{"id":"2ddf238e-a31a-4c3d-9ea4-2a7b49af9770","arxiv_id":"2411.16768","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D Gaussian avatar model that conditions non-rigid deformation on hierarchical skeleton and vertex motion reaches state-of-the-art rendering quality on three human-capture datasets.","lead":"SeqAvatar uses motion history from both the whole skeleton and individual surface points to better animate 3D Gaussian human avatars. It renders high-quality novel views of moving people in real time, beating existing 3DGS-based avatars on several benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fine vertex motion V is deterministically derived from SMPL pose and LBS weights, so it may add no independent motion information; the claimed gains over pose-only conditioning need a control test.","rationale":"Reading the paper in good faith, the strongest claim is that hierarchical motion context, especially fine vertex motion, makes SeqAvatar outperform 3DGS baselines and match or exceed NeRF temporal models at much higher speed. The speed and baseline comparisons are plausible, but the causal story for quality is the fine motion condition. My concern is that fV is not an independent observation of garment motion: it is computed by rigidly skinning the SMPL template (Eq. 10) and taking finite differences, so every velocity value is a function of the same pose sequence that already enters fΔP. The method could still learn useful point-specific motion features from these velocities, since KNN sampling makes fV spatially varying, so I do not claim the mechanism is vacuous a priori. What is missing is a control test that separates genuine local motion cues from extra conditioning capacity. The paper's own Section 6 limitation makes this concern internal, not an outside-consensus dispute. I did not select the lack of code or error bars as load-bearing because those affect reproducibility and confidence, not the validity of the central mechanism. The permutation ablation is cheap, uses the paper's own pipeline, and would either confirm the V-conditioning story or show that the gains in Table 4 (d) are capacity-driven. This matches the reader's weakest assumption, so the conditional verdict should stand pending the test.","tokens_in":18291,"tokens_out":4056,"duration_ms":42548,"concrete_test":"On the I3D-Human setup, retrain the Table 4 (d) configuration with a permutation control: keep fΔP unchanged, but construct the vertex-velocity template F_V by randomly permuting velocities across SMPL vertices (or across time) before Eq. 12. If PSNR/LPIPS remain statistically indistinguishable from Table 4 (d) (32.01/31.23), V contributes no independent local-motion signal; if performance drops to Table 4 (b)/(c) levels, the fine condition carries real information. A secondary check would compare V_t with dense garment tracking on one DNA-Rendering sequence, but the permutation ablation alone settles the redundancy question.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The advertised fine-grained condition fV (Eqs. 11-13) is computed as finite differences of LBS-warped SMPL template vertices, V_t = (To_t - To_{t-s})/s, with To_t = LBS(T, B_t, W) (Eqs. 10-11). For fixed body shape, this field is a deterministic function of the pose history {P_{t-s}, ..., P_t} and the fixed skinning weights; it contains no observed garment or cloth deformation. The coarse skeleton motion fΔP already encodes the same pose differences (Eqs. 7-8), so the unique content of V is a spatially distributed re-encoding of pose history, not a new measurement of local surface motion. The improvement in Table 4 (d-c: PSNR 31.89 -> 32.01, LPIPS 32.17 -> 31.23) could therefore reflect extra capacity rather than the proposed fine-motion mechanism. Section 6 concedes that local velocity cues derive from the coarse SMPL model rather than dense surface tracking, but no experiment checks whether V carries information beyond ΔP. Since the abstract's 'fine-grained vertex motions' is the central basis for claiming superiority over pose-only 3DGS avatars, this redundancy would directly undermine the main claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SeqAvatar, a 3D Gaussian Splatting (3DGS) framework for animatable human avatars from multi-view video. The method augments the standard pose-conditioned non-rigid deformation network with two additional motion conditions: a coarse skeleton-motion embedding ΔP computed as differences of SMPL joint rotations over a temporal window (Eqs. 7-8), and a fine vertex-motion embedding f_V obtained by finite differences of LBS-warped SMPL template vertex positions, aggregated