{"id":"4cb93449-1e18-4ea0-9fe8-e31e7e5ca206","arxiv_id":"2412.00845","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-stage surface-aligned Gaussian representation (adhere-then-detach) improves monocular human avatar synthesis and enables direct mesh extraction.","lead":"A new two-stage method builds animatable 3D avatars from a single monocular video by first gluing Gaussian splats to a body mesh, then releasing them with a soft alignment constraint. It reports faster training, real-time rendering, and better novel-view/pose quality than existing NeRF and Gaussian avatar methods, plus a way to extract clean meshes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central two-stage claim presupposes that the optimized SMPL mesh tracks the true clothed surface; on loose clothing, the Detached-Stage alignment losses could pull Gaussians toward incorrect geometry, and the paper provides no evidence this regime holds.","rationale":"The central mechanism of SAGA is that mesh alignment regularizes Gaussians; this only helps if the mesh approximates the true surface. The reader's weakest assumption is exactly this, and I concur. The two-stage design makes the assumption more acute: the Adhered Stage jointly optimizes mesh vertices and barycentrics, so the mesh can move toward the observed silhouettes, but no geometric supervision (e.g., depth, multi-view consistency) is used to ensure the result tracks loose clothing. The Detached Stage then makes the mesh a hard regularizer through position and orientation losses. If the mesh is wrong, the losses are wrong and the regularizer corrupts geometry and deformation. The paper's own data indicate the benefit is small in the tested regime (Table 5: 0.03 PSNR gain; Table 4: full two-stage only 0.06 PSNR over single-stage variants), so a regime where the mesh is inaccurate could plausibly flip the sign. Other issues, such as the unreproducible '37%/13%' claim and unverified mesh-extraction novelty, are reporting/verifiability problems and do not attack the mechanism as directly. Thus the prescription is a targeted experiment on loose clothing, which would settle whether the central claim generalizes beyond tight athletic wear. Since this is the same concern that motivated the CONDITIONAL verdict, I recommend keeping the verdict UNCHANGED.","tokens_in":23500,"tokens_out":3963,"duration_ms":37337,"concrete_test":"Train SAGA from scratch on a monocular sequence with loose/draped clothing (e.g., a skirt or wide-sleeve garment, with human pose estimates), then compare the full model against the 'w/o alignment loss' ablation from Table 5 on novel-view LPIPS. Also compute the mean Chamfer distance between the optimized canonical mesh and a reference surface (e.g., from multi-view TSDF or a fitted SMPL-X+D). If the full model does not beat w/o-alignment on LPIPS, or the Chamfer distance exceeds the typical scale of Gaussian scales, the proxy-fidelity assumption underlying the Detached Stage fails in that regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the mesh remains a faithful proxy for the true surface after the Adhered Stage. In the Detached Stage, Eq. 12 penalizes distance from each Gaussian center to its bound triangle and Eq. 13 penalizes normal deviation, with Walking-on-Mesh (Sec. 4.3.2) only re-binding to adjacent triangles. If the canonical SMPL mesh deviates strongly from the real clothed surface—which is exactly what happens with loose garments or unusual body shapes—these losses actively bias the optimized Gaussians toward wrong positions and orientations, and the deformation regularizer (bound triangles fixed per Gaussian) propagates that bias to novel poses. The paper never quantifies mesh-surface discrepancy; all test subjects wear relatively tight clothing. The ablations in Table 5 on such subjects show only a marginal effect of the alignment loss (PSNR 31.08→31.11, SSIM 0.9726→0.9728), so the benefit of the core regularization is already small in the tested regime, and could easily become negative when the proxy is inaccurate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAGA, a two-stage surface-aligned Gaussian representation for animatable human avatars reconstructed from monocular video. In the first Adhered Stage, Gaussian centers are parameterized by barycentric coordinates on SMPL mesh triangles and jointly optimized with the mesh vertices, while the Gaussians are flattened and their normals fixed to the triangle