{"id":"28932048-4860-4c70-823f-bb33c517efe1","arxiv_id":"2602.15355","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"DAV-GSWT selects views by diffusion-model uncertainty and hallucinates missing structure so Gaussian Splatting Wang Tiles can be made from sparse captures.","lead":"This paper proposes DAV-GSWT, a pipeline that uses diffusion-model disagreement to pick which camera views to capture and then stitches the rebuilt 3D scene into seamless Gaussian Splatting Wang Tiles from sparse imagery. It claims near-exhaustive visual quality with far fewer views, but it ships no code or data, lacks a same-budget comparison to prior tile methods, and its theoretical guarantees are asserted rather than proved.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Internal contradiction: stated budgets give 68 views (8 + 3×20) vs 200-view exhaustive baseline, a 2.9× reduction, not the claimed 'order of magnitude' in §4.8.","rationale":"The reader's weakest assumption—that diffusion ensemble disagreement is monotonically correlated with reconstruction error—is also a genuine concern, but the numerical contradiction is more direct and can be verified without rerunning experiments. It directly undermines the paper's strongest empirical claim, which is the basis for the data-efficiency contribution. The paper itself even reports p=0.18 against the 200-view baseline, so the quality preservation is statistically inconclusive, and no random-selection baseline at the same 68-view budget is provided to isolate the active-sampling benefit. If the 'order of magnitude' phrasing were corrected to 'about threefold,' the central claim would shrink substantially, and the remaining evidence would still lack a fair active-vs-random comparison. I therefore agree with the reader's REJECT verdict, though I would locate the primary defect in the internally inconsistent view-count claim rather than solely in the uncertainty-correlation assumption. The reader mentioned this contradiction in the rationale but did not make it the weakest assumption, hence 'partial' agreement.","tokens_in":14250,"tokens_out":5613,"duration_ms":56808,"concrete_test":"Recalculate the total views actually used in Table 4: start from N_init=8, k=20/iteration, T=3 iterations (Algorithm 1; §3.6), giving 68 views. Compare this to the 200-view exhaustive baseline reported in Table 4. If 200/68 ≈ 2.94 (not ≈10), then §4.8's 'one order of magnitude fewer captured views' is unsupported and must be corrected; this check can be done from the manuscript and the authors' experiment script.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's headline quantitative claim—'roughly one order of magnitude fewer captured views' (§4.8)—is contradicted by its own stated budgets. §4.2 says DAV starts from 8 views and uses the reported capture budgets; Algorithm 1 / §3.6 / §4.4 set per-iteration budget k=20 and iterations T=3. Total acquired views is therefore 8 + 3×20 = 68. The exhaustive comparison in Table 4 is a 200-view baseline, and the synthetic datasets have 100 views. The reduction is 200/68 ≈ 2.9× (or 100/68 ≈ 1.5×), not 'one order of magnitude' (10×). Thus the central data-efficiency claim is quantitatively overstated by roughly a factor of three. Additionally, the only statistical support in Table 4 shows ΔPSNR = 0.09 dB with p=0.18 (not significant), so the paper does not demonstrate that 68 views match 200 views beyond sampling noise. Without an exact per-configuration view count or a corrected statement, the strongest empirical claim in the abstract/intro and §4.8 is unsupported as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DAV-GSWT, a framework for data-efficient Gaussian Splatting Wang Tiles that combines a pre-trained diffusion prior (Zero-1-to-3) with active view sampling. An uncertainty estimator based on latent-space 2-Wasserstein divergence and LPIPS scores ranks candidate viewpoints, and the top-k are physically captured and fused into the Gaussian field over T iterations. The refined field is partitioned into Wang tiles with semantic-aware seam optimization and uncertainty-guided LOD caching. The central claim is that this pipeline achieves near-exhaustive reconstruction quality with roughly one order of magnitude fewer captured views, supported mainly by Table 4, which reports PSNR 29.41±0.08 for the full method versus 29.50±0.07 for an exhaustive 200-view baseline.","tokens_in":14641,"tokens_out":4246,"duration_ms":46305,"significance":"If substantiated, the method would address a real bottleneck in large-scale procedural rendering: reducing the dense-view requirement of Gaussian Splatting Wang Tiles. The paper contains useful engineering ideas—combining W2 latent disagreement with LPIPS, semantic-aware seam energy, and uncertainty-guided caching—and reports substantial performance details. However, the current evidence does not support the headline data-efficiency claim: the stated budgets imply a 2.9x reduction, not an order of