{"id":"03950d69-29fe-4e4a-8f14-801e8873b077","arxiv_id":"2412.17715","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GaussianPainter produces 3D Gaussians from a point cloud and reference image in one forward pass by constraining Gaussian rotations with predicted surface normals.","lead":"GaussianPainter turns a point cloud plus one reference image into a 3D Gaussian scene in a single feed-forward pass, using predicted surface normals to stabilize the otherwise unstable Gaussian rotation parameters. The method reports higher novel-view quality than prior feed-forward baselines on OmniObject3D and shows qualitative cross-object appearance transfer.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4 reports PSNR 27.5 for unperturbed GT point clouds while Table 1 reports 30.9 on the same OmniObject3D split; the unexplained 3.4 dB gap makes the headline SOTA number not self-consistent.","rationale":"The reader correctly identifies normal accuracy as a risk, but that concern is partially mitigated by the upper-bound study (Table 5) and the rotation ablations (Table 2). The Table 4 versus Table 1 PSNR gap is not mitigated anywhere and directly concerns the headline number. The paper's central claim is state-of-the-art novel view synthesis at 30.9 dB; if the same validation split and crop yield 27.5 dB for the unperturbed input, the claim cannot be reproduced as stated. I would keep a conditional verdict, with an explicit condition that the authors either reproduce 30.9 under the Table 4 protocol or disclose the protocol difference. The absence of error bars and the non-open-source baseline re-implementation are secondary concerns and would not by themselves move the verdict.","tokens_in":13306,"tokens_out":8596,"duration_ms":78515,"concrete_test":"Run the Table 4 protocol at 0% noise on the same OmniObject3D validation split and 600x600 center-crop used for Table 1, or have the authors specify the exact protocol difference. If PSNR reproduces about 30.9, Table 4 must state its different setup; if it reproduces about 27.5, the Table 1 SOTA claim is not reproducible under a consistent protocol. This single check settles whether the central quantitative claim is internally consistent.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 5.4's noisy-point-cloud experiment states: 'These point clouds are generated by perturbing a portion of the GT point clouds with Gaussian noise.' At 0% noise, the input is exactly the unperturbed GT point cloud, yet Table 4 reports PSNR 27.5. Table 1 in Section 5.3 reports PSNR 30.9 for the full model on the OmniObject3D validation split. The center-crop protocol ('we center-cropped ... to 600x600') is described in Section 5.3 as applying to 'our experiments', so a reader cannot tell whether Table 4 uses a different split or omits the crop. If both tables use the same split and protocol, the 3.4 dB discrepancy implies at least one reported number is wrong; if they use different protocols, that difference is not disclosed. Because the abstract and Section 5.3 present PSNR 30.9 as the headline SOTA result, this unresolved numerical self-inconsistency is the most load-bearing concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GaussianPainter, a feed-forward method that converts a point cloud and a reference image into a 3D Gaussian field in a single forward pass. To address the non-uniqueness of Gaussian parameters, the method predicts a surface normal for each point and uses it to define the Gaussian rotation via Rodrigues' formula, leaving the network to predict scales, opacity, and color in a constrained space. A multiscale triplane module injects reference image appearance using DINOv2 features and cross-attention. The method is evaluated on OmniObject3D and Objaverse, reporting state-of-the-art novel view synthesis (PSNR 30.9 vs AGG 28.4 and TriplaneGaussian 25.8) and demonstrating qualitative cross-object appearance transfer.","tokens_in":13535,"tokens_out":5271,"duration_ms":44182,"significance":"If the reported numbers are reliable, the paper makes a useful contribution by showing that a feed-forward Gaussian predictor can be stabilized through normal-guidance, avoiding per-scene optimization while retaining anisotropic Gaussians. The upper-bound study (Table 5) is a thoughtful check on the expressiveness of the constrained rotation representation, and the code release supports reproducibility. However, the numerical self-consistency concerns detailed below must be resolved before the central SOTA claim can be accepted.","major_comments":[{"comment":"Table 4 reports PSNR 27.5 for 0% noisy points, which is exactly the input of unperturbed GT point clouds, while Table 1 reports PSNR 30.9 for the full model on the OmniObject3D validation split. The center-crop protocol in Section 5.3 is described as applying to 'our experiments', so it should also apply to Table 4 unless otherwise stated. If both tables use the same split and