{"id":"20fc3577-31bb-4dbb-9208-0b36c77a4587","arxiv_id":"2506.09836","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DynaSplat is a dynamic Gaussian splatting method with static-dynamic separation, hierarchical motion decomposition, and a distance-angle opacity model; benchmark gains are small and the evaluation lacks error bars and code.","lead":"DynaSplat extends 3D Gaussian splatting to dynamic scenes by sorting Gaussians into static and dynamic sets, then modeling dynamic motion in coarse and fine layers, with an opacity formula tied to distance and viewing angle. The authors report improved benchmark scores and head avatar quality, but the gains are small, code is not released, and some table entries conflict with the abstract's claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The D-NeRF ablation rows for 'w/o Opacity' are implausibly close to the full model, which suggests the opacity component is not actually load-bearing for the headline accuracy claim; verify the ablation table before trusting the contribution claims.","rationale":"The reader's CONDITIONAL verdict is appropriate: the paper's headline claim depends on Tables I-III, which lack error bars and code, and the N3DV row already trails 4D-Rotor-Gaussians in PSNR (31.68 vs 31.80), so the 'surpasses state of the art' phrasing is already softened by the paper's own numbers. My stress-test adds one more precise internal inconsistency: the opacity ablation row on D-NeRF shows no meaningful degradation (SSIM identical, LPIPS 0.035 vs 0.034), even though Section III-D presents physically-based opacity as a core contribution and the qualitative ablation (Fig. 8) claims visible improvements to specular highlights. That flat ablation row is the kind of evidence that should make a careful referee ask whether the module is actually load-bearing. The other two ablations do show consistent drops, so I do not conclude the pipeline fails; I recommend CONDITIONAL because the authors must supply per-scene numbers, the corrected baseline rows, and sensitivity analysis for the hand-set thresholds before the accuracy claim is accepted at face value. I partially agree with the reader: they identified the dynamic-static separation thresholds as the weakest assumption, but the more concretely checkable weakness in this paper is the inconsistency between the opacity module's claimed centrality and its near-zero ablation effect on D-NeRF. The identical baseline rows across Tables I and II are also worth flagging, but they affect the comparison tables more than the internal logic of the method; the opacity ablation affects how we weight the paper's own contribution claims.","tokens_in":9309,"tokens_out":2236,"duration_ms":21204,"concrete_test":"Re-run the 'Without Physically-Based Opacity Computation' ablation on the D-NeRF dataset with the same training budget and report per-scene PSNR/SSIM/LPIPS for the full model and the no-opacity variant. If the per-scene SSIM deltas are all below 0.01 and the LPIPS delta stays below 0.005, the opacity module is not load-bearing for the accuracy claim and the paper should soften its contribution claims; if instead several scenes show large deltas that average out, the table is compressed and the concern resolves. Also re-run the sensitivity of the dynamic-static separation to tau=0.01, epsilon=1.0, gamma=0.5 by varying each over a 2x range and reporting PSNR, to confirm the headline margin is not threshold-dependent.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that DynaSplat 'surpasses state-of-the-art alternatives in accuracy and realism.' Table IV is the paper's only ablation evidence linking the three proposed components to that claim. The 'Without Physically-Based Opacity Computation' row on D-NeRF is 33.65 PSNR, 0.965 SSIM, 0.035 LPIPS, versus 34.39/0.965/0.034 for the full model. SSIM is identical and LPIPS differs by only 0.001 on a benchmark where the claimed SOTA margin over 4D-Rotor-Gaussians is only 0.14 PSNR. The opacity module is presented as a core contribution (Section III-D) and is claimed in the ablation to materially improve rendering of highlights. If dropping it changes SSIM by less than 0.001 and LPIPS by 0.001, the module is not necessary for the accuracy claim; the same table also shows much larger drops for the other two components, so the opacity row sits oddly. The paper provides no per-scene breakdown, no error bars, and no sensitivity analysis for the thresholds tau=0.01, epsilon=1.0, gamma=0.5 or the pruning threshold tau=0.02, all of which gate the dynamic-static