{"id":"d25163a8-ad2a-4e41-a0e1-22e5fd13cabf","arxiv_id":"2412.04459","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"Adaptive sparse voxels rasterized with direction-dependent Morton ordering match 3D Gaussian Splatting quality at real-time frame rates without neural networks.","lead":"SVRaster renders photorealistic 3D scenes in real time by rasterizing adaptive sparse voxels instead of neural networks or Gaussian splats. It matches 3D Gaussian Splatting quality on standard benchmarks while avoiding the popping artifact, and ships public code.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Popping-free claim rests on an unproven entry-face assumption: the Sec. B.3 induction assumes rays enter each voxel through the -x face, so direction-only Morton order is not proven for perspective cameras entering through -y or -z faces.","rationale":"The paper's central value proposition has two pillars: (1) real-time rendering with quality comparable to 3DGS, and (2) freedom from popping artifacts due to correct depth ordering. Pillar (2) is explicitly stated in the abstract and Sec. 1, and the rasterizer's sorting design (Sec. 3.1.2) is the mechanism. The evidence for pillar (1) is strong: public code, thorough ablations, and consistent benchmark results. The evidence for pillar (2) is the proof in Sec. B.3 plus a supplementary video. My stress-test focused on the proof because the 'ray origin doesn't matter' claim is the only load-bearing step that, if false, invalidates the guarantee. The empirical quality metrics would barely change if the ordering were occasionally wrong, so the quantitative results do not test this claim. The reader identified precisely this weakness; I agree. The concern is not that the method fails in practice, but that the paper asserts a mathematical guarantee it has not proven. A corrected proof could (a) show that for all camera origins used in the benchmarks, every ray enters every voxel through the same face, (b) scope the popping-free claim to such configurations, or (c) replace the sorting key with a true depth key. Without one of these, the headline claim overreaches. I therefore recommend keeping the reader's conditional verdict.","tokens_in":27897,"tokens_out":12359,"duration_ms":202713,"concrete_test":"Construct a one-level octree over [0,1]^3 with the paper's order table for signbits=0 (all ray components positive). Take two rays: r1 with origin (-0.1, 0.5, 0.5) and direction (1, 0.1, 0.1), which enters the root through the -x face, and r2 with origin (0.5, -0.1, 0.5) and direction (0.1, 1, 0.1), which enters through the -y face. For each ray, compute the true near-to-far order of the eight child voxels by ray-AABB entry distances (Listing 2). If the two true orders differ while both share signbits=0, then the direction-dependent Morton order in Sec. 3.1.2 cannot be correct for both rays, disproving the base case of Sec. B.3. Optionally, run the released code on a synthetic scene with this camera placement and check for popping or ordering artifacts at voxel boundaries.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 3.1.2 asserts that the correct Morton-order permutation is 'solely dependent on the positive/negative signs of the ray direction (the ray origin doesn't matter).' The induction proof in Sec. B.3 (base case, Fig. 10) only establishes correctness for (+,+,+) rays that enter the octree root through the -x face: it orders the four x-low children before the four x-high children because such a ray hits the -x side first. For a perspective camera placed around or inside the scene, a (+,+,+) ray can enter the root through the -y or -z face, in which case the true near-to-far order of the eight octants is y-major or z-major, not x-major. The same issue recurs inside every descendant voxel: a ray can enter a voxel through any of the three negative faces, depending on which boundary it crossed last, so no single permutation of the 48-bit Morton key can be correct for all rays sharing the same sign bits. Thus the claimed guarantee of 'no popping artifacts' is not established by the provided proof. The empirical quality metrics and the supplementary video may still be fine, but the proof gap directly touches a headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SVRaster, a radiance field representation based on adaptive sparse voxels with trilinear density fields, SH-based view-dependent colors, and a custom CUDA rasterizer that sorts voxels per image tile using a direction-dependent Morton ordering. The authors claim that this ordering guarantees correct depth-sorted rendering