{"id":"3fdb52ff-c83d-4e17-8013-e4c7412e322b","arxiv_id":"2412.12734","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Replacing the per-splat solid color in 2D Gaussian Splatting with a small per-splat texture grid improves reconstruction sharpness and quality on image fitting and 3D scene reconstruction benchmarks.","lead":"This paper upgrades 2D Gaussian Splatting, a method for reconstructing 3D scenes from photos, by giving every flat surface patch its own small texture instead of a single solid color. The result is sharper reconstruction of textured areas such as posters, marble, and vegetation, at similar or better quality with fewer primitives.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-splat textures may overfit training views: the only held-out evidence is one face test camera and near-training NeRF360 validation views, so the central novel-view claim is not yet retired.","rationale":"The central claim is that replacing solid colors with per-splat textures improves both fixed-primitive image fitting and adaptive-primitive 3D reconstruction. The first part is demonstrated by overfitting single images, where more parameters naturally lower the loss; this is not evidence for novel-view generalization. The second part rests on NeRF360 validation images and one face test camera. In both cases, the held-out views are close to the training views: NeRF360 validation frames are interleaved with training frames along the same capture trajectory, and the face test camera sits inside the training hemisphere. Since the texture grids are optimized from random initialization with only a photometric loss, there is no mechanism that prevents them from absorbing view-dependent appearance, especially in regions where geometry is inaccurate or the camera baseline is small. The risk is amplified by the large increase in color parameters and the lack of any cross-splat texture consistency constraint. This is the weakest link in the argument: if the textures overfit, the measured gains are an artifact of the evaluation protocol rather than a genuine improvement in scene representation. The proposed leave-k-out test would directly measure generalization to viewpoints more distant from the training cameras and would settle whether the concern lands. The reader's CONDITIONAL verdict is appropriate because the evidence is plausible but the overfitting risk is not retired.","tokens_in":14744,"tokens_out":9459,"duration_ms":93528,"concrete_test":"Retrain the face model with a leave-k-out protocol: train on 10 of the 14 cameras and test on the 4 omitted cameras (including the central camera), for both 2DGS and Ours (N=4). If the LPIPS advantage of Ours narrows or reverses on the omitted cameras, the texture representation is overfitting to training views. Report per-view metrics and, if feasible, a texture-consistency check by projecting the same 3D surface samples into several views and comparing their uv-sampled colors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 3.3 trains the per-splat texture grids (Eq. 5) jointly with all other parameters using only a photometric loss and no texture-specific regularization. Each 4x4 grid adds 48 color parameters per splat (plus per-splat SH), roughly an order of magnitude more color parameters than 2DGS. Because the texture is anchored to the local uv frame of each primitive and is not required to be consistent across overlapping splats, the optimizer has substantial freedom to use these parameters to memorize the appearance of training views rather than the true surface albedo. This is not purely hypothetical: in the image-overfitting experiments (Sec. 4.1) the task is to reproduce a single image, so view consistency is never exercised. The only genuinely novel-view evidence is the face experiment (Sec. 4.2) with a single held-out camera, and the NeRF360 validation images, which are sampled from the same trajectory as the training frames and therefore close to them in viewpoint. If the textures encode view-dependent patterns, the reported LPIPS gains would not transfer to large-baseline novel views, and the central claim of improved scene reconstruction would fail. The paper provides no cross-view texture consistency check and no multi-seed or leave-k-out analysis to rule this out.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Gaussian Billboards, an extension of 2D Gaussian Splatting (2DGS) in which each 2D primitive is assigned a small N×N texture grid (default 4×4) whose colors are bilinearly interpolated in the primitive's local uv-plane, replacing the per-primitive solid color. The method is trained end-to-end with the same photometric loss and densification/pruning scheme as 2DGS, and is integrated into the gsplat CUDA kernels with manually derived adjoint code. The authors evaluate on single-image overfitting across three images with ablations of the texture resolution N and the spatial extent parameter σ, and on 3D scene reconstruction using the NeRF360 dataset and a multi-view face capture with one held-out test camera. They report consistent LPIPS improvements and visual sharpness gains, with N=4 and σ=0.5 selected as the recommended hyperparameters, and they discuss the relationship to the concurrent GStex method.","tokens_in":14934,"tokens_out":6582,"duration_ms":55976,"significance":"If the central claim holds, the paper is a simple yet effective