separation. Combined with the identical DyNeRF/StreamRF/HyperReel rows in Tables I and II that mirror the source papers' numbers, the support for the headline margin is thin. The concrete check: recompute the no-opacity variant on D-NeRF and report per-scene metric deltas; if the full-model vs no-opacity gap remains below 0.01 SSIM and 0.001 LPIPS, the opacity contribution should be re-scoped to 'visual quality in qualitative examples' rather than a core accuracy contribution. Regardless of the opacity row, the reader's conditional acceptance is appropriate; the attack targets the precision of the 'surpasses state of the art' claim and the completeness of the ablation evidence, not the soundness of the pipeline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DynaSplat, a dynamic 3D Gaussian Splatting method that combines three components: a dynamic-static separation mechanism based on per-Gaussian deformation-offset variance and 2D optical-flow consistency, a hierarchical motion model that blends multiple MLP motion modes with a coarse-to-fine decomposition, and a distance- and angle-dependent opacity computation. The method is evaluated on D-NeRF, N3DV, and a head-avatar benchmark, with quantitative comparisons to prior methods and ablations. The central claim is that DynaSplat surpasses state-of-the-art alternatives in accuracy and realism for dynamic scene reconstruction while remaining efficient.","tokens_in":9671,"tokens_out":4059,"duration_ms":46737,"significance":"If supported, the work would be a useful contribution to dynamic scene reconstruction: it extends 3DGS with a principled dynamic-static separation, a hierarchical motion representation, and a view-dependent opacity model, and it demonstrates competitive quantitative results on two standard benchmarks plus a downstream avatar application. The paper's strengths include a clear architectural description, ablations for each proposed component, and results on multiple datasets. However, the significance is currently limited by evaluation issues: an abstract-level accuracy claim that is contradicted by one of the paper's own tables, duplicated baseline rows across two different datasets, and an ablation row that shows almost no effect for one claimed core component. These issues need to be addressed before the significance claim can be accepted.","major_comments":[{"comment":"The abstract claims DynaSplat 'surpasses state-of-the-art alternatives in accuracy and realism,' but in Table II on the N3DV dataset the proposed method achieves PSNR 31.68, which is lower than the 4D-Rotor-Gaussians PSNR of 31.80 (though SSIM and LPIPS are better). The claim is therefore not supported on the PSNR axis for this dataset. The text should qualify the accuracy claim by metric or remove the blanket 'surpasses' statement.","section":"Abstract and Table II"},{"comment":"Tables I and II contain identical rows for DyNeRF (29.58/-/0.080), StreamRF (28.16/0.850/0.310), HyperReel (30.36/0.920/0.170), and NeRFPlayer (30.69/-/0.110) on two different datasets (D-NeRF synthetic, N3DV real). Identical metric values on different datasets are implausible and suggest the numbers were copied without re-evaluation or proper sourcing. The authors must verify the correct numbers for each dataset, state explicitly whether baselines were retrained or numbers were taken from prior papers, and correct any errors.","section":"Tables I and II"},{"comment":"The 'physically-based opacity computation' is stated as alpha_g = alpha0 * cos(theta_g) / ||P_g - P_camera||^2 without derivation or implementation details. In particular, the paper does not explain how the Gaussian normal n_g is computed, how the opacity is normalized to valid [0,1] ranges, or why opacity should scale inversely with squared camera distance in a physically based volumetric rendering model. This is a load-bearing component of the method and the ablation table, so the authors should either provide a derivation and implementation specification or rename it as a heuristic.","section":"Section III-D, Eq. (14)"},{"comment":"The ablation for 'Without Physically-Based Opacity Computation' on the D-NeRF dataset reports SSIM 0.965 (identical to the full model) and LPIPS 0.035 (0.001 worse than the full model), while the text in Section IV-D-3 claims the opacity computation 'significantly improves' rendering. This discrepancy weakens the contribution claim. The authors should provide per-scene metric deltas and explain why the component has negligible effect on D-NeRF metrics while having a larger