{"id":"b7aff358-73bf-415b-842c-dca7c70c692f","arxiv_id":"2504.12811","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 3D Gaussian rasterizer with an adaptive 3D smoothing filter, view-space bounding, and frustum culling removes aliasing, popping, and projection artifacts at real-time speeds.","lead":"This paper presents a new way to render 3D Gaussian scenes that removes flickering, popping, and stretched-image artifacts while still running fast enough for real-time use. It changes how Gaussians are filtered, bounded, and culled during rendering so results stay stable even for views far from the training camera positions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The tile-culling step (Sec. 3.4) is described as exact, but it minimizes on only the two screen-space-closest planes plus three edges; that subset search bounds the true minimum from above, so tiles with contributing Gaussians can be wrongly discarded, re-introducing the popping the paper claims…","rationale":"The reader identifies the maximum-contribution-along-ray approximation as the weakest assumption. That is a real modeling limitation, but it is an explicit design choice inherited from Hahlbohm et al. and applies to the entire hybrid 3D-evaluation line; it is not an internal inconsistency in the presented method. The more load-bearing concern is the gap between the claimed exact frustum culling and the described subset search in Sec. 3.4. The paper's central artifact-free claim specifically includes eliminating popping caused by incorrect culling (Sec. 3.3, Fig. 5). If the cheap culling heuristic can discard a Gaussian that still contributes to a tile, the method can re-introduce exactly the popping it promises to remove. The check is straightforward with the released code: compare the cheap decision against an exact convex minimization over the full tile frustum. This is a concrete, falsifiable test rather than a dispute about rendering models. If the test shows zero false discards, the concern is resolved and the artifact-free claim stands; if not, the paper must either use exact culling or add a conservative margin and re-measure timings, since exact culling would affect the real-time performance claim. For these reasons the verdict remains CONDITIONAL, matching the reader's verdict.","tokens_in":16043,"tokens_out":18834,"duration_ms":207038,"concrete_test":"Add an assertion mode to the released rasterizer that, for every Gaussian-tile pair on the interpolated test camera paths (and on the large-FOV and close-up sequences), compares the cheap 2-plane/3-edge culling decision against the exact minimum of ρ^2 over the full frustum (all 6 faces, 12 edges, and 8 vertices, or a convex QP solve). Count false discards: cases where the cheap decision discards the tile but the exact minimum is below τρ. If any such false discard overlaps a visible pixel, quantify the resulting per-pixel error and temporal popping; if none occur, the heuristic is safe and the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 claims exact frustum-based culling, but the described implementation computes the minimum of ρ(x)^2 over the tile frustum only on the two screen-space-closest x/y planes and three associated edges, rather than over all six faces and twelve edges. For a convex quadratic, restricting the search to a boundary subset gives an upper bound on the true minimum over the frustum F. The culling test in Eq. (18) discards a tile when this estimated minimum is above τρ; if the true minimum is below τρ, the Gaussian still contributes to that tile and is wrongly discarded. Because the screen-space-closest plane need not be the closest plane in the transformed Gaussian space (the projective map is nonlinear), this is not merely a theoretical edge case. A wrong cull at the image border or on a close-up view is exactly the popping artifact the paper claims to eliminate (Sec. 3.3, Fig. 5). The paper provides no conservative margin and no quantitative check that false culls never occur. This is separate from the inherited maximum-along-ray approximation, which is an explicit modeling choice; the culling heuristic is an internal gap between the 'exact' claim and the implementation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents AAA-Gaussians, a rasterization-based 3D Gaussian renderer that evaluates Gaussians in 3D throughout the pipeline. Three contributions are proposed: an adaptive 3D smoothing filter that dilates Gaussians based on the area perpendicular to the viewing ray, a view-space angle bounding method that stabilizes Gaussians extending behind the image plane, and frustum-based 3D tile culling for hierarchical sorting. The authors evaluate on Mip-NeRF 360, Tanks & Temples, and