{"id":"5fea50e8-373a-4691-9a05-06369b05655e","arxiv_id":"2412.00905","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Ref-GS enhances 2D Gaussian splatting with deferred shading and a spherical mip-map feature grid, yielding sharper reflections and more accurate surface normals.","lead":"Ref-GS is a new way to render 3D scenes that makes reflections and shiny surfaces look realistic in real time. It separates lighting into surface shape, viewing direction, and roughness, then combines them with a small neural network.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Alpha-blended G-buffer normals are used to compute reflection directions, but no argument or experiment shows this preserves surface orientation at edges and transparent regions; the central deferred-shading mechanism is unverified.","rationale":"Ref-GS's novelty is deferred shading after alpha-blending per-Gaussian attributes; the entire directional query mechanism depends on the G-buffer normal being a meaningful surface orientation. In standard deferred shading, the G-buffer stores the frontmost surface; in Gaussian splatting, alpha-blending computes a weighted mixture, which is not equivalent unless exactly one Gaussian fully covers the pixel. The paper neither states that normals are renormalized after blending nor bounds the error at silhouettes. This is not a failure to match a baseline; it is a potential internal inconsistency in the central construction. The experiments do not isolate this effect: the strongest geometry metric is average MAE on six scenes, with per-scene results that are often worse than 3DGS-DR, and no confidence intervals. The reader's weakest assumption matches this concern, so I agree. The requested ablation and error-localization test would settle it. If the depth-gradient normal works as well or better, the claim should be softened; if not, the current CONDITIONAL verdict remains appropriate pending code and significance testing.","tokens_in":18179,"tokens_out":6210,"duration_ms":56353,"concrete_test":"Add an ablation that replaces α-blended normals with the depth-gradient normal \\hat{n} for computing ω_r, keeping everything else fixed, and compare PSNR/LPIPS on held-out views and MAE on Shiny Blender. Additionally, compute, on test views, the pixel-wise angular deviation between the blended G-buffer normal and \\hat{n}; if high-deviation pixels coincide with the largest rendering errors or with transparent/edge regions (e.g., Glass & Ball), the premise fails. If the depth-gradient variant is no worse, the alpha-blended normal is not the cause of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is in §5.1: per-pixel G-buffer normals are obtained by alpha-blending per-Gaussian normals (Eq. 4) and then used to compute the reflection direction ω_r for Sph-Mip (Eq. 6). For an opaque single surface this is fine, but at occlusion boundaries and for transparent/refractive objects the blended normal is a weighted average of normals from different surfaces, not the orientation of any real surface; with partial opacity the vector may even fail to be unit length. The paper provides no normalization step and no argument that α-blending preserves normal consistency. The normal-consistency loss L_n in Eq. (10) pushes per-splat normals toward the depth-gradient normal, but the depth map itself is alpha-composited, so this does not fix G-buffer normals at depth discontinuities. The claimed geometry accuracy (MAE 2.21° on Shiny Blender) is also fragile: per-scene, Ref-GS is worse than 3DGS-DR on Ball, Helmet, Teapot and Coffee, and no variance/error bars are reported. Thus the central claim that deferred shading yields both superior view-dependent effects and accurate geometry rests on an unverified premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Ref-GS, a 2D Gaussian splatting method for jointly recovering view-dependent appearance and geometry. It introduces deferred shading to Gaussian splatting: primitive attributes (diffuse color, feature, roughness, normal) are alpha-blended into a G-buffer and then shaded in a second pass using a spherical Mip-grid (Sph-Mip) that encodes far-field lighting conditioned on reflection direction and roughness. A directional factorization via the per-pixel outer product of a spatial feature and a directional feature is used to model spatially varying inter-reflections. The method is evaluated on synthetic and real datasets (NeRF Synthetic, Shiny Blender, Glossy Synthetic, Shiny Real, Mip-NeRF360, Tanks and Temples, Glass & Ball) against NeRF- and Gaussian-based baselines, with PSNR/SSIM/LPIPS and normal MAE metrics, plus ablations and a speed comparison.","tokens_in":18445,"tokens_out":6422,"duration_ms":52107,"significance":"If the results hold, Ref-GS would be a valuable contribution to reflective and refractive scene reconstruction with 2D Gaussian splatting, achieving real-time rendering with improved view-dependent effects and competitive geometry accuracy. The paper's ablations are internally consistent: removing Sph-Mip, the mipmap, deferred shading, or the directional factorization each degrades metrics. The supplementary material provides per-scene breakdowns, additional datasets, and material-editing visualizations, which strengthen the empirical contribution. However, the central deferred-shading mechanism relies on alpha-blended G-buffer normals for reflection queries, and the paper provides no formal or experimental validation of that premise in challenging regions. The lack of error bars also makes the claimed superiority over close baselines difficult to assess. These issues temper but do not eliminate the