contribution: it decouples geometric and texture capacity in 2DGS, improving reconstruction of textured flat regions while using equal or fewer primitives. The strengths are the clean formulation (Eq. 5), the detailed ablation of N and σ, the transparent integration details and adjoint derivations in the supplementary material, and the qualitative demonstration on challenging scenes. The main weakness is that the novel-view evidence for the central claim is limited to near-training views and a single held-out camera, leaving the possibility of per-splat texture overfitting unresolved. The absence of error bars and the use of the same metrics for hyperparameter selection further weaken the quantitative claim of consistent improvement. Overall, this is a worthwhile enhancement that needs a more rigorous evaluation of view consistency and statistical significance.","major_comments":[{"comment":"The only genuinely held-out novel-view evidence is a single central test camera for the face reconstruction (Fig. 7), while the NeRF360 validation images are typically sampled from the same camera trajectory as the training frames and therefore have limited baseline. Since the per-splat textures are optimized with only a photometric loss and no cross-splat or cross-view consistency constraint (Sec. 3.3), the reported LPIPS gains could in principle come from view-dependent patterns encoded in the texture grids rather than improved scene appearance; without a large-baseline held-out evaluation or a cross-view texture-consistency check, the central claim of improved novel view synthesis is not fully supported.","section":"Sec. 4.2, Figs. 7-8"},{"comment":"The recommended hyperparameters N=4 and σ=0.5 are selected by measuring exactly the same SSIM and LPIPS metrics on which the final comparisons are made, and the ablation uses only three single-image fits with no repeated runs. Given that several final numbers in Fig. 8 show PSNR values slightly below the baseline and SSIM differences below 0.001, the phrase 'consistently outperforms' is too strong without error bars or a separate validation set for hyperparameter selection.","section":"Sec. 4.1, Figs. 5-6"},{"comment":"The texture grid of each primitive is initialized with slight color variations and trained without any regularization that would encourage agreement between overlapping primitives or stability under viewpoint changes. When the underlying geometry is imperfect, the optimizer has substantial freedom to store view-dependent shading in the texture, which is a concrete mechanism for the overfitting concern; the paper should include an analysis or experiment demonstrating that the learned texture grids correspond to stable surface appearance across views.","section":"Sec. 3.3, Eq. (5)"}],"minor_comments":[{"comment":"The acronym PSNR is misspelled as 'PNSR' in several captions and text lines; please correct these typographical errors.","section":"Figs. 4 and 8"},{"comment":"The caption lists only σ values 0.25, 0.5, 1.0, and 2.0, while the x-axis extends to 4.0 and 0.125; please align the caption with the plotted range.","section":"Fig. 5"},{"comment":"Reference [21] has a stray '1' at the end of the page field; please fix the bibliographic entry.","section":"References"},{"comment":"The phrase 'the full color grid residing within the isoline of 0.8 opacity' is unclear; consider rewording to 'the texture square [−σ,σ]² approximately coincides with the 0.8-opacity isoline of the Gaussian'.","section":"Sec. 3.2"},{"comment":"The paper claims the method is orthogonal to other improvements but does not provide a runtime or memory overhead comparison in absolute terms; a short paragraph quantifying the added parameters per splat (e.g., 48 RGB values for N=4) would help readers gauge the cost.","section":"Sec. 3.3"},{"comment":"The diffuse-specular split is introduced without an ablation; consider adding a sentence justifying the use of a single SH set per primitive versus per-texel SH coefficients.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid incremental contribution, but the central claim of improved novel-view synthesis rests on weak evidence. The authors should either provide a large-baseline held-out evaluation or soften the claim. Also, a comparison against GStex, whose code is available as concurrent work, would substantially strengthen the paper. The hyperparameter selection issue is fixable with a proper validation scheme or error bars. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean and honest incremental improvement to 2DGS. Per-splat texture grids replace the solid color, and the paper shows, with consistent LPIPS gains, that this improves image fitting and 3D reconstruction, especially on textured flat surfaces. The idea is simple, the implementation is careful (explicit adjoint derivations, integration into gsplat kernels), and the authors disclose concurrent work by Rong et al. (GStex) up front. That disclosure alone tells you the paper is written in good faith.\n\nThe ablation of sigma and N is genuinely useful. The finding that sigma around 0.5 works best—concentrating the texture within the high-opacity core of the Gaussian—is a real design insight. The training-time numbers for N=8 are a practical caveat that many readers will appreciate.