Deep Blending, reporting large-FOV, multi-resolution, and timing results, and argue that their method removes aliasing, popping, and projection distortions while remaining real-time.","tokens_in":16233,"tokens_out":2153,"duration_ms":22520,"significance":"If the claims hold, the paper would be a useful advance: it provides a unified rasterization approach to several known 3DGS artifacts, with clean derivations in Appendices A and B, an open-source implementation, and a broad evaluation including out-of-distribution views. The strongest evidence is the large-FOV and multi-resolution comparisons, where the proposed method clearly outperforms MCMC and other baselines, and the timing results show that the added culling largely compensates for the cost of hierarchical sorting. The main caveats are that in-distribution metrics are not state-of-the-art, with MCMC achieving higher PSNR on Mip-NeRF 360 and Tanks & Temples, and that the claimed exactness of the frustum culling is not supported by the described implementation.","major_comments":[{"comment":"The text states that 'we compute the point of maximum contribution of the Gaussian inside this 3D frustum' and later calls the culling 'exact frustum culling', but the implementation only projects onto the two screen-space-closest x/y planes and their three associated edges, not all six faces and twelve edges. For a convex quadratic, minimizing over a boundary subset yields an upper bound on the true minimum over the full frustum, so a tile can be discarded when the true minimum is below the threshold tau_rho, meaning a Gaussian that contributes to that tile is wrongly culled. Because the screen-space-closest plane is not necessarily the closest plane in the transformed Gaussian space under the nonlinear projective map, this is not a purely theoretical concern, and it directly affects the claimed removal of popping at image borders and close-up views. The authors should either compute all planes and edges, add a conservative margin so that culling is guaranteed not to remove contributing Gaussians, or present a quantitative check that false culls never occur in the evaluated scenes.","section":"Sec. 3.4, Eq. (18)"},{"comment":"The abstract claims 'state-of-the-art reconstruction quality' and 'significantly outperforms other approaches for out-of-distribution views', but Table 1 shows that MCMC has higher PSNR on Mip-NeRF 360 (28.027 vs. 27.835) and on Tanks & Temples (24.642 vs. 23.582), and the text itself says 'our method outperforms others in nearly all metrics and matches MCMC in overall quality'. The in-distribution claim should be rephrased to match the data, e.g., 'comparable to state-of-the-art on in-distribution views while achieving better out-of-distribution robustness.'","section":"Abstract and Sec. 4.1, Table 1"},{"comment":"The amplitude normalization in Eq. (10) is derived under the explicit modeling choice that a pixel's Gaussian contribution is the maximum value along the viewing ray rather than an integral over the ray. This assumption is inherited from Hahlbohm et al. [11] and is stated clearly, but it is load-bearing for the new anti-aliasing filter: if a future variant of the pipeline switched to full volumetric integration, the perpendicular-area normalization would no longer be calibrated. The paper would be strengthened by stating this dependency explicitly in the limitations and by discussing whether the maximum-along-ray approximation is validated for the artifact-free claim, especially at close distances where multiple high-magnitude contributions may overlap.","section":"Sec. 3.1 and Sec. 3.2, Eq. (9)-(13)"}],"minor_comments":[{"comment":"There is a typo: 'Following Kerblet al. [14]' should be 'Following Kerbl et al. [14]'.","section":"Sec. 3.1"},{"comment":"The Mip-Splatting row is duplicated for Mip-NeRF 360 and Tanks & Temples; one duplicate should be removed.","section":"Table 3"},{"comment":"The notation s1,3 and s2,3 is defined in the text, but the meaning of the subscripts is easy to confuse with the entries of the covariance matrix; using s_{i,j} = <t, T_i ⊙ T_j> consistently in the equations would improve clarity.","section":"Sec. 3.3, Eq. (14)-(15)"},{"comment":"The paper states that popping artifacts close to scene content are difficult to evaluate quantitatively and refers to a video; a static figure with annotated frames before and after the popping event would make the claim more verifiable in the printed manuscript.","section":"Sec. 4.2, 'Close to Scene Camera Location'"},{"comment":"The phrase 'reducing sorting costs' appears in the contribution list and should be 'reducing sorting costs' or 'reducing sorting overhead' for grammatical consistency.","section":"Sec. 3.