protocol, the 3.4 dB gap is an internal inconsistency that invalidates at least one reported number; if they differ, that difference is not disclosed. Moreover, Table 5 reports 29.16 PSNR for per-scene optimized unconstrained rotation, which is lower than the feed-forward 30.9 in Table 1, suggesting further protocol discrepancies across tables. Please reconcile these numbers or explicitly state the evaluation protocol used in each table.","section":"Sec. 5.4, Table 4 vs Sec. 5.3, Table 1"},{"comment":"The text says 'training the rotation prediction model from scratch is quite unstable and the model is hard to be well-converged. Thus, we utilize the pretrained weights of our normal-guided model as initialization and fine-tune the rotation prediction model.' Yet Table 2 lists both 'Rot. (from scratch)' and 'Rot. (fine-tuned)†'. The reader cannot tell whether a from-scratch model was actually trained and under what protocol, or whether the from-scratch row is a separate experiment. If the from-scratch result is included as evidence for the instability claim, its training setup (init, schedule, epochs) must be described. Please clarify the relationship between the text and the two rows.","section":"Sec. 5.4, Table 2 and surrounding text"},{"comment":"The row labeled 'Normal-guided rotation (ours)' in Table 5 is obtained by optimizing normals jointly with the other Gaussian parameters for each scene, not by using the feed-forward normal predictor of GaussianPainter. As an upper-bound study this is a valid way to test the expressiveness of the rotation representation, but the phrase 'ours' is misleading because it does not reflect the end-to-end pipeline where the normal may be imperfect. Please rename the row and add a sentence stating that the feed-forward normal estimation quality is not evaluated by this experiment.","section":"Sec. 5.5, Table 5"}],"minor_comments":[{"comment":"The Instability Score is defined using 'standard variance' which seems to mean standard deviation or variance; the formula should use consistent terminology and specify whether the variance or its square root is used.","section":"Sec. 3"},{"comment":"The rotation axis r = n × z is degenerate when n is parallel to z (the vertical direction); the paper should specify how this case is handled numerically to avoid division by zero.","section":"Sec. 4.1, Eq. (2)"},{"comment":"For the baseline comparisons, the paper states that AGG and TriplaneGaussian are reimplemented or run with GT point clouds substituted; please specify in Sec. 5.3 whether the Objaverse cross-object experiments also use the same substitution, and whether the numbers in Table 1 are obtained from the official code or the reimplementation.","section":"Sec. 5.3, compared methods"},{"comment":"The 'Semantic understanding in the texture injection' subsection relies entirely on qualitative inspection of a few examples and refers to 'OmniObject' instead of 'OmniObject3D'; the claim of semantic understanding would be stronger with a quantitative or at least more systematic evaluation.","section":"Sec. 5.5, semantic understanding"},{"comment":"There are minor terminological and typographical issues: 'KeyWords' should be 'Keywords', 'multi-scale' is sometimes written 'multiscale', and 'OminiObject3D' appears instead of 'OmniObject3D'.","section":"Abstract and throughout"},{"comment":"All quantitative comparisons are reported as single numbers without error bars or multiple-seed statistics; given the small gaps in some ablations (e.g., Table 3, multi-scale 30.6 vs full model 30.9), the authors should state whether these are single runs or averaged over seeds.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the paper is generally well written, but the unresolved discrepancy between Table 4 (27.5 at 0% noise) and Table 1 (30.9) on the same dataset is a serious correctness concern that must be addressed before the paper can be recommended for acceptance. The rotation ablation ambiguity in Section 5.4 further weakens the central argument. If the authors can clarify protocols and correct the tables, the paper may become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things stand out. The normal-guided rotation is a smart, simple fix for the non-uniqueness of Gaussian parameters in feed-forward prediction, and the pilot study in Sec. 3 gives it a real empirical basis. The paper also deserves credit for the upper-bound experiment in Table 5, which directly checks whether the constrained rotation hurts capacity; the small drop there suggests the design is safe.\n\nBut there is a numerical inconsistency that sits on the main claim. Table 1 gives the full model 30.9 PSNR on the OmniObject3D validation split. Table 4, which measures robustness to noisy point clouds, reports 27.5 PSNR at 0% noise — i.e., the exact same unperturbed GT point clouds. The paper says the center-crop protocol applies to 'our experiments,' so unless Table 4 uses a different split or omits the crop, the two numbers cannot both be correct. The paper never explains the gap. That makes the headline SOTA number shaky.