normals, enforcing strict surface adherence while allowing the Gaussians to flow on the mesh. In the second Detached Stage, the Gaussians are freed but kept under a position-alignment loss (Eq. 12) and a normal-alignment loss (Eq. 13) relative to their bound triangles, and a Walking-on-Mesh strategy (Algorithm 1) re-binds Gaussians that drift outside their triangles. A non-rigid deformation module and a pose-conditioned colorization MLP handle motion and appearance change. Experiments on ZJU-MoCap, MonoCap, and PeopleSnapshot report state-of-the-art or competitive novel-view and novel-pose results with roughly 12-minute training and 60+ FPS rendering, and the paper claims the first direct high-quality mesh extraction from deformable Gaussians learned from monocular video.","tokens_in":23733,"tokens_out":18973,"duration_ms":155622,"significance":"If the results hold, the adhere-then-detach idea is a genuinely useful contribution to the monocular avatar literature: it offers a clean way to obtain the geometric regularization of mesh-bound Gaussians without the expressivity loss of rigid binding, and it is supported by a structured set of ablations. Concrete strengths include that the paper runs official code of all baselines under unified protocols, reports training and rendering efficiency concretely, provides per-component ablations (Tabs. 4-7), measures the overhead of the walking-on-mesh module (1.7 ms), and evaluates on three datasets. The evaluation is not circular: the mesh alignment is a regularizer rather than a fitted target, and the reported metrics are on held-out views and poses. However, the quantitative support is thinner than the text suggests in several places: the headline percentage improvements are not reproducible from the tables, the ablation of the central alignment loss is at noise level, and the robustness of the core assumption (that the SMPL mesh tracks the true clothed surface) is untested in the regime where it could fail. These issues are fixable within the manuscript's scope.","major_comments":[{"comment":"The sentence \"our method can notably surpass them on challenging subject by 37% and 13% respectively\" is not reproducible from Tab. 1. For every subject and metric in that table, the relative improvement of SAGA over InstantNVR is at most approximately 29% (LPIPS on subject 394: (40.00-28.48)/40.00) and over GauHuman at most approximately 12% (LPIPS on subject 387: (38.72-34.20)/38.72), and on PSNR GauHuman actually beats SAGA on three of the six subjects (377, 387, 394). The authors should either identify the exact computation behind the 37% and 13% figures or correct the claim, since as written the headline margin is not supported by the reported numbers.","section":"6.3.1, Tab. 1"},{"comment":"The central quantitative evidence for the Gaussian-Mesh Alignment Regularization, one of the paper's main contributions, is a mean PSNR gain of 0.03 dB (31.08 vs 31.11), an SSIM gain of 0.0002, and an LPIPS gain of 0.0000 over six subjects. These margins are at or below typical run-to-run variation, and the loss weights entering L_geo (lambda_pa and lambda_na in Sec. 4.2.2) are not reported anywhere in the manuscript. Please provide per-subject results, the exact weights, and ideally multiple seeds or a statistical test; without these, the ablation cannot support the statement in Sec. 6.5.2 that the alignment loss \"effectively regularizes the detached Gaussians\" on the strength of the quantitative evidence alone.","section":"6.5.2, Tab. 5"},{"comment":"A load-bearing assumption of the two-stage scheme is that the optimized SMPL mesh remains a faithful proxy for the clothed surface during the Detached Stage, and this assumption is untested in the regime where it can fail. All ZJU-MoCap and PeopleSnapshot subjects wear tight-fitting clothing, yet Eqs. (12)-(13) pull every detached Gaussian toward its bound triangle; if a garment deviates strongly from body shape (skirts, baggy trousers, dresses), these losses would bias Gaussians toward the template rather than the true surface, and Walking-on-Mesh only re-binds within the same template's adjacent triangles. The paper should quantify the mesh-surface discrepancy (for example, silhouette error or distance of the learned canonical mesh to the reconstructed depth) and/or add a subject with loose clothing. I do not see a circularity problem here, since the mesh acts as a regularizer and the metrics are on held-out