{"id":"77e09dff-9cda-4d32-aa6e-6bc5d3d71bd4","arxiv_id":"2504.12800","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CAGE-GS deforms a source 3DGS model toward a target shape in any of five input formats, using a learned deformation cage and Jacobian-based Gaussian covariance updates to preserve texture.","lead":"CAGE-GS reshapes a 3D Gaussian Splatting model so it matches a target given as text, image, mesh, point cloud, or another 3DGS, using a learned cage plus a Jacobian-based covariance update that keeps textures sharp. A generalist reader may care because it makes 3D content editing, animation, and scene creation faster and less manual.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"kNN Jacobian sharing is the load-bearing assumption for texture fidelity; Fig. 10 alone does not establish it under large or high-curvature deformations.","rationale":"The paper's strongest claim is that Jacobian-based covariance updating preserves texture and outperforms baselines. Equation (7) is the standard first-order covariance transport and is internally consistent. The fragile step is not Eq. (7) itself but the decision to compute J only at 10,000 sampled Gaussians and propagate by kNN. For that step to deliver the claimed texture fidelity, the deformation gradient must be nearly constant on the scale of each neighborhood. Cage MVC deformations are smooth but have significant gradient variation near cage features and under large deformations; the user study and DINO numbers in Table 1 have no error bars, and Figure 10 is one visual example with no worst-case or large-deformation quantitative check. The paper's stated limitation that straight lines, planes, and parallel structures are not guaranteed is honest and consistent with this concern. I do not see an internal contradiction in the math, and the method has a plausible pipeline with ablations, so the reader's CONDITIONAL verdict is appropriate. The concrete test compares kNN propagation against exact per-Gaussian Jacobians; if the metrics are close, the concern is resolved, and if they are not, the fidelity claim should be revised.","tokens_in":11565,"tokens_out":4883,"duration_ms":52046,"concrete_test":"Re-run the paper's ShapeNet chair, car, and table deformations with a large target deformation (e.g., lambda=1.0 and a target with strong local bending) and compare two variants on the same optimized cage: (A) the paper's kNN Jacobian sharing with m=10,000, and (B) exact per-Gaussian Jacobians computed at every Gaussian center. Report DINO, CD, and rendered artifacts, plus the mean and max Frobenius error between assigned and exact Jacobians, stratified by distance to cage vertices. If variant A matches variant B within metric noise, the locality assumption holds; if DINO drops or visible artifacts appear, the kNN step is the load-bearing failure and the fidelity claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central texture-preservation claim rests on Eq. (7), Sigma' = J R S S^T R^T J^T, with J evaluated only at m=10,000 sampled Gaussians and then assigned to unsampled Gaussians by kNN. This is a locality assumption: the deformation gradient must be nearly constant over each kNN neighborhood, which typically contains 10-50 Gaussians on 100k-500k models. MVC-based cage deformation is smooth but not affine; near cage edges or vertices, and for large lambda interpolation, the Jacobian varies rapidly, so a borrowed first-order Jacobian can mis-rotate or mis-scale anisotropic splats and produce exactly the blur and elongation artifacts the method claims to remove. The only supporting evidence, Fig. 10, is a single visual example; Table 2 measures runtime, not fidelity. The paper's own limitation (no guarantee on straight lines, planes, and parallel structures) is consistent with this weakness but is not a substitute for a stress test. This does not contradict Eq. (7) mathematically; it means the headline fidelity advantage over NeuralCage and GSDeformer is under-supported exactly where the Jacobian update is the differentiator.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CAGE-GS proposes a cage-based deformation pipeline for 3D Gaussian Splatting (3DGS) models. The method first predicts a deformed cage from source and target point clouds using a per-model optimized neural network (Sec. 3.2), then updates Gaussian centers via mean value coordinates (Eq. 2), and finally updates the covariances as Sigma' = J R S S^T R^T J^T using the Jacobian of the deformation, computed for 10,000 sampled Gaussians and shared with the rest via kNN (Eqs. 6-7, Sec. 3.3). Experiments on ShapeNet and real-captured objects compare with CC, NeuralCage, and GSDeformer, reporting similar Chamfer distance yet higher DINO similarity and user preference (Table 1). Ablations isolate the contributions of the cage deformation, the Jacobian update, and the sampling strategy.","tokens_in":11671,"tokens_out":7502,"duration_ms":73570,"significance":"The paper addresses a timely problem: structure-aware deformation of 3DGS while preserving texture. The Jacobian-based covariance update is a natural extension of cage deformation to Gaussian parameters, and is mathematically sound under the first-order approximation. Supporting multiple target representations (text, image, point cloud, mesh, 3DGS) is practically useful. The per-model optimization avoids large training data requirements. However, the evaluation currently overstates