\n\nNow the soft spots, in proportion. The evidence is all single-run, no error bars or seeds. Several PSNR and SSIM numbers are slightly worse than baseline, and the paper relies on LPIPS as the main differentiator. The hyperparameters N=4 and sigma=0.5 are chosen by measuring the same metrics used in the final evaluation, on three images. That is not fatal, but it is circularity at the mild end.\n\nThe overfitting concern raised in the stress-test is less damning than it looks. The per-splat texture is shared across all training views, and a splat seen from multiple viewpoints is constrained by all of them through the same photometric loss. That is an implicit multi-view consistency regularizer. The face experiment with a held-out center camera and consistent metrics is weak evidence, but not the only evidence. Still, the paper would be stronger with a large-baseline held-out evaluation and a check on cross-view texture consistency.\n\nThe bigger gap is that the only quantitative baseline is the original 2DGS. Given that GStex exists, readers need at least a qualitative or quantitative comparison to know what the per-primitive texture idea adds beyond that concurrent formulation. That is an addressable revision, not a flaw in the core claim.\n\nBottom line: for anyone working on Gaussian splatting, this is a useful, adoptable modification. It deserves a serious referee. I would send it to review, with the expectation of a modest revision: add multiple seeds, compare with GStex, and report at least one large-baseline novel-view number.","headline":"A clean, honestly reported incremental improvement to 2DGS that replaces per-splat solid colors with small texture grids; the evidence is consistent if not over-strong, and the paper deserves a serious referee.","tokens_in":15511,"tokens_out":1995,"would_cite":true,"duration_ms":17895,"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":"Giving each 2D Gaussian splat a tiny color texture sharpens reconstruction without adding primitives.","keywords":["2D Gaussian Splatting","Gaussian Billboards","per-primitive texture","bilinear interpolation","novel view synthesis","scene reconstruction","texture mapping"],"falsifier":"Train Gaussian Billboards and baseline 2DGS on a scene with a large flat textured surface observed by many training cameras, then evaluate on several held-out cameras over multiple random initializations; if the novel-view SSIM or LPIPS advantage over 2DGS disappears, or if the learned color grids contain view-dependent structure visible as per-camera patterns, the central claim fails.","tokens_in":14471,"feed_emoji":"🎨","tokens_out":4329,"duration_ms":36977,"temperature":0.7,"pith_summary":"Gaussian Billboards is an extension of 2D Gaussian Splatting that gives each flat, oriented splat its own small color texture instead of a single solid color. The paper argues that because most scenes carry more texture information than geometric information, a per-splat texture lets the same or fewer primitives reconstruct sharp detail, especially on flat but richly textured surfaces such as posters, marble, foliage, and faces. The authors show in overfitting experiments that texture resolution N=4 with spatial extent σ=0.5 improves PSNR, SSIM, and LPIPS, and that on NeRF360 scenes and face captures the textured splats improve visual quality without changing training hyperparameters. A sympathetic reader would take the central claim to be: spatially varying per-splat color is a simple, orthogonal upgrade to 2DGS that improves reconstruction sharpness and quality.","feed_headline":"Per-splat textures sharpen 2D Gaussian rendering","feed_subtitle":"A small color grid per splat recovers fine detail on flat surfaces, matching or beating solid-color splats.","key_machinery":"The central object is the per-primitive texture grid, an N×N array of RGB colors parameterized by the same (u,v) tangent-plane coordinates that already define each 2D Gaussian splat. During rasterization the pixel color of a primitive is computed by bilinear interpolation with border clamping: c(u,v)=bilinear(C, u/σ, v/σ), so the texture rotates and scales with the splat's tangent vectors t_u, t_v and scales s_u, s_v. The machinery works because gradients flow through the interpolation to the grid and to the uv coordinates, hence to position, scale, and orientation, and the fused CUDA kernels of gsplat carry the extra per-primitive parameters in shared memory. View-dependent appearance is handled by a diffuse-specular split: the grid stores the Lambertian albedo and a single per-primitive set of spherical harmonics coefficients adds a view-dependent residual.","core_discovery":"On the paper's own terms, the discovery is that replacing the constant color c_k of each 2D Gaussian primitive with a spatially varying color obtained by bilinearly interpolating an N×N color grid C_k in the primitive's uv-frame improves reconstruction capability both when the number of primitives is fixed and when densification and pruning adapt the count during optimization. The color at a point is c(u,v)_k = bilinear(C_k, u/σ, v/σ), with texture coordinates clamped to [−σ, +σ]^2; σ=0.5 places most of the texture inside the high-opacity core of the Gaussian. The paper reports that this modification yields consistently better LPIPS on NeRF360 validation images, better SSIM in five of seven scenes, sharper flat textured regions, and, on a face capture, similar metrics with roughly 64,000 primitives instead of 90,000. The