significance.","major_comments":[{"comment":"The G-buffer normal N is produced by alpha-blending per-Gaussian normals via Eq. (4), and this blended normal is then used to compute the reflection direction ω_r in Eq. (6) for the Sph-Mip query. The paper gives no argument that alpha-blending preserves surface orientation at depth discontinuities or for transparent/refractive objects, and it does not normalize the blended normal; with partial opacity the vector need not be unit length. The normal-consistency loss in supplementary Eq. (10) supervises per-splat normals against the depth-gradient normal, but the depth map itself is alpha-composited, so it cannot correct the G-buffer normal at occlusion boundaries. Because the specular term depends on ω_r, this is a load-bearing unverified premise. Please provide a theoretical justification, a normalization step, or an experiment quantifying the error of the blended normal at edges and on transparent scenes (e.g., using the available ground-truth normals of Shiny Blender).","section":"Section 5.1 (Eqs. 4-6)"},{"comment":"The claim of accurate geometry recovery is based on an average MAE of 2.21° on Shiny Blender, but the per-scene results in supplementary Table 5 show that Ref-GS is worse than 3DGS-DR on Ball (1.05 vs 0.85), Helmet (1.99 vs 1.67), Teapot (0.69 vs 0.53), and Coffee (3.61 vs 2.21). No variance or multiple-seed results are reported anywhere in the paper, and the rendering margins over 3DGS-DR are small (e.g., Shiny Blender PSNR 34.80 vs 33.94 in Table 1). Without error bars or a significance test, the statement that Ref-GS achieves 'superior photorealistic rendering ... while also accurately recovering geometry' (abstract) is not robustly supported. Please report standard deviations over at least three runs, or at least a paired per-scene analysis.","section":"Section 6.4 and Table 5 (supplementary)"}],"minor_comments":[{"comment":"The limitations paragraph in the appendix appropriately acknowledges that the method is slower than 2DGS and is hard to integrate into standard CG engines; this honesty is appreciated, but the abstract's phrase 'superior photorealistic rendering' should be nuanced by the fact that on NeRF Synthetic the PSNR (33.20) is below 3DGS (33.30) and 3iGS (33.60) in Table 2.","section":"Supplementary B"},{"comment":"The sentence 'Our Ref-GS achieves effective reconstruction of high-frequency reflection and fraction' appears to contain a typo; 'fraction' should likely be 'refraction'.","section":"Section 1"},{"comment":"There is a typo in the sentence 'The goal of the shallow MLP fΘ is is to non-linearly map...' (duplicate 'is').","section":"Supplementary A.1"},{"comment":"The implementation details give the Sph-Mip grid size and MLP width but do not specify the learning rate or the number of training views per scene; specifying these would improve reproducibility.","section":"Section 6.3"},{"comment":"The phrase 'we treat M, N, K as a G-buffer and pass it a standard rasterization render for shading' is unclear; the G-buffer should be passed to the shader rather than to a render.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent combination of existing ideas (deferred shading from 3DGS-DR, tensorial factorization from TensoRF, and Mip-grid encoding). Its main novelty is the Sph-Mip and the outer-product factorization, which are empirically useful. The central weakness is the unverified normal-blending assumption; I believe this can be addressed with additional experiments, so I recommend major revision rather than rejection. I also note that the paper does not release code, which is a growing expectation in this area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Ref-GS is a solid, incremental but useful contribution to the 2D Gaussian splatting literature. The new combination is deferred shading with a roughness-aware spherical mip-grid and an outer-product factorization of spatial and directional features. That combination isn't in the prior work, and the ablations are internally consistent: removing the mip-grid, the mipmap levels, deferred shading, or the factorization each hurts. The real-world results on reflections are visually convincing, and the geometry recovery is plausible.\n\nThe soft spots are real but not fatal. The paper reports no error bars or multiple seeds, and the PSNR margins over 3DGS-DR are often small; on Shiny Blender the average is 34.80 vs 33.94, but per-scene you lose on Ball, Helmet, Teapot, and Coffee. That means the headline claim of \"superior\" rendering needs a significance test before it's credible. The alpha-blended normal issue the stress-test flags is worth probing: at occlusion edges and for transparent objects, the blended normal isn't the orientation of any surface, and the reflection direction can be off. But this is a common approximation in deferred shading, and the network is trained end-to-end, so it can compensate; the paper should discuss it and ideally analyze it, but I don't see it as a load-bearing flaw.\n\nThe bigger practical problem is that no code is released. That, combined with small margins, makes independent verification hard. The paper would be much stronger with code and a few seed runs. The citation pattern looks fair: 2DGS and TensoRF are properly credited, and reliance on the group's own prior work is legitimate since those are published.