magnitude, and the only statistical comparison to the exhaustive baseline is not significant (p=0.18). The absence of a same-budget GSWT baseline and the lack of code/data/protocol details further undermine the contribution.","major_comments":[{"comment":"The claim of 'roughly one order of magnitude fewer captured views' is contradicted by the manuscript's own stated budget. §3.6/§4.4 set k=20, T=3, and §4.2 states DAV starts from 8 views, giving 8+3×20=68 total views. Against the 200-view exhaustive baseline in Table 4 this is a 2.9x reduction; against the 100-view synthetic datasets it is 1.5x. Moreover, Table 4 reports p=0.18 for the full method vs. the 200-view exhaustive baseline, so the claimed quality preservation is not statistically significant. The headline data-efficiency claim must be corrected to the actual reduction, and the significance test must be reported honestly.","section":"§4.8, §3.6, §4.4"},{"comment":"The central ablation table lacks a same-budget GSWT baseline. To establish that active diffusion-guided selection, rather than simply having any 68-view set, is responsible for the result, the authors must compare against GSWT trained with 68 views sampled randomly or uniformly. Figure 7 mentions 'random' strategies but no corresponding quantitative table is provided. Without this control, the active-selection mechanism is not validated; the observed PSNR could be attributable to the additional views alone or to the diffusion refiner. This is a load-bearing omission for the paper's core claim.","section":"Table 4, §4.8"},{"comment":"Several theoretical statements are asserted without proof. A.1 claims a proof that ulat is monotonically correlated with ground-truth reconstruction error under 'locally convex' latent-manifold assumption, but no proof follows; the assumption itself is not established for unbounded terrain. A.2 asserts submodularity and monotonicity of the aggregate uncertainty U without demonstration, and the greedy approximation ratio is invoked without the required conditions. A.3 states submodularity of the seam energy without proof. A.5 states the concentration bound ||I_syn - I_gt||_inf ≤ σ^2 sqrt(2 log(1/δ)) without derivation; the σ^2 scaling is also dimensionally suspect. Since these results are used to justify the active-selection and hallucination guarantees, they must either be proven rigorously or removed and replaced with empirical validation.","section":"Appendix A.1–A.3, A.5"},{"comment":"The evaluation protocol is underspecified. For real scenes, no ground-truth test views are described, yet PSNR values are reported; it is unclear whether PSNR is computed against held-out frames, reconstructed images, or synthetic renderings. The paired t-test in Table 4 lacks details on the paired observations (e.g., per-scene or per-seed values). The paper also states no code or data will be released, which hinders reproducibility and verification of the central claims. The authors should specify the exact test-view protocol, provide per-scene results, and make at least the datasets and baseline implementations available.","section":"§4.2, §4.6, Table 4"}],"minor_comments":[{"comment":"Equation (5) appears to define W(s,t) as a ratio of two weighted sums, which is unusual for a graph-cut pairwise cost. Appendix Eq. (11) writes a similar expression with numerator and denominator swapped. Please clarify the intended weighted combination and ensure the two equations are consistent.","section":"§3.4, Eq. (5)"},{"comment":"The caption states 'p <0.01 vs. next best in each block,' but the exhaustive baseline row reports p=0.18. Clarify whether the p-values are two-tailed tests against the full method and why the exhaustive row is excluded from the '<0.01' statement.","section":"Table 4"},{"comment":"Figure 1 mentions 'image-space LPIPS gradients' and 'latent 2-Wasserstein divergence W2(Z)', but the text does not clearly explain how these two uncertainty terms are combined recursively. Also, the arrow from G_t-1 to the diffusion model is not described in the text.","section":"§3.2, Figure 1"},{"comment":"The blending function α(d) is claimed to achieve C1 continuity at boundary limits when Δ is sufficiently large, but the piecewise-linear form is only C0. Please check the smoothness claim.","section":"Appendix A.4, Eq. (12)"},{"comment":"The abstract claims the framework 'significantly reduces the required data volume' without reporting a statistical test on view counts. Please qualify this statement with the actual reduction and confidence.