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically solid in its derivations and evaluations, but the overstatement of state-of-the-art in-distribution quality and the gap between the 'exact' culling claim and the subset-search implementation are both fixable. I recommend major revision rather than rejection because the central ideas are sound and the issues are local: rephrase the quality claims and either make the culling truly conservative or provide evidence that the subset search does not cause false culls. The self-citation to StopThePop is appropriate as a building block and not used as evidence for the new contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible, useful 3DGS rendering paper. The three contributions—perpendicular amplitude scaling for anti-aliasing, view-space angle bounding, and 3D frustum tile culling—are genuinely new relative to Mip-Splatting, Hahlbohm et al., and StopThePop. The derivations in Appendices A and B are clean and parameter-free, and the out-of-distribution evaluations (large FOV, multi-resolution) show real, consistent gains. The ablations are careful, timings are reported honestly, and the code is public.\n\nBut the paper overstates one thing: Section 3.4 describes \"exact frustum culling,\" yet the implementation minimizes ρ(x)^2 only on the two screen-space-closest planes and three edges, not the full frustum boundary. As the stress-test note says, that subset search upper-bounds the true minimum, so a tile can be discarded when the real minimum falls below the threshold. That would reintroduce the popping artifact the paper claims to eliminate. The stress-test is on target here. It's not a fatal flaw—the core method is sound—but it's an internal gap between claim and implementation, and it needs either a conservative margin, a full search, or quantitative evidence that false culls never occur.\n\nThe inherited maximum-along-ray approximation is an explicit modeling choice, not a hidden flaw, but it does mean the new amplitude renormalization is calibrated to that approximation. That's fine for this line of work, but worth stating more plainly.\n\nOther soft spots are minor. The abstract says \"state-of-the-art quality,\" but on Tanks & Temples MCMC still has higher PSNR (24.64 vs 23.58); the paper explains this reasonably in Section 4.1, but the abstract is still a bit strong. The \"100 FPS\" claim in the conclusion isn't supported by the worst-case timing: M360 Outdoor runs at 10.66 ms, which is about 94 FPS. And the popping/culling improvements are only demonstrated qualitatively, not benchmarked.\n\nOverall, this is a genuine step forward for rasterization-based 3DGS rendering, and it deserves serious peer review. I'd send it out, and ask for a revision that either makes the culling conservative or benchmarks the approximation, plus a toned-down abstract and a corrected FPS statement. Then it's a solid accept.","headline":"A solid, well-derived 3DGS rendering paper whose 'exact' culling claim is the one real gap between the math and the text.","tokens_in":16905,"tokens_out":3864,"would_cite":true,"duration_ms":39747,"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":"3D Gaussian rendering goes artifact-free at 100+ FPS","keywords":["3D Gaussian Splatting","anti-aliasing","view-consistent rendering","3D Gaussian evaluation","frustum culling","hierarchical rasterization","out-of-distribution views","real-time rendering"],"falsifier":"Render a scene of overlapping semi-transparent anisotropic Gaussians with full volumetric ray tracing as ground truth and with this rasterizer at the same out-of-distribution poses; if the images differ in edge opacity or color by more than a small threshold, the maximum-along-ray model that the filter is built on is falsified. A cheaper check on a single Gaussian: render one highly elongated ellipsoid edge-on at two zoom levels and verify that its opacity stays constant under the perpendicular-area normalization, which is what Eq. (10) predicts.","tokens_in":15767,"feed_emoji":"🖼️","tokens_out":7599,"duration_ms":73991,"temperature":0.7,"pith_summary":"3D Gaussian Splatting renders scenes by projecting 3D Gaussians to 2D splats, and this paper argues that those projections are the root of the artifacts that appear when a trained scene is viewed from unusual angles, distances, or fields of view. AAA-Gaussians instead evaluates each Gaussian in full 3D throughout the entire rasterization pipeline, replacing the 2D splat approximation with a maximum-contribution-along-the-ray model. On top of that it adds an adaptive 3D smoothing filter whose amplitude scaling depends only on the Gaussian's area perpendicular to the view ray, so zooming out and moving close