views and poses; the issue is empirical coverage of the proxy assumption, not the logic of the method. The small effect of the alignment loss in Tab. 5 in the tested regime makes the risk of a negative effect in the untested regime plausible rather than merely hypothetical.","section":"4.2.2, Eqs. (12)-(13)"},{"comment":"The stage-ablation protocols are underspecified, which limits what Tab. 4 can establish. For the \"w/o adhered stage\" row it is not stated how and where the Gaussians are initialized, nor when the Detached Stage begins (iteration 0 or iteration 3k). For the \"w/o detached stage\" row it is not stated whether training continues to 15k iterations under the adhered parameterization or stops at 3k, and since the non-rigid deformation module is active only in the Detached Stage, the ablation varies stage duration jointly with module activation. Please specify the exact training schedule and initialization used for each row.","section":"6.5.1, Tab. 4"},{"comment":"The claim of \"direct high-quality mesh extraction\" and the \"first successful attempt\" (abstract, Sec. 1, Sec. 7) is supported only by qualitative depth visualizations and TSDF-fusion examples (Figs. 16-17), with no quantitative geometry metric. Moreover, because the representation is built on a deformed SMPL template, the extracted mesh inherits the SMPL prior by construction, so the improvement over 3DGS-Avatar in Fig. 17 may largely reflect that prior rather than the proposed alignment mechanism itself. Please add a quantitative geometry comparison (for example, multi-view depth consistency, silhouette IoU, or distance to a reference obtained from the 22 held-out ZJU-MoCap views) or temper the claim accordingly.","section":"6.6"}],"minor_comments":[{"comment":"Typographical and naming issues: \"neglectable\" (Sec. 6.5.3) should be \"negligible\"; \"OutOfTrianlge\" in Algorithm 1; \"representaion\" in the Tab. 4 caption; \"SpattingAvatar\" vs \"SplattingAvatar\" should be unified (Sec. 6.2 and Fig. 8 caption); and \"Peoplesnapshot\" vs \"PeopleSnapshot\" capitalization is inconsistent.","section":"Throughout"},{"comment":"Eq. (12) indexes the sum from i=0 to N while Gaussians are indexed i=1..N elsewhere. In Eq. (17), after clamping the barycentric coordinates to [0,1], the coordinates are not renormalized to sum to 1; the normalization step should be stated explicitly.","section":"Eqs. (12), (17)"},{"comment":"The per-frame latent vector psi in Eq. (24) is described as a \"per-frame latent vector\" but it is not explained how it is obtained (a learnable embedding per training frame?) or how it behaves for novel poses; please clarify.","section":"Sec. 4.4, Eq. (24)"},{"comment":"The 42% LPIPS gain over SplattingAvatar on female-3-casual is arithmetically consistent with Tab. 3, but the explanation that PSNR/SSIM \"favor the smoothed blurry results\" is asserted without supporting evidence; either provide a quantitative demonstration (for example, a patch-level analysis) or soften the claim.","section":"Sec. 6.3.2, Tab. 3"},{"comment":"Novel pose synthesis is one of the two headline tasks but is evaluated only qualitatively (Figs. 9, 10, 12); adding a quantitative metric on the AIST++/AMASS animations would make the generalization claim testable.","section":"Sec. 6.4"},{"comment":"Key hyperparameters (epsilon in Eq. (7), the loss weights lambda_mask, lambda_LPIPS, lambda_lap, lambda_normal, and the weights of the alignment losses, as well as the Gaussian count N) are deferred to the supplementary, which is not included in this submission; please include them in the main text or make the supplement available.","section":"Sec. 5.2"},{"comment":"In Tab. 2, 3DGSAvatar achieves better LPIPS than SAGA on Olek (0.0115 vs 0.0116) and Vlad (0.0158 vs 0.0165), so the caption claim \"outperforms the comparison methods on most subjects\" should acknowledge this exception or be rephrased per metric.","section":"Tab. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the two-stage idea is worth pursuing, but the reporting must be tightened before acceptance: several numeric claims (37%/13%, the 200x speedup) do not match the reported tables, and the key alignment-loss ablation shows only a 0.03 dB gain. I would ask the authors at revision to add per-subject results, exact loss weights and ablation schedules, and a robustness experiment on looser clothing; none of these should be prohibitive. The \"first mesh extraction\" novelty claim should also be checked against concurrent work, such as the mesh-related aspects of SplattingAvatar and HAHA [59], before the final version is published."