to each Gaussian via KNN (Eqs. 10-13). A spatio-temporal multi-scale sampling strategy (STMS, Eq. 14) varies the time interval to combine long-term trends with inter-frame details. Experiments on DNA-Rendering, I3D-Human, and ZJU-MoCap report improved PSNR/SSIM/LPIPS over 3DGS-Avatar, GART, GauHuman, and the NeRF-based Dyco, at roughly 62 FPS versus Dyco's 0.7 FPS.","tokens_in":18515,"tokens_out":12640,"duration_ms":106509,"significance":"The work has practical value: it is a simple and modular motion-conditioning scheme that demonstrably improves 3DGS avatar rendering quality on loose-clothing, complex-motion benchmarks (Tables 1-2) while preserving real-time rendering (Table 8). Credit is due for the multi-dataset evaluation, the component-wise ablations (Tables 4, 6, 7), the efficiency analysis, and the candid limitation statement in Section 6. The conceptual novelty is modest: the fine motion condition is a spatially distributed re-encoding of pose history rather than an observation of true surface motion, and the incremental gains over pose-only conditioning are small (on the order of 0.1-0.3 dB). The significance of the central claim therefore rests on whether the reported gains are attributable to the motion semantics of the new conditions rather than to added model capacity; the control experiments requested below will determine this.","major_comments":[{"comment":"The stress-test concern about the fine vertex-motion condition lands. V is a deterministic function of the SMPL pose history, the fixed template T, and the fixed skinning weights W: V_t is computed by LBS-warping the template (Eqs. 10-11), and the KNN sampling in Eq. (12) uses canonical Gaussian positions only, so V contains no measured garment or surface deformation beyond what the pose history implies. The observed gain from adding V (Table 4, (c)->(d): PSNR 31.89->32.01, LPIPS 32.17->31.23) is small and could be produced by the added parameters of E_knn and E_V rather than by the semantic content of the velocities, a possibility the paper itself leaves open in Section 6 ('our local velocity cues are derived from the coarse SMPL model rather than dense surface tracking'). Because the abstract credits 'fine-grained vertex motions' as the basis for superiority over pose-only 3DGS avatars, the authors should add a control that keeps capacity but destroys the spatial semantics of V, e.g., randomly permuting vertex velocities across the template before KNN sampling, or replacing f_V with a per-Gaussian learned latent of the same dimension (R^96). If the improvement in row (d) persists under either control, the proposed fine-motion mechanism is not supported.","section":"§4.1, Eqs. (10)-(13); Table 4"},{"comment":"All quantitative results are single-run and no variance estimates are reported. This matters because several differences that carry the paper's claims are small: adding V in Table 4 changes PSNR by +0.12 dB, and on ZJU-MoCap (Table 5) SeqAvatar trails GauHuman on PSNR (31.02 vs 31.04) and SSIM (0.9619 vs 0.9620). Without standard deviations over multiple seeds, these margins cannot be distinguished from run-to-run noise, and the ablative support for the central contribution is correspondingly weak. Please report mean and standard deviation over at least three runs for the headline comparisons and for the ablation rows in Table 4.","section":"Tables 1-5"},{"comment":"The protocol for extending the monocular baselines 3DGS-Avatar, GART, and GauHuman to multi-view input is not described; the text only states that they are extended 'under the same settings.' The fairness of the comparison depends on details such as training views, iterations, adaptive densification settings, loss weights, and whether pose refinement is enabled. Without this information, the consistent gains in Tables 1-3 cannot be fully attributed to the proposed method rather than to differences in baseline tuning. Please document the adaptation protocol in the supplement.","section":"§5.2"},{"comment":"The abstract's unconditional statement that SeqAvatar 'significantly outperforms 3DGS-based approaches' is stronger than the ZJU-MoCap evidence supports: in Table 5 the method does not exceed GauHuman on PSNR (31.02 vs 31.04) or SSIM (0.9619 vs 0.9620), and is better only on LPIPS (28.89 vs 31.81). The advantage appears to be dataset-dependent, being pronounced on the loose-clothing, complex-motion DNA-Rendering and I3D-Human sets but at parity on the controlled ZJU-MoCap set. Please calibrate the claim to this scope, e.g., by stating that the method significantly outperforms 3DGS baselines on complex-motion datasets and is