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem, and the combination of active view selection with diffusion priors for Gaussian Splatting Wang Tiles is plausible. However, the current manuscript's central quantitative claim is unsupported as written: the stated experimental budget contradicts the 'order-of-magnitude' claim, the key ablation lacks a same-budget baseline, and the theoretical appendix asserts results without proofs. These are fixable in principle by correcting the claims, adding the missing baselines, and providing full proofs or empirical validation, but they require substantial additional work. I would not recommend acceptance in the present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core system — diffusion-disagreement active view selection feeding Gaussian Splatting Wang tiles — is a genuine integration. The W2+LPIPS score is a small variation on ensemble disagreement, but the loop design is coherent and the rendering/caching side looks like solid engineering. The 2AFC seam study is a nice touch. But the headline data-efficiency claim does not survive contact with the paper's own numbers.\n\nAlgorithm 1 and §4.4 set k=20, T=3, starting from 8 views, giving 68 total views. §4.8 says 'roughly one order of magnitude fewer' than the 200-view exhaustive baseline. That's 2.9x, not 10x — and against the 100-view synthetic sets it's 1.5x. The direct comparison in Table 4 shows 68-view PSNR 29.41±0.08 vs 200-view 29.50±0.07, p=0.18, a null result. So the paper does not show sparse capture matches dense capture; it shows they're within noise. There's also no same-budget baseline against the prior GSWT method, no random or curiosity selection baseline, and no code or data. These are not minor omissions.\n\nThe theory appendix is assertion-heavy. Submodularity, monotonic convergence, and the diffusion error bound are stated without derivation, and Eq. (13)'s σ² scaling looks off by inspection. The monotonicity of the uncertainty score with reconstruction error is the load-bearing assumption; the paper 'proves' it only under a local-convexity assumption that is unlikely to hold for unbounded terrain. The experiments are too weak to rescue that.\n\nI'd still send this to peer review rather than desk reject. The idea has merit and referees could push the authors to fix the accounting, add proper baselines, and release code and data. Right now I would not cite it as evidence that active sampling gives an order-of-magnitude data reduction, because it doesn't.","headline":"The integration is real, but the claimed order-of-magnitude view reduction is contradicted by the paper's own budget math.","tokens_in":15123,"tokens_out":4249,"would_cite":false,"duration_ms":42223,"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":"DAV-GSWT claims that diffusion-driven active view selection reconstructs seamless Gaussian Splatting terrain tiles from roughly an order of magnitude fewer captured views, matching exhaustive capture quality.","keywords":["3D Gaussian Splatting","Wang Tiles","active view selection","uncertainty quantification","diffusion priors","sparse-view reconstruction","procedural terrain generation","real-time rendering"],"falsifier":"Measure per-pose reconstruction error on a held-out test set after adding each candidate view to the field, and correlate it with the paper's uncertainty score u(θ) across a dense pose grid on a real unbounded-terrain scene. If the Spearman rank correlation is zero or negative — or if random and uniform view selection at the same budget match the active loop's PSNR — the monotonic-correlation premise fails and the data-efficiency gain vanishes. A cheaper version sits in the paper's own tables: the full method (29.41 ± 0.08) and the exhaustive 200-view baseline (29.50 ± 0.07) are statistically","tokens_in":14117,"feed_emoji":"🏔️","tokens_out":9642,"duration_ms":89305,"temperature":0.7,"pith_summary":"DAV-GSWT sets out to remove the data bottleneck in Gaussian Splatting Wang Tiles, the tiling scheme that turns 3D Gaussian exemplars into seamless infinite terrain. Its central claim: a pre-trained diffusion prior can guide capture, because the disagreement among stochastic predictions of a candidate view signals where the Gaussian field most needs new measurements. Selecting the top-k disagreeing viewpoints and fusing those images yields tile-ready fields at a fraction of the usual capture budget — the paper reports parity with a 200-view exhaustive scan at roughly one order of magnitude fewer views. If right, vast photorealistic worlds could be built from a handful of drone or handheld images, with the active loop also driving seamless tile boundaries and real-time LOD rendering.","feed_headline":"Roughly ten times fewer views build seamless 3D terrain tiles","feed_subtitle":"An active-capture loop ranks views by diffusion uncertainty, matching a 200-view scan from a tenth of the images.","key_machinery":"The engine of the method is a viewpoint uncertainty score: after M stochastic forward passes of the diffusion prior with attention dropout, each spatial location of the latent map is treated as a diagonal Gaussian, and the pairwise 2-Wasserstein divergence across samples is summed over space and channels, then combined with a perceptual-similarity (LPIPS) term between two decoded images. This single scalar ranks candidate camera poses; the top-k drive the physical capture budget. Work