effect on N3DV, or re-scope the opacity contribution to specific visual effects.","section":"Section IV-D, Table IV"},{"comment":"Equations (11) and (12) are duplicates: both define G_dynamic with identical conditions, and Section III-B-c on obtaining the motion flow map is not integrated into the formal definition. This duplication is confusing and leaves unclear whether the motion flow map is part of the thresholding pipeline or only a verification step. The authors should merge the definitions and clarify the exact computational flow.","section":"Section III-B"},{"comment":"The head-avatar experiment is presented as a significant application, but no dataset name, data preprocessing, training protocol, or evaluation split is provided for the 'benchmark dataset' in Table III. Without this information, the reported improvements over INSTA, FlashAvatar, and GaussianBS cannot be independently assessed. Please specify the dataset and protocol.","section":"Section IV-C, Table III"}],"minor_comments":[{"comment":"The blending parameters beta_m in the Adaptive Motion Networks are introduced but no constraint (e.g., softmax or sum-to-one) is specified; please clarify how the mixture is normalized.","section":"Section III-C, Eq. (13)"},{"comment":"The temporal importance w_i is defined as a maximum over all views and timestamps; please clarify whether this is recomputed during training or fixed, and how the pruning threshold tau=0.02 interacts with the removed opacity reset operation.","section":"Section III-D, Eq. (15)"},{"comment":"Reference [7] for 3D Gaussian Splatting lists an incorrect author sequence ('R. Kerbl, A. Ruiz, T. Leimkühler, C. Buehler, S. Saito, A. Vedaldi, ...'); the standard citation should be Kerbl, Kopanas, Leimkühler, and Drettakis. References [8] and [22] appear to describe the same work with different publication venues; please unify.","section":"References"},{"comment":"There are several typographical issues: 'GroudTruth' in Figure 5, 'shown is Fig. 2' in Section III-D, and 'lightweight three MLP' in Section III-B should be 'three-layer MLP'. Please proofread.","section":"Throughout"},{"comment":"The related work section is very brief and does not discuss several recent dynamic Gaussian Splatting methods beyond the ones directly compared; adding a short paragraph on these would help position the contributions.","section":"Section II"}],"recommendation":"major_revision","confidential_remarks":"The duplicated rows in Tables I and II and the overclaim in the abstract are the most serious issues and should be checked carefully by the editor. The paper has a plausible core method and the ablations do show internal consistency for most components, but the evaluation reporting needs substantial correction before the claims can be trusted. I do not see evidence of deliberate fabrication, but the copied-looking baseline rows require the authors to provide raw per-scene numbers and source references."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nPunchline: this is a workable dynamic Gaussian splatting pipeline, not a breakthrough. The new bits are the combination of offset-variance and flow-consistency to split static/dynamic Gaussians, the coarse-to-fine hierarchical motion (neighbor average + fine MLP), and a view/distance-dependent opacity term. Each piece is a modest twist on known deformable-GS ideas, but the combination is new and the ablations mostly do their job. I believe the central mechanism holds up; the problem is the evaluation.\n\nThe biggest issue is the numbers. Table II (N3DV) shows DynaSplat at 31.68 PSNR against 31.80 for 4D-Rotor-Gaussians, yet the abstract says 'surpasses state-of-the-art in accuracy.' That is a direct contradiction on the headline metric. Worse, the baseline rows for DyNeRF, StreamRF, HyperReel, and NeRFPlayer are identical in both the D-NeRF and N3DV tables. That is almost certainly a copy-paste error, and it makes me distrust every baseline number until the authors provide corrected tables or code.\n\nThe stress-test note flags the D-NeRF no-opacity ablation as too close (SSIM identical, LPIPS +0.001). True, but look at the same row on N3DV: SSIM drops from 0.967 to 0.930, LPIPS from 0.069 to 0.150. So the opacity module is not dead weight. However, the ablation removes the whole Section III-D bundle, including Temporal Importance Filtering, so it doesn't isolate Eq. (14). That needs an ablation split.