and therefore eliminates popping artifacts, while achieving real-time frame rates and novel-view synthesis quality comparable to 3D Gaussian Splatting, all without neural networks or Gaussian primitives. The method is evaluated on Mip-NeRF360, Tanks&Temples, Deep Blending, DTU, and ScanNet++, with extensive ablations, a fast-training and a fast-rendering variant, and demonstrations of TSDF fusion and marching cubes on the same voxel grid. Code is publicly released.","tokens_in":28213,"tokens_out":8843,"duration_ms":89211,"significance":"If the correctness claims hold, this is a significant contribution: it shows that a fully explicit voxel representation, with no neural components and no sparse-point prior, can reach a quality-speed trade-off competitive with 3DGS while inheriting the well-defined volume and ordering properties of grids. The experimental work is thorough: per-scene breakdowns, ablations over most hyperparameters, memory and model-size comparisons, a ScanNet++ third-party benchmark evaluation, and released code that supports reproducibility. The demonstrated compatibility with classic grid algorithms (volume fusion, voxel pooling, marching cubes) is a genuine strength that opens practical extensions. The main reservation concerns the proof of the artifact-free ordering claim, which is central to the paper's headline contribution.","major_comments":[{"comment":"The ordering claim that the correct Morton-order permutation is solely a function of the ray direction signs and not of the ray origin is not established by the given proof. The base case in B.3 (Fig. 10) orders the four x-low octants before the four x-high octants for a (+,+,+) ray, which is only valid if the ray enters the octree root through the -x face. For a perspective camera positioned around or inside the scene, a (+,+,+) ray can enter through the -y or -z face, making the true near-to-far order y-major or z-major. Even for rays entering through the -x face, the order among the x-low children is not fixed: it depends on the entry y,z coordinates and the relative slopes, since the ray can cross the y=center or z=center planes before or after x=center. Therefore no single permutation of the 48-bit Morton key can be correct for all rays that share the same sign bits, and the induction step in B.3 does not close this gap. This directly affects the headline promise of being 'free from popping artifacts' in the abstract and Section 1. Please either prove the claim under explicit restrictive assumptions (e.g., all rays enter each voxel through its -x,-y,-z faces in a prescribed order), or revise the claim to be a heuristic that is empirically validated, or extend the sort key to be entry-face-aware and duplicate voxels accordingly.","section":"Section 3.1.2 / Appendix B.3"},{"comment":"The supersampling scale is reported inconsistently: the main text (Sec. 4.1) states hss=1.5, while the supplementary ablation (Table 7) marks hss=1.10 as the adopted setup and its caption says 'We use hss = 1.1 for speed-quality trade-off.' Since hss directly affects the reported rendering FPS in Table 1, the inconsistency makes the main FPS numbers difficult to reproduce. Please clarify which value was actually used for the main results and align the presentation.","section":"Section 4.1 / Supplementary Table 7"}],"minor_comments":[{"comment":"The columns in Table 4 are misaligned: the 1024^3 resolution appears to have LPIPS='OOM' and the PSNR and FPS entries are shifted into the adjacent columns, making the ablation results unreadable. Please reformat the table so each resolution is aligned with its own metrics.","section":"Table 4 / Section 4.3"},{"comment":"The sentence 'We handle the corner case when multiple Morton orders are required in supplementary materials' is only partially accurate: B.1 describes how voxels are duplicated for different ray sign bits, but the entry-face dependence caused by varying ray origins within a tile is not addressed.","section":"Section 3.1.2 / Supplementary B.1"},{"comment":"The text contains a typo: 'Plenxoels' should be 'Plenoxels'.","section":"Section 4.3"},{"comment":"The definition of vrate as the ratio vs / v_interval is dimensionally a number of pixels, but the text describes it as a 'sampling rate'; please clarify the interpretation in the surrounding paragraph.","section":"Section 3.2 / Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is empirically strong, well evaluated, and the code release adds credibility. The blocker is the proof of the central ordering guarantee: the provided induction is not valid for general perspective