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SAGA is a credible incremental step, and if the authors fix a few things it deserves publication. The two-stage adhere-then-detach scheme with flowing-on-mesh barycentrics and the walking-on-mesh re-binding is new against GoMAvatar and SplattingAvatar, and the ablations in Tables 4–6 support the mechanism: removing either stage or the alignment loss degrades results, and the walking strategy beats the UV-based walk. The efficiency numbers are real and useful. The central idea is not circular: evaluation uses held-out views and poses.\n\nThe soft spots are real but manageable. First, the text says SAGA surpasses InstantNVR and GauHuman \"by 37% and 13%\" on challenging subjects. I can't reproduce that from Table 1. On subject 377, LPIPS is 26.03 vs 18.61 (≈28%) for InstantNVR, and 19.13 vs 18.61 (≈3%) for GauHuman. Averaging doesn't give 37/13 either. This is exactly the kind of discrepancy that erodes trust. Second, the \"first direct high-quality mesh extraction from deformable Gaussians\" is an overclaim. The paper renders depth maps and does TSDF fusion; that's a standard indirect route, not direct extraction, and it is compared only against 3DGS-Avatar, with no quantitative geometry metrics. Third, no code, no data, and the key hyperparameters and the projection derivation sit in an unavailable supplementary. That is a reproducibility gap for a methods paper.\n\nOn the stress-test worry: the concern that the Detached-stage alignment losses bias Gaussians when the SMPL proxy is wrong (e.g., loose clothing) is reasonable but unproven. The mesh is optimized in stage one, so it tracks the body at least to the limit of SMPL topology. The ablations on tight clothing show the alignment loss barely moves average PSNR (31.08→31.11), which cuts both ways: the regularizer isn't harmful there, but its benefit is mostly qualitative and the loose-garment regime is untested. I'd ask the authors to add a subject with baggy clothing or quantify mesh-to-surface distance, but I would not reject the paper on this basis.\n\nWho benefits: anyone building 3DGS avatars from monocular video. The paper is a solid within-subfield contribution with honest ablations, but it needs a corrected numerical claim, a more modest mesh-extraction statement, and at least a detailed appendix if not code. I'd send it to peer review with a major-revision recommendation.","headline":"SAGA is a credible two-stage Gaussian-mesh avatar with supportive ablations, but the 37%/13% claim doesn't survive contact with its own table, and the 'first direct mesh extraction' is really depth-fusion; worth reviewing, not ready as-is.","tokens_in":24303,"tokens_out":3986,"would_cite":true,"duration_ms":37264,"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 a two-stage adhere-then-detach mesh alignment turns 3D Gaussian Splatting into animatable human avatars that generalize to novel views and poses and support direct mesh extraction.","keywords":["3D Gaussian Splatting","monocular human reconstruction","animatable avatar","mesh-guided regularization","novel view synthesis","novel pose synthesis","mesh extraction","barycentric coordinates"],"falsifier":"Run SAGA on a monocular sequence of a person in loose clothing, then compare the rendered depth and the volumetric depth-fusion mesh against a synchronized multi-view scan; if the template mesh is far from the true surface, the alignment losses should pull the Gaussians to the wrong depth and produce visible artifacts in exactly the loose regions.","tokens_in":23255,"feed_emoji":"🕺","tokens_out":9804,"duration_ms":84414,"temperature":0.7,"pith_summary":"This paper claims that the way to keep 3D Gaussian Splatting (3DGS), a point-cloud-like rendering representation, from overfitting in monocular human capture is to let a coarse body mesh discipline the Gaussians, but only for a while. Its method, SAGA, first adheres the Gaussians to the mesh so they learn a clean surface, then detaches them in a second stage so they can still fit wrinkles, shadows, and fine details. A soft alignment loss keeps the detached Gaussians near their bound triangles, and a Walking-on-Mesh routine updates those triangles as the Gaussians drift. If the claim holds, monocular video can produce avatars