,"prompt_tokens_details":{"text_tokens":865,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":372,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":865,"tokens_out":119,"duration_ms":3971,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T20:18:59.174007+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A head-to-head human preference study on held-out large-scale outdoor scenes in which WD-R, under identical splat budgets and without further hyper-parameter retuning, is no longer preferred over the original L1+SSIM loss or over Perceptual-GS.","supporting_citations":[],"review_version":1}