\n\nAlso, 'physically-based opacity' is a stretch. Eq. (14) is an inverse-square-plus-cosine heuristic with a hand-set alpha0. No derivation is provided. It may work fine, but the label overclaims. The hand-set thresholds tau, epsilon, gamma, and tau_prune have no sensitivity analysis, and there are no per-scene metrics or error bars. These are fixable in a revision.\n\nI don't see circularity or invented entities. The math is simple and coherent. The paper is likely a solid incremental contribution once the evaluation is made honest. It deserves peer review, but the verdict should be major revision, not acceptance. I wouldn't cite it until the tables are corrected, but I'd put it on the reading-group shortlist.\n\nRecommendation: send it out, but ask the reviewers to focus on baseline verification, the ablation split for opacity, and a softened accuracy claim.","headline":"Plausible dynamic-GS pipeline with a real evaluation problem: the SOTA claim is contradicted by its own N3DV PSNR table, and the baseline rows are duplicated across datasets; still worth sending to reviewers if the authors clean it up.","tokens_in":10275,"tokens_out":3893,"would_cite":false,"duration_ms":40885,"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":"DynaSplat claims dynamic scene reconstruction is best done by first splitting the scene's Gaussians into static and dynamic sets using deformation statistics and camera-subtracted motion flow, then modeling the moving set with a…","keywords":["dynamic scene reconstruction","3D Gaussian splatting","dynamic-static separation","hierarchical motion decomposition","optical flow consistency","physically-based opacity","head avatar reconstruction","novel view synthesis"],"falsifier":"On a sequence with a strongly moving camera whose pan direction parallels an independently moving object, residual camera-flow errors would land on static Gaussians: if DynaSplat flags a large share of the background as dynamic, its score should fall toward its own no-separation ablation (33.12 PSNR on D-NeRF). A second check is to sweep the separation threshold $\\tau$ from $10^{-4}$ to $10^{-1}$ on D-NeRF and see whether the reported 34.39 PSNR and the margin over the closest competitor (34.25) survive the sweep.","tokens_in":9062,"feed_emoji":"🎥","tokens_out":14069,"duration_ms":127163,"temperature":0.7,"pith_summary":"Dynamic scene reconstruction asks for an accurate, temporally coherent 3D model built from video of moving content, and DynaSplat's answer is that the standard deformable-Gaussian pipeline wastes capacity by deforming every point even when most of a scene is still. The paper's claim is that a two-track design fixes this: classify each Gaussian as static or dynamic early in training, keep a single lightweight network for the static majority, and spend the expressive motion model only on the moving minority, whose motion is further decomposed into one coarse global component and one fine per-Gaussian residual. On the D-NeRF and N3DV benchmarks the method reports the best or near-best PSNR, SSIM, and LPIPS among the compared methods, and on a head-avatar benchmark it reports higher fidelity than specialized avatar baselines while training in about 11 minutes and rendering at 324 frames per second. If the claim holds, complex dynamic scenes and high-fidelity digital avatars can be reconstructed from video at interactive speeds without giving up accuracy.","feed_headline":"Separating motion from stillness sharpens 3D video reconstruction","feed_subtitle":"DynaSplat labels each part static or moving, models motion in two layers, and beats prior baselines on D-NeRF and N3DV.","key_machinery":"The load-bearing mechanism is the classification rule that produces the partition $G = G_{\\text{dynamic}} \\cup G_{\\text{static}}$: a Gaussian is dynamic if its offset variance $\\operatorname{Var}(\\Delta\\mu_g) > \\tau$ and if, for more than a fraction $\\gamma = 0.5$ of frames, its projected position falls on a pixel whose motion-flow magnitude exceeds $\\epsilon = 1.0$. On top of that partition sits the hierarchical motion identity $\\Delta X_g = \\Delta X^{\\text{coarse}}_g + \\Delta X^{\\text{fine}}_g$, where the coarse term averages motion over neighboring Gaussians and the fine term is an MLP applied to the coarse estimate plus a learned per-point feature, with four motion modes blended by learnable weights $\\{\\beta_m\\}$. The third mechanism is physically-based opacity $\\alpha_g = \\alpha_0 \\cos\\theta_g / \\|P_g - P_{\\text{camera}}\\|_2^2$ in place of the opacity reset used in vanilla 3D Gaussian Splatting, paired with a temporal importance filter that prunes Gaussians whose contribution to