cameras, and the claim as stated is falsifiable with a simple counterexample. This is fixable within the manuscript's scope by either adding entry-face information to the sort key or by carefully restating the claim and adding an experiment that verifies or qualifies the behavior. I would advise the editor to ask the authors to address this before acceptance, and to consider requesting a short verification experiment with camera trajectories that pass through or around the scene to test whether popping actually occurs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a solid, well-executed paper that deserves a serious referee. The headline result—explicit adaptive sparse voxels rasterized in real time at 3DGS-like quality—is credible and well supported by public code, thorough ablations, and per-scene numbers. The one significant soft spot, as you flagged, is the popping-free proof in Sec. B.3: it doesn't support the claim that the correct Morton order depends only on ray-direction signs.\n\nWhat's actually new: leaf-only adaptive octree voxels (no parent pointers), a direction-dependent Morton permutation for sorting, and the rasterizer that makes this possible. They report >4 dB PSNR over Plenoxels and roughly 3DGS-quality rendering with better LPIPS, without using SfM points for initialization. The mesh extraction and 2D-feature fusion demonstrations are genuinely useful extensions, and the code is out. The ablations are honest—they show many hyperparameter sweeps and disclose failure cases (exposure variation, texture over-explanation).\n\nThe proof issue. Sec. 3.1.2 and the supplementary claim the sort order depends only on the ± signs of the ray direction. The induction in Sec. B.3 is written for (+,+,+) rays and effectively assumes the ray enters the root through the -x face: it puts the four x-low children before x-high because the ray 'is going toward +x.' For a perspective camera located around or inside a scene, a (+,+,+) ray can enter the root through the -y or -z face, in which case the correct near-to-far order is y- or z-major, not x-major. The same logic applies at every level of the octree, since the entry face of a child voxel depends on which boundary was crossed last. So the theorem as stated—'the ray origin doesn't matter'—is false. The empirical results may still look fine, and the authors themselves note ordering artifacts are not a major factor in the numbers (Sec. 4.3), but the headline popping-free guarantee is not proven as stated. That is a conditional, not a fatal, flaw: the videos could still be correct, and a corrected proof or a scoped claim (e.g., for cameras outside a bounding region with known entry face) may well salvage it.\n\nBottom line: for anyone working on radiance-field rendering, this is one of the more interesting neural-free/gaussian-free representations in a while. The paper deserves peer review and will likely get in with a revision that either fixes the proof or carefully scopes the ordering claim. I'd bring it to reading group and would probably cite it if I were working on voxel-based NVS.","headline":"Strong, well-executed sparse-voxel rasterization paper whose empirical claims are credible, but the popping-free guarantee rests on a proof that silently assumes a particular entry face and is not general.","tokens_in":28778,"tokens_out":5390,"would_cite":true,"duration_ms":51399,"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":"By rasterizing adaptive sparse voxels with direction-dependent Morton ordering, a radiance field renders in real time with quality comparable to 3D Gaussian splatting and without Gaussian-style popping artifacts.","keywords":["sparse voxels","radiance fields","rasterization","Morton ordering","novel view synthesis","real-time rendering","volume rendering","popping artifacts"],"falsifier":"Render a mixed-level voxel scene from a camera above or inside the foreground octree and compare against a reference image computed by z-sorting the actual voxel intersections per ray; any pixel whose compositing order differs from the reference contradicts the origin-independence claim. A cheaper analytic check enumerates rays entering the root through the −y and −z faces and compares the first octant they hit against the order the eight sign-based permutations imply.","tokens_in":27646,"feed_emoji":"🧊","tokens_out":13404,"duration_ms":123410,"temperature":0.7,"pith_summary":"This paper tries to establish that a radiance field can be rendered in real time by rasterizing explicitly stored sparse voxels, with no neural network, no 3D Gaussian primitive, and