it does: it converts an expensive, passive data-collection process into a closed active loop in which each captured view is chosen to shrink the disagreement of the generative prior. A second named mechanism, t","core_discovery":"The paper seeks to establish that viewpoint selection for sparse reconstruction can be driven by the internal disagreement of a pre-trained diffusion prior rather than by geometric heuristics. Starting from a coarse Gaussian field built by a fast structure-from-motion pass, the method runs a single-image-to-3D diffusion model with attention dropout at each candidate pose, forming an ensemble of latent samples; the pairwise 2-Wasserstein divergence of those latents, plus a perceptual-similarity term, scores every candidate pose, and the top-k are sent to a physical capture agent, iterated a few times. The paper reports 29.41 ± 0.08 dB PSNR from this loop (about 60 captured views) against 29.5","pith_inferences":["My inference: the same disagreement-driven active loop should transfer to other procedural content — building facades, road networks, foliage — wherever a diffusion prior can forecast the next view; the paper only demonstrates terrain tiles.","My inference: if diffusion disagreement does track reconstruction error, the method points to a general pattern for low-data 3D sensing — let a generative prior propose candidate measurements and a cheap sensor verify them — which could matter for field robotics and rapid exploration beyond virtual worlds.","My inference: with the default budget (8 initial views plus 3 iterations of 20 captures = 68 views) against a 200-view baseline, the realized view reduction is closer to threefold; the 'order of magnitude' phrasing is the paper's summary, not what the default arithmetic itself shows."],"forward_implications":["Capture budgets for terrain exemplars drop by roughly an order of magnitude: about 60 active views match a 200-view exhaustive scan within statistical noise (p = 0.18), so seamless infinite-terrain rendering no longer requires dense capture.","Latent-space ensemble disagreement (W2 + LPIPS) is a better view-selection signal than image-space gradient uncertainty, cutting uncertainty-evaluation cost and improving seam quality.","Removing the semantic weight from the seam graph cut costs about 0.8 dB PSNR, so semantic awareness materially contributes to tile-boundary fidelity.","Interactive rendering survives the sparse pipeline: per-frame render times stay in the 5–15 ms range with millions of splats, supported by uncertainty-guided pre-sorted caching and continuous LOD blending.","The framework works on real drone footage as well as synthetic terrains, so the data-efficiency gain is not confined to controlled renderings."],"fun_headline_variants":["Diffusion uncertainty picks views: 10× fewer images for 3D tiles","Active sampling from diffusion priors cuts capture views 10×","Diffusion-guided view selection: 200-view quality from ~60 shots","Seamless terrain tiles from a tenth of the images via diffusion disagreement","Diffusion disagreement steers capture: 3D tiles at 10× less data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire data-efficiency gain rests on the assumption that the disagreement among stochastic diffusion predictions at a viewpoint is a reliable proxy for how much the Gaussian field would actually improve from capturing that view — an assumption the paper proves only when the diffusion prior's latent manifold is locally convex, a condition it does not establish for unbounded terrain (Appendix A.1).","fun_headline_variants_meta":{"raw":{"variants":["Diffusion uncertainty picks views: 10× fewer images for 3D tiles","Active sampling from diffusion priors cuts capture views 10×","Diffusion-guided view selection: 200-view quality from ~60 shots","Seamless terrain tiles from a tenth of the images via diffusion disagreement","Diffusion disagreement steers capture: 3D tiles at 10× less data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3099,"prompt_tokens":685,"completion_tokens":2414,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":2329}},"tokens_in":429,"tokens_out":2414,"duration_ms":15697,"temperature":1.0,"reasoning_tokens":2329,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T22:51:57.954012+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure per-pose reconstruction error on a held-out test set after adding each candidate view to the field, and correlate it with the paper's uncertainty score u(θ) across a dense pose grid on a real unbounded-terrain scene. If the Spearman rank correlation is zero or negative — or if random and uniform view selection at the same budget match the active loop's PSNR — the monotonic-correlation premise fails and the data-efficiency gain vanishes. A cheaper version sits in the paper's own tables: the full method (29.41 ± 0.08) and the exhaustive 200-view baseline (29.50 ± 0.07) are statistically","supporting_citations":[],"review_version":2}