generality and efficiency: the geometric alignment metric is partly circular because the cage is fitted to the target, the kNN Jacobian sharing is not quantitatively validated, and no runtime comparison with baselines is provided. These issues are addressable with additional experiments and clarifications.","major_comments":[{"comment":"The cage prediction module is optimized per model against the target point cloud (Sec. 3.2: 'we remove the pre-training process in NeuralCage and perform the optimization process per model'). Consequently, the Chamfer distances in Table 1 measure fit quality against the target, not deformation transfer or generalization. The comparison with NeuralCage and GSDeformer is therefore difficult to interpret unless those baselines are also optimized per model against the same target. Please state explicitly how each baseline was run (pre-trained weights vs. per-model optimization) and, if they are not per-model optimized, add an ablated baseline with the same treatment. In addition, to support the claimed deformation-transfer capability, please provide a held-out experiment where the network is tested on a source/target pair not used during the per-model optimization, or clarify that the method is a per-instance fitting approach rather than a learned transfer.","section":"Sec. 3.2, Table 1"},{"comment":"The claim that the kNN-based Jacobian sharing 'barely affect[s] rendering quality' is supported only by a single visual example (Fig. 10) and by runtime data (Table 2); no quantitative fidelity metric is reported for the sampled versus full-Jacobian variants. Since the entire texture-preservation advantage relies on the assumption that the deformation Jacobian is nearly constant over each kNN neighborhood, please provide a quantitative ablation (e.g., DINO similarity or rendered-view PSNR/LPIPS) across several objects and deformation magnitudes, including a large-magnitude case (lambda near 1.0) and a case with strong local curvature or non-uniform scaling. If the assumption breaks down in such regimes, state the resulting limitation explicitly.","section":"Sec. 3.3, Fig. 10, Table 2"},{"comment":"The abstract claims the method 'significantly outperforms existing techniques in both efficiency and deformation quality,' yet no runtime comparison with any baseline is reported. The only efficiency evidence is the internal sampling ablation in Table 2, which compares the method with itself (with and without sampling). Please report the end-to-end runtime of GSDeformer and NeuralCage on the same data and hardware, or temper the efficiency claim to avoid overstatement.","section":"Abstract, Sec. 4.1"}],"minor_comments":[{"comment":"The manuscript contains several typos and grammatical errors, including 'our method outperforming existing techniques' (end of Sec. 1), 'plains' for 'planes' (Sec. 5), and inconsistent spelling of 'Jacobi' vs. 'Jacobian'. A careful proofread is recommended.","section":"Throughout"},{"comment":"The value of k in the k-nearest neighbors assignment is never specified. Please report the value used and analyze its sensitivity, since it directly controls the locality approximation.","section":"Sec. 3.3"},{"comment":"In the covariance decomposition, Eq. (7) is correct, but the preceding text and equations typeset Sigma as 'RSS TRT' with missing superscripts; please correct the notation for clarity.","section":"Sec. 3.1"},{"comment":"For text and image targets, the paper states that a learning-based method generates a rough 3D proxy, but only AtlasNet is named (for images). Please specify the method used for text inputs and provide implementation details for reproducibility.","section":"Sec. 4.1"},{"comment":"The user study report ('60 participants') lacks details on the number of questions, the statistical test used, and confidence intervals. Please add these to support the preference claim.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable incremental contribution that combines NeuralCage-style cage prediction with 3DGS and a Jacobian-based covariance update. The main risk is that the evaluation protocol makes the method appear stronger than the evidence supports: the per-model fitting makes the CD metric circular, the kNN Jacobian sharing is under-validated exactly where the method differentiates itself, and the efficiency claim is not benchmarked against baselines. These issues are fixable with additional experiments and honest framing. I do not see a fundamental flaw in the core derivation, so rejection is not warranted, but the claims need to be scaled to the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CAGE-GS is a competent engineering paper. It puts together NeuralCage-style learned deformation cages and a Jacobian-based covariance update for 3DGS, and it supports a useful range of target representations. Nothing here is conceptually new on its own, but the package is new and effectively demonstrated. The math is straightforward and the paper reads honestly.\n\nWhat it does well: the pipeline is clearly described, the ablation isolates the Jacobian update as the source of texture improvement, and the authors state a real limitation (no guarantees on straight lines, planes, or parallel structures). The multi-representation support is a nice practical feature.