no structure-from-motion point prior. The authors claim their system beats the previous fully explicit voxel model, Plenoxels, by more than 4 dB PSNR and more than ten times the frame rate, while matching 3D Gaussian splatting in quality on standard benchmarks. The load-bearing idea is that the correct front-to-back order for voxels of mixed sizes depends only on the signs of the ray direction, so eight fixed Morton-order permutations sort an entire image tile without per-ray sorting. If the claim holds, voxel grids become a practical real-time primitive that inherits the well-defined volume and ordering properties that Gaussian splatting lacks.","feed_headline":"Rasterized sparse voxels match Gaussian splatting, no popping","feed_subtitle":"Morton-order sorting puts voxels front-to-back: real-time radiance fields without neural nets or SfM priors.","key_machinery":"The direction-dependent Morton ordering: each voxel's octree Morton code, from bit-interleaving its grid index, is remapped through one of eight hard-coded bit permutations chosen by the three sign bits of the ray direction, so that a single 48-bit sort key orders a whole tile's voxels near-to-far. The argument is by induction on octree level: in the base case the most significant bit separates the near half-space along each axis in turn, and subdividing a voxel appends the same base-case order inside its children, so mixed-level voxel sets stay correctly ordered after pruning and subdivision. The surrounding system is adaptive sparse-voxel optimization: initialization from constant empty space, pruning by maximum blending weight, subdivision guided by alpha-weighted loss gradients, corner-shared trilinear densities activated by exponential-linear, and per-voxel spherical-harmonic colors shared across all covered pixels.","core_discovery":"The paper's central claim is that a radiance field stored in adaptive explicit sparse voxels can be rasterized at interactive frame rates with quality on par with 3D Gaussian splatting while removing the popping artifacts caused by approximate Gaussian depth sorting. Scenes are represented by octree-layout leaf voxels up to level 16, a finest grid resolution of $65536^3$; each voxel carries a trilinear density field whose corner values are shared with neighbors and a spherical-harmonic color held constant inside the voxel. Rendering projects voxels to image tiles and sorts each tile's voxels by a direction-dependent Morton code, eight fixed permutations of the octree Morton bits selected by the sign pattern of the ray direction, which the authors prove by induction gives near-to-far order for voxels of mixed levels. Because space is partitioned into disjoint voxels and the order is exact, the paper states the rendering is free from popping artifacts. Against the previous fully explicit voxel model it reports over 4 dB higher PSNR and more than ten times the frame rate, and on the Mip-NeRF360 benchmark it reaches 121 FPS with PSNR 27.33, SSIM 0.822, and LPIPS 0.185, using no neural network, no Gaussians, and no SfM points.","pith_inferences":["The induction proof in the supplement covers rays entering the root octree through one face; for cameras inside or beside the scene, correct front-to-back order can depend on the entry point, so the popping-free guarantee is not yet established for all viewpoints the method is used from.","The paper itself concedes that the exact ordering barely changes average numerical scores, so the practical payoff is in fly-through video consistency, a dimension the benchmark tables do not measure.","The direction-dependent Morton scheme transfers to any octree-partitioned set of primitives, suggesting it is a general cure for center-sorting artifacts rather than a voxel-specific trick.","A concrete next test is initializing voxels from sensor depth through the paper's sparse-voxel TSDF-Fusion (which the authors point to as future work), replacing empty-space initialization and potentially closing the remaining geometry gap on surfaces."],"forward_implications":["Fully explicit voxels render 121 FPS on Mip-NeRF360 with LPIPS 0.185, better than 3DGS's 0.216 on the same benchmark, while using no COLMAP sparse-point prior.","The same trained voxels plug directly into Volume Fusion, Voxel Pooling, and Marching Cubes, so mesh extraction and lifting 2D features to 3D need no conversion step.","Mesh reconstruction from the density field alone reaches DTU chamfer distance 0.76 with 5-minute training, competitive with surface-specialized NeRF variants that use SDF parametrization.","The two speed variants span a