\n\nWho is this for? People working on view-dependent effects in Gaussian splatting, reflective scene reconstruction, or real-time rendering. It deserves serious peer review; a good referee would ask for error bars, code, and a closer look at the normal blending at discontinuities. I would not desk-reject it. Send it out.","headline":"Solid incremental contribution to 2D Gaussian splatting with deferred shading; the geometry claim is plausible but needs code and error bars before I'd trust the small margins.","tokens_in":19011,"tokens_out":2486,"would_cite":true,"duration_ms":23214,"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":"Ref-GS defers shading until after Gaussian attributes are blended, then queries a roughness-aware spherical Mip-grid with the reflected direction and uses an outer product of spatial and directional features to render view-dependent…","keywords":["2D Gaussian splatting","deferred shading","view-dependent effects","specular reflection","refractive surfaces","spherical mip grid","directional factorization","geometry recovery"],"falsifier":"Take a synthetic scene with a thin translucent object or two overlapping transparent surfaces where front and back surfaces both land in one pixel, render it with a ground-truth path tracer, and compare Ref-GS's predicted normal and specular highlight against the true reflection. If the blended normal lies between the two surfaces, the highlight will appear at the wrong location or blur; that would show the deferred normal is not a single consistent surface orientation.","tokens_in":17972,"feed_emoji":"✨","tokens_out":6747,"duration_ms":60768,"temperature":0.7,"pith_summary":"Ref-GS is a method for 2D Gaussian splatting that targets view-dependent appearance: reflections, highlights, and refraction. Its central move is to stop querying color per Gaussian primitive and instead alpha-blend per-splat attributes (diffuse color, appearance features, roughness, normal) into a screen-space G-buffer, then shade after blending. The claim is that this deferred shading removes the ambiguity between a splat's orientation and the viewing direction that makes prior per-primitive reflection queries fail, and that it preserves accurate geometry instead of trading rendering quality for surface noise. The paper grounds the claim in synthetic and real-world experiments across reflective, glossy, and refractive scenes.","feed_headline":"Deferred shading resolves reflection ambiguity in Gaussian splatting","feed_subtitle":"Blend splat attributes first, then shade: rough and shiny surfaces get sharp reflections and clean geometry.","key_machinery":"The load-bearing mechanism has three parts. (1) Deferred Gaussian rendering: per-Gaussian attributes are $\\alpha$-blended first, producing a G-buffer with diffuse color, feature $K$, roughness $M$, and normal $N$; shading happens afterward on this screen-space surface. (2) Sph-Mip grid: a learnable multi-level spherical feature grid whose two angular axes index the reflection direction $\\omega_r$ and whose third axis indexes roughness, so the same grid serves both sharp and rough surfaces with trilinear interpolation. (3) Directional factorization: the per-pixel outer product $K \\otimes S$ of the spatial feature and the directional feature feeds a shallow MLP, letting a single shared lighting representation combine with local material and geometry information to reproduce spatially varying inter-reflections.","core_discovery":"On its own terms, the paper's discovery is that directional encoding belongs on the rendered surface, not on each Gaussian. Ref-GS first renders a G-buffer by blending the per-Gaussian diffuse color, feature vector, roughness, and normal along each ray; it then computes a reflection direction from the blended normal and viewing direction, encodes that direction with a learnable multi-level spherical feature grid (Sph-Mip grid) that is also indexed by roughness, and predicts specular color from the directional feature plus the per-pixel outer product of the spatial and directional features through a lightweight MLP. The paper argues that this pipeline resolves the ambiguity in which per-primitive SH coefficients and orientations could otherwise undo the reflection transform, and that the factorization keeps per-primitive features small (four channels) and the renderer fast while supporting spatially varying near-field effects.","pith_inferences":["If the alpha-blended normal assumption fails at edges of transparent or overlapping surfaces, a natural extension is to keep a small set of candidate normals per pixel and select the one with the highest transmittance-weighted confidence before computing the reflection direction.","Because the Sph-Mip grid encodes lighting separately from the spatial features, swapping the grid for a different environment should give relighting of the same scene without retraining the geometry; the paper does not demonstrate this but the factorization invites it.","The outer-product factorization is a rank-one tensor decomposition; adding more rank terms could trade a little speed for better handling of multi-bounce reflections, a testable extension the paper leaves implicit."],"forward_implications":["Reflective and refractive surfaces, including windshields, spheres, and glass objects, can be reconstructed with cleaner normals and meshes than prior Gaussian baselines, not just rendered more convincingly.","Specular highlights stay sharp in novel views because the reflection direction is computed from a blended surface rather than from individual splats.","Rough and smooth materials are handled by one representation: interpolation along the roughness axis of the Sph-Mip grid adjusts the directional feature without adding extra networks.","The outer-product factorization keeps per-Gaussian features to four dimensions, making the rendering pipeline fast enough for real-time novel view synthesis at more than 45 FPS