\n\nSoft spots, in order of seriousness. First, the paper cites Gao et al. (real-time large-scale deformation of Gaussian splatting) but never compares against it. That is the closest prior art for cage-based 3DGS deformation, and the novelty increment remains unproven without a direct comparison. Second, the quantitative support is thin: no error bars, the CD advantage over cage baselines is at the fourth decimal, and the DINO metric is presented without any validation that it tracks perceived texture quality. Third, and most important, the texture-fidelity claim rests on the kNN Jacobian sharing described in Sec. 3.3. The deformation gradient is assumed nearly constant over each kNN neighborhood, and that is exactly where large or high-curvature deformations could break. The only evidence is one visual example in Fig. 10 and a runtime table; there is no quantitative check on worst-case or large-magnitude deformations. The stress-test note is fair. None of this contradicts the math, but it means the headline texture-fidelity advantage is under-supported where it matters most.\n\nThe authors' own limitation paragraph is consistent with this weakness, so I don't read it as a hidden defect.\n\nBottom line: this is a solid paper worth engaging with. It is not a field reorg, but it is a useful step for 3DGS content creation. Send it to peer review. Ask the authors to add a comparison with Gao et al., report variance across shapes, and stress-test the kNN Jacobian sharing with quantitative metrics on large deformations and high-curvature regions. If those come back clean, the paper would be a reasonable accept.","headline":"Solid engineering contribution combining learned cages with Jacobian covariance updates; the texture-fidelity claim needs a stronger stress test on the kNN Jacobian sharing.","tokens_in":12332,"tokens_out":2147,"would_cite":true,"duration_ms":21461,"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":"The paper claims that a learned cage plus a Jacobian-based covariance update can deform a 3D Gaussian Splatting scene to a target shape while preserving texture detail.","keywords":["3D Gaussian Splatting","cage-based deformation","deformation transfer","Jacobian matrix","covariance update","mean value coordinates","texture preservation","3D editing"],"falsifier":"Deform a source scene toward a target with a sharp bend, set the interpolation to lambda=1.0, and compare high-curvature regions against a version that computes a separate Jacobian for every Gaussian; if the nearest-neighbor version shows visible blurring or stretching, the near-constant local deformation assumption fails.","tokens_in":11253,"feed_emoji":"🎨","tokens_out":9410,"duration_ms":84890,"temperature":0.7,"pith_summary":"Deforming a 3D Gaussian Splatting scene to match a user-specified target usually degrades texture, because moving Gaussian centers without adjusting their covariance blurs or stretches the rendering. This paper proposes learning a deformation cage from the target shape and then updating each Gaussian's covariance with the Jacobian of the deformation field. The central claim is that this combination matches the geometric alignment of earlier cage-based deformers while preserving texture better, as measured by directional perceptual similarity and a 63.3 percent user-study preference. The method accepts target shapes given as text, image, point cloud, mesh, or another 3DGS model, and a sampling scheme brings deformation time to about eight minutes for a 200k-Gaussian model.","feed_headline":"Learned cages deform 3D scenes without smearing textures","feed_subtitle":"A Jacobian-based covariance update keeps Gaussian splats sharp while the cage controls the shape.","key_machinery":"The load-bearing object is the deformation cage together with the Jacobian of the cage-induced map: a coarse control mesh predicted from source and target point clouds, with mean value coordinates converting cage-vertex motion into smooth displacements of every Gaussian center. The Jacobian $J = \\partial \\mu'_{\\mathrm{sample}} / \\partial \\mu_{\\mathrm{sample}}$ is computed at sampled centers and applied through $\\Sigma' = J R S S^T R^T J^T$ to rotate and scale each Gaussian's covariance so the splat follows the local deformation; nearest-neighbor assignment spreads each Jacobian to unsampled Gaussians. This keeps the anisotropic shape of splats coherent, which is what preserves texture detail after deformation.","core_discovery":"The paper's central claim is that cage-based deformation transfer can be made compatible with 3D Gaussian Splatting without sacrificing texture fidelity. Gaussian centers are bound to a learned cage through mean value coordinates and displaced by $p' = \\sum_i \\omega_i(p) v_{s\\to t}^i$; each covariance is then transformed as $\\Sigma' = J R S S^T R^T J^T$, where $J = \\partial \\mu'_{\\mathrm{sample}} / \\partial \\mu_{\\mathrm{sample}}$ is the Jacobian of the deformed positions. The paper further claims that computing Jacobians on 10,000 sampled Gaussians and assigning each to nearby unsampled Gaussians by nearest neighbors preserves rendering quality while cutting runtime from roughly 170 minutes to about 7–8 minutes. Quantitatively, it reports a Chamfer distance tied with the cage-based baselines (0.0997 vs. 0.0998), a higher directional perceptual similarity (0.402 vs. 0.385 and 0.374), and 63.3 percent user preference.","pith_inferences":["Beyond the paper's claims, the nearest-neighbor Jacobian assignment implies a testable failure mode: large-magnitude or sharply bending deformations, such