comparable on controlled settings.","section":"Abstract; Table 5"}],"minor_comments":[{"comment":"Loss weights are denoted λ1, λ2, λ3 in Eq. (25) but reported as λ0, λ1, λ2 in Section 11; please unify the notation and state which weight corresponds to which loss term for each dataset.","section":"Eq. (25); §11"},{"comment":"The figures use a calligraphic symbol for vertex velocity while the text uses V and F_V, and the delta notation in the caption of Fig. 1 is not defined; please harmonize the notation between figures and equations.","section":"Figs. 1-2"},{"comment":"The baseline set is inconsistent across tables: GART appears only in Table 1, and Dyco appears only on I3D-Human and ZJU-MoCap; please state whether the same baselines were evaluated on all datasets and give the reason if not, e.g., availability of pretrained models or data splits.","section":"§5.2, Tables 1-3"},{"comment":"The KNN neighborhoods are computed from canonical Gaussian positions x_i, which are optimized and may drift from the SMPL template vertices; please state whether the neighborhoods are recomputed during optimization or fixed once at initialization, since this affects the meaning of the 'local region' being sampled.","section":"§4.2, Eq. (12)"},{"comment":"There is a typo in the supplementary material: '3DGS-Avaar' should be '3DGS-Avatar'.","section":"§12.1 (Supp.)"},{"comment":"The caption reads 'Mehods' and the table does not state the dataset, resolution, and hardware details beyond 'a single 4090 GPU'; please specify these details so the 62 FPS figure is reproducible.","section":"Table 8"}],"recommendation":"major_revision","confidential_remarks":"The paper shares two co-authors with the main NeRF-based comparator, Dyco [8]; the comparison itself appears honest and quantitative, but the editor may wish to be aware of the overlap. The contribution is incremental engineering rather than a new paradigm: the core idea is conditioning a non-rigid MLP on temporal pose differences and per-vertex SMPL velocities. The principal risk to the central claim is the absence of a capacity-control experiment for the vertex-motion condition, which the authors can address within the scope of a revision; the small margins in Tables 4 and 5 without error bars compound this risk. Regarding fit to the journal, the manuscript is written in a conference format and the experimental section would benefit from the requested statistical and control analyses before journal publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid, incremental 3DGS avatar paper. The core idea is to augment the pose-residual condition (borrowed from Dyco) with per-Gaussian velocities computed from LBS-warped SMPL template vertices, and to sample these conditions at multiple temporal scales. The experiments are thorough: three datasets, many baselines, ablations, and even a computational cost table. The authors also state their main limitation plainly in Section 6.\n\nWhat is actually new: the vertex-velocity field, sampled with KNN from the SMPL template, is a simple way to give each Gaussian a spatially varying motion vector rather than a single global skeleton feature. The multi-scale temporal sampling is a modest but reasonable addition. The paper doesn't open a new capability, but it is a useful refinement for the 3DGS avatar pipeline.\n\nThe main soft spot is the one the stress-test flags: V is a deterministic function of pose history and fixed skinning weights, so it carries no observed garment deformation. It does, however, provide local directional information that the global ΔP lacks, so it is not strictly redundant. The real question is whether the improvement is due to the specific velocity content or just to extra MLP capacity. The ablation gain from adding V is small (about 0.12 dB PSNR and 0.9 LPIPS on I3D-Human), and there is no control experiment with a same-dimensional but motion-agnostic condition. That's a fixable but important omission.\n\nOther issues are conventional: no code release, no error bars, hyperparameters selected on the benchmarks used for final results, and the abstract's claim of 'significantly outperforms 3DGS-based approaches' is uneven on ZJU-MoCap, where the method is on par with GauHuman. The baseline extension to multi-view is not described in detail. Also, the pose-residual condition comes from Dyco, and two authors overlap; that's not a flaw by itself, but it makes the multi-view adaptation of monocular baselines more consequential, and the paper glosses over it.