at 800x800 resolution.","The separated diffuse, specular, and roughness maps support material editing, such as changing roughness or diffuse color, without retraining."],"supporting_citations":[{"why":"Supplies the 2D Gaussian splatting base representation, ray-splat intersection, and training losses that Ref-GS extends.","marker":"[11]"},{"why":"Provides the original Gaussian splatting formulation and forward-rendering baseline that Ref-GS compares against.","marker":"[14]"},{"why":"Contributes the reflection-direction encoding idea and the Shiny Blender and Shiny Real datasets used for evaluation.","marker":"[35]"},{"why":"Is the deferred-reflection Gaussian baseline whose spherical-domain training strategy Ref-GS adopts.","marker":"[46]"},{"why":"Is the Gaussian shading-function baseline that Ref-GS must beat in rendering and geometry comparisons.","marker":"[12]"},{"why":"Provides the factorized tensorial illumination baseline and comparison for reflective scenes.","marker":"[34]"},{"why":"Supplies the low-rank tensor factorization idea that motivates the outer-product geometry-light factorization.","marker":"[7]"},{"why":"Provides the Glass & Ball refractive-object dataset used to test generalization to transparent materials.","marker":"[5]"},{"why":"Provides the Glossy Synthetic dataset and reflective-object reconstruction method used for geometry and rendering comparisons.","marker":"[22]"},{"why":"Provides real-world scenes such as Garden that contain reflections and are used for evaluation.","marker":"[4]"}],"fun_headline_variants":["Ref-GS: Deferred shading for unambiguous reflections in Gaussian splatting","Deferred shading on splat surfaces sharpens reflections and geometry","Roughness-aware directional factorization in Gaussian splatting","Ref-GS: Blend attributes first, then shade for accurate reflections","Surface encoding resolves Gaussian splatting's reflection ambiguity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method assumes that when several Gaussian splats overlap in a pixel, blending their normals gives the true surface direction at that pixel; where that fails, the reflected ray is pointed the wrong way and the specular result is ungrounded.","fun_headline_variants_meta":{"raw":{"variants":["Ref-GS: Deferred shading for unambiguous reflections in Gaussian splatting","Deferred shading on splat surfaces sharpens reflections and geometry","Roughness-aware directional factorization in Gaussian splatting","Ref-GS: Blend attributes first, then shade for accurate reflections","Surface encoding resolves Gaussian splatting's reflection ambiguity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001623,"raw_usage":{"total_tokens":6403,"prompt_tokens":838,"completion_tokens":5565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":5478}},"tokens_in":454,"tokens_out":5565,"duration_ms":39997,"temperature":1.0,"reasoning_tokens":5478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:52:24.416146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic scene with a thin translucent object or two overlapping transparent surfaces where front and back surfaces both land in one pixel, render it with a ground-truth path tracer, and compare Ref-GS's predicted normal and specular highlight against the true reflection. If the blended normal lies between the two surfaces, the highlight will appear at the wrong location or blur; that would show the deferred normal is not a single consistent surface orientation.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically accu- rate radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D Gaussian splatting base representation, ray-splat intersection, and training losses that Ref-GS extends."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Provides the original Gaussian splatting formulation and forward-rendering baseline that Ref-GS compares against."},{"cited_title":"Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields","cited_arxiv_id":null,"evidence_quote":"Contributes the reflection-direction encoding idea and the Shiny Blender and Shiny Real datasets used for evaluation."},{"cited_title":"3d gaussian splat- ting with deferred reflection","cited_arxiv_id":null,"evidence_quote":"Is the deferred-reflection Gaussian baseline whose spherical-domain training strategy Ref-GS adopts."},{"cited_title":"Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces","cited_arxiv_id":null,"evidence_quote":"Is the Gaussian shading-function baseline that Ref-GS must beat in rendering and geometry comparisons."},{"cited_title":"3igs: Factorised tensorial illumination for 3d gaussian splatting","cited_arxiv_id":null,"evidence_quote":"Provides the factorized tensorial illumination baseline and comparison for reflective scenes."},{"cited_title":"Tensorf: Tensorial radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies the low-rank tensor factorization idea that motivates the outer-product geometry-light factorization."},{"cited_title":"Eikonal fields for refractive novel-view synthesis","cited_arxiv_id":null,"evidence_quote":"Provides the Glass & Ball refractive-object dataset used to test generalization to transparent materials."},{"cited_title":"Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images","cited_arxiv_id":null,"evidence_quote":"Provides the Glossy Synthetic dataset and reflective-object reconstruction method used for geometry and rendering comparisons."},{"cited_title":"Mip-nerf 360: Unbounded anti-aliased neural radiance fields","cited_arxiv_id":null,"evidence_quote":"Provides real-world scenes such as Garden that contain reflections and are used for evaluation."}],"review_version":1}