\n\nOther soft spots are smaller. The 'Rot. (from scratch)' ablation in Table 2 is actually initialized from the normal-guided model and fine-tuned, so it does not test the claimed difficulty of learning rotations from scratch — the paper admits this, but it undercuts the comparison. Cross-object appearance transfer rests only on qualitative figures, and there are no error bars anywhere. None of these are fatal on their own, but they all push in the same direction: treat the specific numbers as provisional until the authors clarify the protocol.\n\nWho is this for? Anyone working on feed-forward 3D generation or 3DGS prediction. The idea is worth borrowing even if the exact numbers need verification. It deserves a serious referee — the core mechanism is novel and the analysis is mostly honest — but the referee should ask for the Table 1/Table 4 discrepancy to be resolved before publication. I would not desk reject it; I would send it to review with that concern flagged.","headline":"Normal-guided rotation is a genuine improvement, but the 30.9 vs 27.5 PSNR discrepancy between the main table and the robustness table undermines the headline result.","tokens_in":14062,"tokens_out":3074,"would_cite":false,"duration_ms":28202,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that normal-guided rotation resolves the non-uniqueness that makes feed-forward Gaussian prediction unstable, achieving the best compared novel-view results (PSNR 30.9) and cross-object appearance transfer.","keywords":["3D Gaussian splatting","surface normal","point cloud painting","feed-forward 3D generation","non-uniqueness","novel view synthesis","appearance transfer","triplane injection"],"falsifier":"Train a second feed-forward model with identical capacity and data but unconstrained rotations and a stabilized training scheme; if it matches or exceeds PSNR 30.9 on the benchmark used in the main comparison, then the claimed non-uniqueness bottleneck is not the limiting factor and normal guidance is not necessary.","tokens_in":1520,"feed_emoji":"🎨","tokens_out":1643,"duration_ms":91594,"temperature":0.7,"pith_summary":"GaussianPainter aims to turn any point cloud into a renderable 3D Gaussian field — a set of points carrying rotation, scale, opacity, and color that can be rendered in real time — from a single reference image in one feed-forward pass. The paper's central argument is that feed-forward Gaussian prediction is unstable because the parameter space is non-unique: substantially different Gaussian fields can render nearly identical images, so a network cannot settle on consistent parameters. Its fix is to constrain each Gaussian's rotation with a predicted surface normal, reducing rotation to a 2-degree-of-freedom alignment and making the remaining parameters learnable. A multiscale triplane module injects the reference image's appearance into the point-cloud decoder. With this design the method reports the best compared novel-view synthesis results on a large object benchmark (PSNR 30.9, SSIM 0.945, LPIPS 0.134) and demonstrates transferring appearance between objects of different shapes.","feed_headline":"Normal guidance unlocks feed-forward Gaussian painting: 30.9 dB","feed_subtitle":"Surface normals fix the unstable rotations that blocked one-pass 3D Gaussian prediction and style transfer.","key_machinery":"The load-bearing mechanism is normal-guided rotation: for each occupied point the network predicts a unit normal $n$, and the Gaussian's rotation matrix $R$ is defined as the rotation that aligns the world $z$-axis with $n$, with rotation axis $r = n \\times z$ and angle $\\theta = \\angle(z, n)$, computed via the standard axis-angle rotation formula. This collapses the Gaussian rotation to two degrees of freedom and leaves the network to predict only scales, opacity, and spherical-harmonic colors. Normals are trained through an Isotropic Normal Rendering module that renders them into a 2D normal map and supervises with L1 and SSIM losses, avoiding the need for direct 3D normal labels. Appearance is injected separately by a multiscale triplane — three axis-aligned feature planes indexed by 3D position — built from self-supervised visual Transformer features, cross-attended with occupancy features, and added into the point-cloud decoder at several scales.","core_discovery":"The paper establishes that the rotation parameters, rather than opacity or spherical-harmonic colors, are the main source of non-uniqueness in Gaussian fields: a pilot study re-initializes rotations, scales, opacities, and SH coefficients while fixing locations, and reports instability scores showing rotations and scales diverge most across re-initializations. Because rotations and scales are entangled, GaussianPainter treats stable rotation as the key subproblem and