no longer cause flicker or over-transparency. The result is a rasterizer that matches the best published quality on standard test views while sharply outperforming them on out-of-distribution views, at frame times of 4-11 ms on a consumer GPU.","feed_headline":"3D Gaussian rendering goes artifact-free at 100+ FPS","feed_subtitle":"Evaluating Gaussians in 3D instead of 2D splats keeps image quality steady when views change.","key_machinery":"The load-bearing object is the 3D Gaussian evaluated along a viewing ray at its point of maximum contribution, inherited from hybrid transparency rendering via screen-space planes. Three new mechanisms carry the argument: (1) an adaptive 3D smoothing filter that combines the training-time sampling frequency $\\hat v_{\\text{train}}$ with the current view's frequency $\\hat v' = \\min(\\hat v_{\\text{train}}, \\hat v)$ and renormalizes each Gaussian by the perpendicular covariance determinant, which is what removes aliasing without over-transparency; (2) perspective-correct bounding that computes tangent angles $\\theta_{1,2}, \\phi_{1,2}$ of the cutoff ellipsoid in view space and clamps them to $[-\\pi/2+\\epsilon, \\pi/2-\\epsilon]$, which is what prevents popping for Gaussians crossing the image plane; and (3) frustum-based 3D tile culling, which constructs per-tile frusta and keeps a tile only if the minimum $\\rho(x)^2$ over the frustum is below the opacity threshold $\\tau_\\rho$, which is what makes hierarchical sorting fast enough for real time.","core_discovery":"The paper's central claim is that one rasterization pipeline can eliminate all three classic 3DGS artifacts—aliasing, popping, and projection distortion—at once by treating Gaussians as true 3D objects rather than 2D splats. Concretely, the paper replaces the 2D screen-space Mip filter with a 3D smoothing filter that dilates each Gaussian by the training sampling frequency and renormalizes amplitude using only the covariance projected onto the plane perpendicular to the ray, giving a closed-form scale factor $\\sqrt{|\\Sigma|\\,d^\\top\\Sigma^{-1}d/|\\hat\\Sigma|\\,d^\\top\\hat\\Sigma^{-1}d}$. It bounds Gaussians in view space by solving for the tangent angles of their cutoff ellipsoid, so Gaussians reaching behind the near plane are bounded stably instead of being discarded, and it lifts tile-based culling into 3D by constructing per-tile frusta and discarding tiles where the minimum $\\rho(x)^2$ inside the frustum exceeds the threshold. With these components the method reports state-of-the-art or matching metrics on Mip-NeRF 360, Tanks & Temples, and Deep Blending, large improvement over all baselines at 3x larger field of view, and effectively unchanged quality at half, full, and double resolution.","pith_inferences":["The perpendicular-area normalization is not specific to Gaussian splatting: any renderer that shades a primitive by its brightest point along a ray could reuse the same scale factor, so the formula is a candidate building block for other ray-max approximations.","The view-space bounding construction, although derived for a pinhole camera, is expressed in angles around the camera origin rather than screen coordinates; adapting it to fisheye or other central cameras may be straightforward, and would let the no-popping guarantee carry over to those models.","If future work moves to full volumetric integration along the ray, the max-along-ray approximation behind the filter calibration would need to be revisited: the paper's own derivation assumes only the maximum contribution matters, so a testable prediction is that scenes with many overlapping translucent Gaussians will be the first place the amplitude calibration deviates from a true volume render."],"forward_implications":["Rendering a trained scene at resolutions and fields of view far outside the training distribution no longer degrades: the large-FOV experiment holds PSNR essentially constant while all compared methods drop by several dB.","Zooming toward an object or pulling back no longer produces popping or flicker, because the 3D filter and view-space bounds keep contribution stable as the camera moves.","The pipeline keeps the MCMC densification training setup, so artifact-free rendering does not require a new representation or a new optimizer.","Timing measurements place the full method at 5.8-10.7 ms per frame, i.e. above 100 FPS on an RTX 4090, with culling recovering most of the cost of per-pixel sorting.","By removing view-inconsistent 'cheating' in the optimizer, the method exposes the view-dependent color encoding as the next bottleneck, so more expressive encodings should yield disproportionately larger gains."],"supporting_citations":[{"why":"Defines the 3DGS