\n\nAll that said, this is honestly written and careful. The central mechanism is plausible, the experiments are above the field average, and the limitations are acknowledged. It deserves a serious referee; with a capacity-matched control and code release I would be comfortable seeing it accepted.\n\nRecommended: send to peer review. For a reading group, it's a fine example of a modern 3DGS avatar paper, but not an exciting one.","headline":"Solid incremental 3DGS avatar work; the vertex-velocity condition is plausible, but the paper never shows it beats a same-capacity pose-only control.","tokens_in":19101,"tokens_out":3450,"would_cite":true,"duration_ms":32433,"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":"Conditioning Gaussian avatar deformation on both skeleton residuals and per-vertex velocities captures loose-clothing motion at real-time speed.","keywords":["3D Gaussian Splatting","animatable human avatar","non-rigid deformation","hierarchical motion context","SMPL template velocities","multi-scale temporal sampling","novel view synthesis","pose sequence conditioning"],"falsifier":"Train the same pipeline twice on a video of a garment whose motion is dominated by inertia or external forces, such as a flaring skirt during a spin or a coat blowing in wind: once with SMPL-derived vertex velocities and once with vertex velocities taken from dense surface tracking of the real garment. If the second does not outperform the first, or if the first shows no gain over a pose-only baseline, the claim that fine vertex motion is the load-bearing improvement fails.","tokens_in":18048,"feed_emoji":"👤","tokens_out":7174,"duration_ms":65478,"temperature":0.7,"pith_summary":"This paper argues that pose-only conditioning is the bottleneck for animatable Gaussian avatars: a single frame's pose cannot represent the many appearances the same pose can take during complex motion, especially in loose clothing. SeqAvatar adds a hierarchical motion context with two levels: coarse skeleton motion, obtained from differences of SMPL poses across frames, and fine-grained point-wise motion, obtained from finite-difference velocities of SMPL template vertices warped by linear blend skinning. Both conditions are sampled at several temporal scales and over neighbouring template vertices, then fed to the MLP that predicts each Gaussian's non-rigid deformation. The claim is that this restores real-time rendering while matching or surpassing slower NeRF-based temporal models and beating pose-only 3DGS baselines on DNA-Rendering, I3D-Human, and ZJU-MoCap.","feed_headline":"Vertex-velocity cues push 3DGS avatars past pose-only baselines","feed_subtitle":"Skeleton deltas plus per-vertex velocities capture loose clothing at interactive rates, matching slower NeRF models.","key_machinery":"The load-bearing object is the vertex motion template field $F_{\\mathcal{V}}$: a per-SMPL-vertex velocity array computed by warping the template to observation space with linear blend skinning and taking finite differences $V_t = (T^o_t - T^o_{t-s})/s$. Each canonical Gaussian looks up the $\tau$ nearest template vertices' velocities, encodes them with an MLP, and concatenates the result with the coarse skeleton motion embedding obtained from pose residuals $\\Delta P_t = \\delta(P_t, P_{t-s})$. Spatio-temporal multi-scale sampling varies the interval $s$ across a set $S = \\{s_0, s_0+\\Delta s, \\dots\\}$, giving the deformation MLP both global body movement and local region motion over several temporal windows. This is what lets one network resolve the same pose at different moments into different non-rigid deformations.","core_discovery":"The central claim is that an explicit Gaussian representation can carry motion conditions at two granularities, and that the fine granularity is what recovers appearance details far from the skeleton. The coarse condition is a sequence of per-joint rotation differences between adjacent frames, encoded into a 32-dimensional embedding. The fine condition is a motion template field that stores, for every SMPL vertex, its finite-difference velocity under linear blend skinning; each Gaussian samples its $\tau$ nearest template vertices' velocities and encodes them into a 96-dimensional embedding. A spatio-temporal multi-scale sampling strategy builds these embeddings from several time intervals so that long-term motion trends and inter-frame details enter together. The non-rigid MLP then predicts position, scale, and rotation offsets for each Gaussian from its position, pose, and the two motion embeddings, and the deformed Gaussians are warped by LBS and splatted at real-time rates.","pith_inferences":["Replacing the SMPL-derived velocity proxy with dense surface tracking or physics-based simulation is the