rendered pixels is negligible at every view and time. Together these carry the argument: the split focuses capacity on motion, the additive decomposition captures global trends plus local detail, and the opacity law handles occlusion, depth ordering, and semi-transparent surfaces.","core_discovery":"The central claim is that dynamic-static separation is the missing ingredient in deformable Gaussian splatting, where a scene is represented as thousands of small 3D ellipsoidal Gaussians, each with a position, rotation, scale, color, and opacity, projected and blended to render each frame. DynaSplat first trains a canonical-and-deformation model for a few thousand iterations, then labels a Gaussian as dynamic only when two independent signals agree: the temporal variance of its learned position offset exceeds a threshold $\\tau$, and its projected positions keep landing on pixels where a motion flow map — optical flow with the camera-induced component removed — shows real movement. Static Gaussians retain a lightweight three-layer MLP; dynamic Gaussians are handed to a hierarchical model whose final motion is the sum of a coarse neighborhood-averaged offset and a fine MLP residual computed from that coarse estimate plus a learnable per-Gaussian feature, with several motion modes blended by learned weights. Opacity is no longer a free parameter: it is computed as $\\alpha_g = \\alpha_0 \\cos\\theta_g / \\|P_g - P_{\\text{camera}}\\|_2^2$, and pruning is done by temporal importance across training views rather than by spatial attributes. With these pieces the paper reports PSNR 34.39 on D-NeRF against 34.25 for the closest Gaussian competitor, SSIM 0.967 and LPIPS 0.069 on N3DV, and PSNR 33.24 with 324 FPS inference on the head-avatar benchmark, and the ablation tables attribute the totals to all three components.","pith_inferences":["The most robust evidence for the headline claim may be the perceptual and structural gains on N3DV (LPIPS 0.069 and SSIM 0.967 versus 0.142 and 0.935 for the closest Gaussian competitor) rather than the 0.14 dB PSNR lead on D-NeRF, which is within the usual run-to-run spread of these pipelines.","Because the separation rule is a hand-set classifier over flow and deformation statistics, the fixed thresholds are the part of the design most likely to need per-scene tuning; replacing them with a learned separator is a natural and testable extension of the paper's own thesis.","The camera-subtracted flow map is the only motion cue for the dynamic-static split, so the method's real-world robustness is untested in exactly the regimes where optical flow struggles — heavy occlusion, textureless surfaces, and reflections — and those cases will decide how general the claim is.","The opacity formula is a simple physical prior that does not depend on motion at all; applying the same distance-and-angle opacity inside static Gaussian splatting would be a cheap way to test whether that contribution transfers beyond dynamic scenes."],"forward_implications":["Dynamic scenes can be rendered at interactive rates with the accuracy of slower methods: the head-avatar experiment reports 324 FPS inference after about 11 minutes of training, with PSNR 33.24 and LPIPS 0.0281 against specialized baselines.","Model capacity is spent where motion lives: static regions keep a single lightweight MLP while confirmed moving Gaussians alone receive the multi-mode hierarchical network, so background fidelity does not consume deformation parameters.","Motion that mixes a global body transform with local deformations — head pose plus facial expression, for instance — is exactly where the coarse-plus-fine sum pays off, which the paper argues is why the method suits avatar reconstruction.","Removing the physically-based opacity from the N3DV ablation alone drops SSIM from 0.967 to 0.930 and LPIPS from 0.069 to 0.150, indicating geometry-aware opacity and temporal importance pruning carry much of the perceptual gain on real-world video.","The D-NeRF table places the method ahead of non-Gaussian dynamic-scene methods too: the best non-Gaussian scores are K-Planes at 31.05 and HyperReel at 30.36, versus the proposed 34.39."],"supporting_citations":[{"why":"Supplies the 3D Gaussian representation, rendering equations, and optimization schedule that DynaSplat extends to dynamic scenes.","marker":"[7]"},{"why":"The D-NeRF synthetic dynamic-scene dataset on which the headline PSNR 34.39 and the Table I comparisons are measured.","marker":"[13]"},{"why":"The N3DV multi-view video dataset used for Table II, and