proposes normal-guided rotation. Each point receives a surface normal estimated from occupancy features and rendered to a 2D normal map for supervision; the Gaussian's rotation is then defined by rotating the vertical axis onto that normal using the standard axis-angle formula. With rotation fixed, the network predicts only scales, opacity, and spherical-harmonic colors in a constrained space. The paper shows this outperforms isotropic Gaussian prediction and direct rotation prediction, and that the resulting model can transfer appearance across objects by learning part-level semantics.","pith_inferences":["A testable extension is to apply the same instability-score diagnostic to other feed-forward predictors, such as relighting or material parameters, to identify which parameter channels need geometric constraints.","Because the normal-guided rotation leaves one rotational degree of freedom unmodelled — roll around the normal — objects with strongly directional surface detail such as hair, fur, or brushed metal may require an additional per-point roll parameter.","The apparent semantic correspondence in cross-object transfer suggests the triplane injection is not merely copying 2D patches; probing with adversarial reference images could reveal whether part-level understanding is robust or an artefact of dataset bias.","If normals were supplied by geometry rather than learned, the method could factorize geometry and appearance, potentially painting point clouds into Gaussians without any reference image."],"forward_implications":["Feed-forward painting can replace per-object optimization for object-level Gaussian generation, turning any point cloud, whether scanned, generated, or noisy, into a renderable Gaussian field in one pass.","Anisotropic Gaussians with normal-guided rotation outperform isotropic Gaussians, so constraining rotation need not give up the expressiveness that makes 3D Gaussian splatting attractive.","Appearance transfer generalizes across objects with different shapes when part semantics are understood, enabling style transfer between unrelated 3D assets.","Multi-reference control and varying visual-feature layers allow style interpolation and diverse painting outcomes at inference time without retraining.","The method degrades gracefully on noisy point clouds, with rendered quality falling only modestly as the fraction of perturbed points rises to 90 percent."],"supporting_citations":[{"why":"Defines the differentiable 3D Gaussian representation and rasterizer that the method's output must render through.","marker":"(Kerbl et al. 2023)"},{"why":"Documents the non-uniqueness problem and provides the isotropic-Gaussian alternative that normal guidance is designed to beat.","marker":"(Xu et al. 2024)"},{"why":"Provides the feed-forward baseline and the point-cloud generator used to test generalization to generated points.","marker":"(Zou et al. 2023)"},{"why":"Supplies the visual features that the triplane appearance-injection module lifts into 3D.","marker":"(Oquab et al. 2023)"},{"why":"Provides the object dataset with point clouds, multi-view renders, and normal maps used for training and evaluation.","marker":"(Wu et al. 2023)"}],"fun_headline_variants":["Normal-guided rotations stabilize one-pass 3D Gaussian painting","First feed-forward method paints point clouds into 3D Gaussians","Fix rotation to paint point clouds into 3D Gaussians in one pass","Surface normals solve rotation chaos in fast 3D Gaussian style transfer","One-pass 3D Gaussian painting via normal-constrained rotations"],"cache_read_input_tokens":16256,"weakest_assumption_plain":"The method assumes that surface normals estimated from sparse occupancy features, supervised only through rendered 2D normal maps, are accurate enough to fix Gaussian rotations without losing the anisotropic expressiveness that makes 3D Gaussian splatting powerful.","fun_headline_variants_meta":{"raw":{"variants":["Normal-guided rotations stabilize one-pass 3D Gaussian painting","First feed-forward method paints point clouds into 3D Gaussians","Fix rotation to paint point clouds into 3D Gaussians in one pass","Surface normals solve rotation chaos in fast 3D Gaussian style transfer","One-pass 3D Gaussian painting via normal-constrained rotations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1327,"prompt_tokens":949,"completion_tokens":378,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":565,"tokens_out":378,"duration_ms":3625,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:14:25.637294+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a second feed-forward model with identical capacity and data but unconstrained rotations and a stabilized training scheme; if it matches or exceeds PSNR 30.9 on the benchmark used in the main comparison, then the claimed non-uniqueness bottleneck is not the limiting factor and normal guidance is not necessary.","supporting_citations":[],"review_version":1}