authors frame the result as merging classic billboard texturing with modern splatting optimization.","pith_inferences":["If small per-splat textures remain view-consistent under a purely photometric loss, the same idea could be applied to other per-primitive attributes such as opacity or normal maps, potentially sharpening geometry as well as color.","The shared-memory ceiling that currently limits N to 8 suggests a testable extension: storing textures in a compressed or hashed form, or sharing a small atlas across splats, could lift the resolution limit without the large training slowdown at N=8.","The texture grid decouples appearance from geometry more cleanly than solid colors, so texture-space editing or relighting of individual splats becomes a natural next step; the paper does not explore this."],"forward_implications":["With a fixed primitive budget, e.g. 10,000 or 100,000 splats, image-fitting PSNR, SSIM, and LPIPS improve monotonically from N=1 (solid color) up to N=8, with N=4 recommended for the best quality-to-cost trade-off.","Under standard 2DGS densification and pruning with unchanged hyperparameters, Gaussian Billboards improve novel-view quality on NeRF360 scenes, most visibly on flat textured regions like marble counters and vegetation.","On a multi-view face capture, the textured representation reaches comparable metrics with fewer primitives (about 64,000 versus 90,000) and a slightly faster test-view render.","The diffuse-specular split keeps the texture view-independent, so the method remains compatible with other 2DGS and 3DGS improvements as an orthogonal change."],"supporting_citations":[{"why":"The 2D Gaussian Splatting baseline whose per-primitive solid color is replaced; the central comparison for all experiments.","marker":"[6]"},{"why":"3D Gaussian Splatting provides the foundation of the scene representation, training, and densification heuristics that the method builds on.","marker":"[9]"},{"why":"gsplat is the CUDA rasterization library whose fused forward and backward kernels are extended to carry per-primitive color grids.","marker":"[21]"},{"why":"The Mip-NeRF 360 dataset supplies the 3D scene reconstruction benchmark used for the main quantitative evaluation.","marker":"[1]"},{"why":"Concurrent work on per-primitive texturing of 2DGS; the paper positions its contribution against it and notes differences in texture scheduling and resolution.","marker":"[16]"},{"why":"EWA volume splatting supplies the analytic projection approximation used to rasterize Gaussian primitives to screen space.","marker":"[27]"}],"fun_headline_variants":["Textured splats refine flat-surface detail","Color grids per splat improve Gaussian rendering","2DGS gets a texture map with each primitive","Spatially varying color enhances 2D Gaussian splats"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a small per-splat texture, optimized from scratch with only a photometric loss and no texture-specific regularization, will remain view-consistent rather than memorizing per-view color patterns in the local uv frame; if it memorizes, the sharpness gains would not transfer to novel views.","fun_headline_variants_meta":{"raw":{"variants":["Textured splats refine flat-surface detail","Color grids per splat improve Gaussian rendering","2DGS gets a texture map with each primitive","Spatially varying color enhances 2D Gaussian splats"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1818,"prompt_tokens":960,"completion_tokens":858,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":806}},"tokens_in":576,"tokens_out":858,"duration_ms":8152,"temperature":1.0,"reasoning_tokens":806,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:47:11.763624+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Gaussian Billboards and baseline 2DGS on a scene with a large flat textured surface observed by many training cameras, then evaluate on several held-out cameras over multiple random initializations; if the novel-view SSIM or LPIPS advantage over 2DGS disappears, or if the learned color grids contain view-dependent structure visible as per-camera patterns, the central claim fails.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically accu- rate radiance fields","cited_arxiv_id":null,"evidence_quote":"The 2D Gaussian Splatting baseline whose per-primitive solid color is replaced; the central comparison for all experiments."},{"cited_title":"3D gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"3D Gaussian Splatting provides the foundation of the scene representation, training, and densification heuristics that the method builds on."},{"cited_title":"Mip-NeRF 360: Unbounded anti-aliased neural radiance fields","cited_arxiv_id":null,"evidence_quote":"The Mip-NeRF 360 dataset supplies the 3D scene reconstruction benchmark used for the main quantitative evaluation."},{"cited_title":"GStex: Per-primitive tex- turing of 2D gaussian splatting for decoupled appearance and geometry modeling","cited_arxiv_id":null,"evidence_quote":"Concurrent work on per-primitive texturing of 2DGS; the paper positions its contribution against it and notes differences in texture scheduling and resolution."},{"cited_title":"EW A vol- ume splatting","cited_arxiv_id":null,"evidence_quote":"EWA volume splatting supplies the analytic projection approximation used to rasterize Gaussian primitives to screen space."}],"review_version":1}