the DyNeRF method both as dataset source and as a baseline.","marker":"[25]"},{"why":"Followed to build the 2D motion flow map: off-the-shelf optical flow with the camera-induced component removed, on which the dynamic-static consistency check depends.","marker":"[9]"},{"why":"The optical flow estimator that produces the motion flow map used to verify candidate dynamic Gaussians.","marker":"[21]"},{"why":"Defines the canonical-plus-deformation architecture (DeformNet) that the early training stage and static track inherit, and appears as the Def-3D-Gauss baseline.","marker":"[19]"},{"why":"The closest competitor on both benchmark tables (D-NeRF PSNR 34.25); the claim of surpassing current methods is judged chiefly against it.","marker":"[30]"},{"why":"The strongest head-avatar baseline on PSNR in Table III (31.07 versus the proposed 33.24), used to demonstrate the avatar application.","marker":"[34]"},{"why":"The avatar baseline with the best reported error before the proposed method (PSNR 32.52), the main comparison for the application claim.","marker":"[35]"}],"fun_headline_variants":["Splitting static and moving Gaussians sharpens 3D video","Two-stage motion modeling lifts Gaussian splatting realism","DynaSplat: static-dynamic split yields crisper 3D scenes","Masking static parts improves dynamic 3D reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The two-track model is only as good as the dynamic-static split, and that split inherits every error in the optical flow and the camera-motion subtraction, with hand-set thresholds ($\\tau = 0.01$, $\\epsilon = 1.0$, $\\gamma = 0.5$) for which no sensitivity analysis is reported.","fun_headline_variants_meta":{"raw":{"variants":["Splitting static and moving Gaussians sharpens 3D video","Two-stage motion modeling lifts Gaussian splatting realism","DynaSplat: static-dynamic split yields crisper 3D scenes","Masking static parts improves dynamic 3D reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000378,"raw_usage":{"total_tokens":2063,"prompt_tokens":1052,"completion_tokens":1011,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":938}},"tokens_in":668,"tokens_out":1011,"duration_ms":10522,"temperature":1.0,"reasoning_tokens":938,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:39:45.388672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a sequence with a strongly moving camera whose pan direction parallels an independently moving object, residual camera-flow errors would land on static Gaussians: if DynaSplat flags a large share of the background as dynamic, its score should fall toward its own no-separation ablation (33.12 PSNR on D-NeRF). A second check is to sweep the separation threshold $\\tau$ from $10^{-4}$ to $10^{-1}$ on D-NeRF and see whether the reported 34.39 PSNR and the margin over the closest competitor (34.25) survive the sweep.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian representation, rendering equations, and optimization schedule that DynaSplat extends to dynamic scenes."},{"cited_title":"D-nerf: Neural radiance fields for dynamic scenes,","cited_arxiv_id":null,"evidence_quote":"The D-NeRF synthetic dynamic-scene dataset on which the headline PSNR 34.39 and the Table I comparisons are measured."},{"cited_title":"Motiongs: Exploring explicit motion guidance for deformable 3d gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Followed to build the 2D motion flow map: off-the-shelf optical flow with the camera-induced component removed, on which the dynamic-static consistency check depends."},{"cited_title":"Memflow: Optical flow estimation and prediction with memory,","cited_arxiv_id":null,"evidence_quote":"The optical flow estimator that produces the motion flow map used to verify candidate dynamic Gaussians."},{"cited_title":"4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes,","cited_arxiv_id":null,"evidence_quote":"The closest competitor on both benchmark tables (D-NeRF PSNR 34.25); the claim of surpassing current methods is judged chiefly against it."},{"cited_title":"Flashavatar: High-fidelity head avatar with efficient gaussian embedding,","cited_arxiv_id":null,"evidence_quote":"The strongest head-avatar baseline on PSNR in Table III (31.07 versus the proposed 33.24), used to demonstrate the avatar application."},{"cited_title":"3d gaussian blendshapes for head avatar animation,","cited_arxiv_id":null,"evidence_quote":"The avatar baseline with the best reported error before the proposed method (PSNR 32.52), the main comparison for the application claim."}],"review_version":1}