representation and affine projection that the paper replaces with 3D evaluation.","marker":"[14]"},{"why":"Supplies the screen-space plane 3D evaluation and maximum-contribution model that the paper inherits and extends.","marker":"[11]"},{"why":"Provides hierarchical sorting and per-tile culling whose 2D logic the paper lifts to 3D frusta.","marker":"[28]"},{"why":"Introduces the 3D smoothing filter and training sampling frequency that the paper's adaptive filter corrects.","marker":"[36]"},{"why":"Gives the ellipsoid plane-fitting used for exact bounds, which the paper moves from screen space to view space.","marker":"[30]"},{"why":"Provides the error-analysis and FOV stress test setup that the paper adopts for out-of-distribution evaluation.","marker":"[13]"},{"why":"Represents the Unscented Transform projection approach whose distortion artifacts the paper's 3D evaluation avoids.","marker":"[34]"},{"why":"Supplies the MCMC densification used as the training backend for the method.","marker":"[16]"}],"fun_headline_variants":["Rendering Gaussians as true 3D kills aliasing, popping, and distortion","True 3D Gaussian evaluation delivers artifact-free views","Aliasing, popping, distortion? Solved in 3D Gaussian rendering","One 3D pipeline ends aliasing, popping, and distortion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole anti-aliasing calibration rests on the approximation that a Gaussian's contribution to a pixel is its maximum value along the ray, not the integral of density along that ray; if true volumetric integration is required for correctness, the new perpendicular amplitude scaling would be miscalibrated.","fun_headline_variants_meta":{"raw":{"variants":["Rendering Gaussians as true 3D kills aliasing, popping, and distortion","True 3D Gaussian evaluation delivers artifact-free views","Aliasing, popping, distortion? Solved in 3D Gaussian rendering","One 3D pipeline ends aliasing, popping, and distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000732,"raw_usage":{"total_tokens":3301,"prompt_tokens":995,"completion_tokens":2306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":2229}},"tokens_in":611,"tokens_out":2306,"duration_ms":17870,"temperature":1.0,"reasoning_tokens":2229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:21:33.957417+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a scene of overlapping semi-transparent anisotropic Gaussians with full volumetric ray tracing as ground truth and with this rasterizer at the same out-of-distribution poses; if the images differ in edge opacity or color by more than a small threshold, the maximum-along-ray model that the filter is built on is falsified. A cheaper check on a single Gaussian: render one highly elongated ellipsoid edge-on at two zoom levels and verify that its opacity stays constant under the perpendicular-area normalization, which is what Eq. (10) predicts.","supporting_citations":[{"cited_title":"3D Gaussian Splatting for Real-Time Radi- ance Field Rendering","cited_arxiv_id":null,"evidence_quote":"Defines the 3DGS representation and affine projection that the paper replaces with 3D evaluation."},{"cited_title":"Efficient Perspective- Correct 3D Gaussian Splatting Using Hybrid Transparency,","cited_arxiv_id":null,"evidence_quote":"Supplies the screen-space plane 3D evaluation and maximum-contribution model that the paper inherits and extends."},{"cited_title":"StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering","cited_arxiv_id":null,"evidence_quote":"Provides hierarchical sorting and per-tile culling whose 2D logic the paper lifts to 3D frusta."},{"cited_title":"Mip-Splatting: Alias-free 3D Gaussian Splat- ting","cited_arxiv_id":null,"evidence_quote":"Introduces the 3D smoothing filter and training sampling frequency that the paper's adaptive filter corrects."},{"cited_title":"GPU-Based Ray-Casting of Quadratic Surfaces","cited_arxiv_id":null,"evidence_quote":"Gives the ellipsoid plane-fitting used for exact bounds, which the paper moves from screen space to view space."},{"cited_title":"On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy","cited_arxiv_id":null,"evidence_quote":"Provides the error-analysis and FOV stress test setup that the paper adopts for out-of-distribution evaluation."},{"cited_title":"3DGUT: Enabling Dis- torted Cameras and Secondary Rays in Gaussian Splatting,","cited_arxiv_id":null,"evidence_quote":"Represents the Unscented Transform projection approach whose distortion artifacts the paper's 3D evaluation avoids."},{"cited_title":"3D Gaussian Splatting as Markov Chain Monte Carlo","cited_arxiv_id":null,"evidence_quote":"Supplies the MCMC densification used as the training backend for the method."}],"review_version":1}