with photorealistic rendering, geometry clean enough for direct mesh extraction, training in about 12 minutes, and rendering at 60+ FPS. The paper reports that this two-stage design outperforms both naive Gaussian avatars and rigidly mesh-bound avatars on novel view and pose synthesis.","feed_headline":"Two-stage Gaussian alignment sharpens animatable human avatars","feed_subtitle":"Stick splats to a body mesh first, then set them free: sharper novel views and poses in 12-minute training.","key_machinery":"The load-bearing mechanism is the bound-triangle relationship between a Gaussian and a mesh face, expressed through barycentric coordinates. A Gaussian's center is a weighted sum of the three vertices of its bound triangle, and its flat normal is aligned with the triangle normal; jointly optimizing barycentric coordinates and vertex positions lets Gaussians flow on the mesh instead of being pinned in place. In the Detached Stage, the same binding is retained as a soft regularizer: a position loss (minimum distance to the triangle) and a normal loss (absolute cosine distance) keep geometry on-mesh, while a retraction step in Stage 1 and a Walking-on-Mesh step in Stage 2 correct Gaussians that slide out of their triangle. This combination is what lets the mesh transfer its smooth, well-defined geometry and natural skinning weights to the Gaussians without sacrificing the expressivity of 3DGS.","core_discovery":"On the paper's own terms, the central discovery is a two-stage surface-aligned Gaussian representation for monocularly reconstructed, animatable humans. In the Adhered Stage, each Gaussian center is a barycentric combination of the vertices of a bound triangle of the SMPL body mesh; the Gaussian is flattened along one axis and its normal is locked to the triangle normal, while both barycentric coordinates and mesh vertices are optimized so the Gaussian can flow over the surface. In the Detached Stage, the center is freed and reparameterized directly, but the Gaussian keeps its triangle binding and is pulled back toward that triangle by a position-alignment loss (minimum distance to the triangle, not just a plane) and an orientation-alignment loss (cosine distance between the Gaussian normal and the triangle normal). A Walking-on-Mesh routine replaces the bound triangle with the nearest adjacent one when a Gaussian drifts outside, keeping the regularization honest. The paper argues this adhere-then-detach schedule enforces well-defined geometry and frame-consistent deformation, and for the first time permits high-quality mesh extraction directly from deformable Gaussians learned from monocular video.","pith_inferences":["The same adhere-then-detach schedule should transfer to other articulated subjects, such as animals or deformable objects, whenever a coarse template mesh is available, because the mechanism is template regularization rather than human-specific skinning.","A direct stress test the paper leaves open is loose or flowing clothing; if the template mesh deviates far from the true surface, the alignment losses would pull Gaussians to the wrong location, so an adaptive alignment weight based on estimated mesh-to-surface error would be a natural extension.","The triangle-binding across frames could also serve as an explicit temporal smoothness prior: counting how often Walking-on-Mesh reassigns a Gaussian and penalizing frequent reassignment would regularize non-rigid motion beyond what the paper reports.","Multi-view or depth supervision at test time could relax the dependence on the template mesh and let the same representation capture geometry on bodies that the parametric body prior fits poorly."],"forward_implications":["Animating an avatar into a new pose should stop producing the broken zippers, armpit fractures, and needle-like joint artifacts shown for unregularized Gaussians, because the mesh binding supplies smoother deformation and more natural skinning weights.","A trained avatar should be usable as a mesh, not only as a renderer: fusing rendered depth maps should give a clean surface that can be retargeted, edited, or relit.","The cost profile should stay practical for interactive use: about 12 minutes of training on a single GPU and real-time rendering at 60+ FPS at 512x512 resolution.","Warm-starting with a brief strict-adhered stage should be enough to inherit the mesh's geometry discipline, avoiding the long training runs that fully rigid mesh-bound avatars