as interpolation at lambda=1.0 near cage edges, should show more blur or stretching in high-curvature regions than in the paper's chair or car examples.","Beyond the paper's claims, the gap between geometric metrics (Chamfer distance) and perceptual metrics (directional similarity, user preference) suggests texture fidelity and geometric alignment can be evaluated separately; reporting both would let future methods make that trade-off explicit.","Beyond the paper's claims, the covariance update depends only on the deformation Jacobian, so the same recipe could transfer to other rasterized representations with covariances, such as 2D Gaussian splatting or surfel-based renderers.","Beyond the paper's claims, an end-to-end version that learns Jacobian fields directly, which the paper lists as future work, could remove the nearest-neighbor approximation and extend the method to articulated or extreme deformations."],"forward_implications":["A source 3DGS scene can be reshaped to match a target supplied as text, image, point cloud, mesh, or another 3DGS model, without manual cage editing.","Deformation magnitude can be controlled continuously by interpolating between the source and deformed cages, yielding smooth partial deformations at any intermediate value.","Updating covariances with the deformation Jacobian preserves texture better than moving Gaussian centers alone, which otherwise blurs or elongates splats.","Sampling 10,000 Gaussians for Jacobian computation cuts runtime from roughly 170 minutes to about 7–8 minutes on typical models while rendering quality is reported to be barely affected.","The method does not guarantee that straight lines, planes, or parallel structures remain unchanged, especially in artificial shapes."],"supporting_citations":[{"why":"Defines the 3D Gaussian Splatting representation (centers, covariances, colors, opacities) that the method deforms.","marker":"[21]"},{"why":"Supplies mean value coordinates used to interpolate cage deformation to Gaussian centers.","marker":"[20]"},{"why":"Provides the cage-learning approach the paper adapts to predict deformed cages from source and target point clouds.","marker":"[44]"},{"why":"Supplies the Jacobian matrix concept used to update Gaussian covariances under deformation.","marker":"[33]"},{"why":"Cage-based 3DGS deformation baseline against which the method is compared and extended.","marker":"[15]"},{"why":"Cycle-consistency deformation baseline used in qualitative and quantitative comparisons.","marker":"[12]"},{"why":"Shape dataset used to construct source and target models for experiments.","marker":"[6]"}],"fun_headline_variants":["Cage-based Gaussian splatting deforms 3D scenes 20x faster","Cages control 3D scene shape while Jacobians keep textures sharp","CAGE-GS: Deform 3D scenes with cages, no texture smearing","One cage to deform them all: 3D Gaussian splatting made fast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the deformation field is locally nearly linear, so a single Jacobian computed at a sampled Gaussian center accurately describes the motion of all nearby unsampled Gaussians.","fun_headline_variants_meta":{"raw":{"variants":["Cage-based Gaussian splatting deforms 3D scenes 20x faster","Cages control 3D scene shape while Jacobians keep textures sharp","CAGE-GS: Deform 3D scenes with cages, no texture smearing","One cage to deform them all: 3D Gaussian splatting made fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000673,"raw_usage":{"total_tokens":3069,"prompt_tokens":957,"completion_tokens":2112,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2024}},"tokens_in":573,"tokens_out":2112,"duration_ms":14049,"temperature":1.0,"reasoning_tokens":2024,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:22:41.687129+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deform a source scene toward a target with a sharp bend, set the interpolation to lambda=1.0, and compare high-curvature regions against a version that computes a separate Jacobian for every Gaussian; if the nearest-neighbor version shows visible blurring or stretching, the near-constant local deformation assumption fails.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Defines the 3D Gaussian Splatting representation (centers, covariances, colors, opacities) that the method deforms."},{"cited_title":"Mean value coor- dinates for closed triangular meshes","cited_arxiv_id":null,"evidence_quote":"Supplies mean value coordinates used to interpolate cage deformation to Gaussian centers."},{"cited_title":"Neural cages for detail-preserving 3d deformations","cited_arxiv_id":null,"evidence_quote":"Provides the cage-learning approach the paper adapts to predict deformed cages from source and target point clouds."},{"cited_title":"A modern ap- proach to classical theorems of advanced calculus, 1965","cited_arxiv_id":null,"evidence_quote":"Supplies the Jacobian matrix concept used to update Gaussian covariances under deformation."},{"cited_title":"Gsdeformer: Direct, real-time and extensible cage-based deformation for 3d gaussian splat- ting, 2024","cited_arxiv_id":null,"evidence_quote":"Cage-based 3DGS deformation baseline against which the method is compared and extended."},{"cited_title":"Unsupervised cycle- consistent deformation for shape matching","cited_arxiv_id":null,"evidence_quote":"Cycle-consistency deformation baseline used in qualitative and quantitative comparisons."}],"review_version":1}