practical trade-off frontier: about 258 FPS at a modest quality loss, and about 4.5-minute training at full render speed.","Ordering correctness holds for all mixed-level voxel configurations, so fly-through rendering avoids the popping artifacts that center-sorted Gaussian splatting produces."],"supporting_citations":[{"why":"the tile-based CUDA rasterizer this method rebuilds on, the main quality and speed baseline, and the source of the popping-artifact problem.","marker":"[20]"},{"why":"the previous fully explicit voxel grid model, the baseline for the claimed 4+ dB PSNR and 10x FPS improvements.","marker":"[11]"},{"why":"provides the main Mip-NeRF360 benchmark scenes and the distortion loss used in training.","marker":"[2]"},{"why":"documents the popping artifact caused by center-based sorting that this paper's Morton ordering is designed to avoid.","marker":"[39]"},{"why":"supplies the exponential-linear activation, the per-point RGB loss, and the direct voxel grid optimization lineage.","marker":"[45]"},{"why":"the Marching Cubes procedure adapted to sparse voxels for mesh extraction.","marker":"[27]"},{"why":"the volumetric fusion method adapted as sparse-voxel TSDF-Fusion for surface extraction and depth initialization.","marker":"[7]"}],"fun_headline_variants":["Sparse voxel rasterization beats Gaussian splatting popping","Real-time radiance fields via rasterized sparse voxels","Morton-sorted voxels end Gaussian popping artifacts","No neural nets, no Gaussians: voxel rasterization at 121 FPS","Adaptive sparse voxels: 65536^3 detail, real-time rendering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The popping-free guarantee rests on a proof that assumes rays enter the scene's root octree through one specific face; for cameras placed around or inside the scene, the front-to-back order can depend on where the ray enters, and that case is not covered by the proof.","fun_headline_variants_meta":{"raw":{"variants":["Sparse voxel rasterization beats Gaussian splatting popping","Real-time radiance fields via rasterized sparse voxels","Morton-sorted voxels end Gaussian popping artifacts","No neural nets, no Gaussians: voxel rasterization at 121 FPS","Adaptive sparse voxels: 65536^3 detail, real-time rendering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1321,"prompt_tokens":1011,"completion_tokens":310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":218}},"tokens_in":627,"tokens_out":310,"duration_ms":3134,"temperature":1.0,"reasoning_tokens":218,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:24:27.993724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a mixed-level voxel scene from a camera above or inside the foreground octree and compare against a reference image computed by z-sorting the actual voxel intersections per ray; any pixel whose compositing order differs from the reference contradicts the origin-independence claim. A cheaper analytic check enumerates rays entering the root through the −y and −z faces and compares the first octant they hit against the order the eight sign-based permutations imply.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"the tile-based CUDA rasterizer this method rebuilds on, the main quality and speed baseline, and the source of the popping-artifact problem."},{"cited_title":"Plenoxels: Radiance fields without neural networks","cited_arxiv_id":null,"evidence_quote":"the previous fully explicit voxel grid model, the baseline for the claimed 4+ dB PSNR and 10x FPS improvements."},{"cited_title":"Barron, Ben Mildenhall, Dor Verbin, Pratul P","cited_arxiv_id":null,"evidence_quote":"provides the main Mip-NeRF360 benchmark scenes and the distortion loss used in training."},{"cited_title":"Stopthepop: Sorted gaussian splatting for view-consistent real-time rendering","cited_arxiv_id":null,"evidence_quote":"documents the popping artifact caused by center-based sorting that this paper's Morton ordering is designed to avoid."},{"cited_title":"Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction","cited_arxiv_id":null,"evidence_quote":"supplies the exponential-linear activation, the per-point RGB loss, and the direct voxel grid optimization lineage."},{"cited_title":"Lorensen and Harvey E","cited_arxiv_id":null,"evidence_quote":"the Marching Cubes procedure adapted to sparse voxels for mesh extraction."},{"cited_title":"A volumetric method for building complex models from range images","cited_arxiv_id":null,"evidence_quote":"the volumetric fusion method adapted as sparse-voxel TSDF-Fusion for surface extraction and depth initialization."}],"review_version":1}