require.","Compared with a rigid fixed-on-mesh approach, the paper reports both higher rendering quality and a roughly 150x training speedup, showing that the two-stage release of constraints is what makes mesh alignment affordable."],"supporting_citations":[{"why":"Defines 3D Gaussian Splatting, the base representation whose splatting and rendering equations SAGA uses, and identifies the overfitting to view-dependent regions that motivates mesh regularization.","marker":"[20]"},{"why":"Provides the SMPL parametric body mesh used as the template surface and as the source of skinning weights for canonical-to-pose deformation.","marker":"[36]"},{"why":"Introduces surface-aligned Gaussian splatting for static scenes and motivates direct mesh extraction, the idea SAGA extends to dynamic monocular humans.","marker":"[28]"},{"why":"Presents the rigid fixed-on-mesh Gaussian avatar baseline that overconstrains Gaussians; SAGA's two-stage design is explicitly positioned against its limited flexibility.","marker":"[35]"},{"why":"Presents a mesh-embedded Gaussian avatar with UV coordinates and a UV-based walking strategy; SAGA compares its Walking-on-Mesh against this approach.","marker":"[58]"},{"why":"A monocular Gaussian avatar with non-rigid deformation and pose-dependent color; it is the main naive-Gaussian baseline that SAGA claims overfits and therefore outperforms.","marker":"[21]"},{"why":"Another naive Gaussian monocular avatar baseline; SAGA uses it to show that unregularized Gaussians oversmooth or overfit under novel views and poses.","marker":"[22]"},{"why":"Provides the ZJU-MoCap dataset and the NeuralBody baseline used for novel view synthesis evaluation.","marker":"[7]"},{"why":"Provides the HumanNeRF baseline with non-rigid deformation; SAGA compares against it as a high-quality but slow NeRF alternative.","marker":"[8]"}],"fun_headline_variants":["Adhere, then detach: sharper avatar views","Two-stage Gaussian alignment improves avatars","Stick to mesh, release for realistic avatars","Surface-aligned Gaussians enhance pose synthesis","Walking-on-mesh detach boosts avatar quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the template body mesh staying close to the real clothed surface once the first stage is done, which is untested for loose clothing and unusual body shapes.","fun_headline_variants_meta":{"raw":{"variants":["Adhere, then detach: sharper avatar views","Two-stage Gaussian alignment improves avatars","Stick to mesh, release for realistic avatars","Surface-aligned Gaussians enhance pose synthesis","Walking-on-mesh detach boosts avatar quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000812,"raw_usage":{"total_tokens":3629,"prompt_tokens":1079,"completion_tokens":2550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":2480}},"tokens_in":695,"tokens_out":2550,"duration_ms":23532,"temperature":1.0,"reasoning_tokens":2480,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:56:21.945217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SAGA on a monocular sequence of a person in loose clothing, then compare the rendered depth and the volumetric depth-fusion mesh against a synchronized multi-view scan; if the template mesh is far from the true surface, the alignment losses should pull the Gaussians to the wrong depth and produce visible artifacts in exactly the loose regions.","supporting_citations":[{"cited_title":"Splattingavatar: Realistic real-time human avatars with mesh- embedded gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Presents a mesh-embedded Gaussian avatar with UV coordinates and a UV-based walking strategy; SAGA compares its Walking-on-Mesh against this approach."},{"cited_title":"Gauhuman: Articulated gaussian splatting from monocular human videos,","cited_arxiv_id":null,"evidence_quote":"Another naive Gaussian monocular avatar baseline; SAGA uses it to show that unregularized Gaussians oversmooth or overfit under novel views and poses."},{"cited_title":"Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans,","cited_arxiv_id":null,"evidence_quote":"Provides the ZJU-MoCap dataset and the NeuralBody baseline used for novel view synthesis evaluation."},{"cited_title":"HumanNeRF: Free-viewpoint rendering of moving people from monocular video,","cited_arxiv_id":null,"evidence_quote":"Provides the HumanNeRF baseline with non-rigid deformation; SAGA compares against it as a high-quality but slow NeRF alternative."}],"review_version":1}