clear next experiment; if it further sharpens garments, the coarse template is the bottleneck rather than the conditioning principle.","The same velocity-conditioned deformation could apply to any explicit representation with a template and skinning weights, including animals or clothed characters whose pose-to-appearance map is similarly ambiguous.","One could drive avatars from video by feeding observed vertex velocities directly into the deformation MLP, allowing re-enactment and animation to share the same machinery without first regressing pose.","Because the vertex motion template is a cheap pre-computation, the conditioning could be extended to higher-resolution templates or denser velocity fields without affecting the real-time inference cost."],"forward_implications":["Adding the coarse skeleton-motion condition alone improves over pose-only non-rigid deformation, and adding the fine vertex-motion condition improves further, which supports the hierarchy as the cause of the gains.","Multi-scale sampling over several time intervals gives better robustness than any single interval, so temporal context should be gathered at multiple scales rather than at one fixed step.","The method keeps real-time rendering while matching or beating the slower NeRF-style temporal baseline, so temporal conditioning does not have to sacrifice interactivity.","Novel-pose and out-of-distribution animations trained on one sequence render plausible results on poses from unseen sequences, indicating that the motion conditions generalize rather than merely memorize training poses."],"supporting_citations":[{"why":"Supplies the pose-sequence temporal baseline and the I3D-Human dataset splits; the coarse skeleton-motion condition extends its pose-residual idea.","marker":"[8]"},{"why":"The explicit 3D Gaussian representation that makes real-time splatting and per-primitive velocity conditions possible.","marker":"[27]"},{"why":"Pose-conditioned non-rigid deformation baseline whose LBS weight-offset MLP, mask loss, and evaluation protocol the method adopts and extends.","marker":"[22]"},{"why":"The pose-conditioned 3DGS avatar baseline it must beat; also provides ZJU-MoCap splits and the as-isometric-as-possible losses.","marker":"[56]"},{"why":"SMPL template vertices and pose parameters from which the vertex motion template field is constructed.","marker":"[43]"},{"why":"SMPL-X template used for DNA-Rendering experiments, providing the 10475-vertex Gaussian initialization.","marker":"[51]"},{"why":"DNA-Rendering dataset with loose clothing and complex motion used for the main multi-view novel-view comparisons.","marker":"[9]"},{"why":"ZJU-MoCap benchmark used to test generalization to simpler controlled motions.","marker":"[53]"}],"fun_headline_variants":["SeqAvatar: skeleton deltas plus vertex velocities for realistic 3DGS avatars","Fine-grained vertex motion unlocks loose clothing for real-time 3DGS avatars","Hierarchical motion context: coarse joints plus fine vertices speed up avatars","Velocity-aware 3DGS matches NeRF quality at interactive rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's fine-grained motion signal comes from the coarse SMPL body model, not from tracking the actual clothing, so everything rests on the assumption that SMPL vertex motion faithfully represents how loose garments and local surfaces really move.","fun_headline_variants_meta":{"raw":{"variants":["SeqAvatar: skeleton deltas plus vertex velocities for realistic 3DGS avatars","Fine-grained vertex motion unlocks loose clothing for real-time 3DGS avatars","Hierarchical motion context: coarse joints plus fine vertices speed up avatars","Velocity-aware 3DGS matches NeRF quality at interactive rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000531,"raw_usage":{"total_tokens":2546,"prompt_tokens":924,"completion_tokens":1622,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":1539}},"tokens_in":540,"tokens_out":1622,"duration_ms":10114,"temperature":1.0,"reasoning_tokens":1539,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:33:38.922980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline twice on a video of a garment whose motion is dominated by inertia or external forces, such as a flaring skirt during a spin or a coat blowing in wind: once with SMPL-derived vertex velocities and once with vertex velocities taken from dense surface tracking of the real garment. If the second does not outperform the first, or if the first shows no gain over a pose-only baseline, the claim that fine vertex motion is the load-bearing improvement fails.","supporting_citations":[{"cited_title":"Within the Dynamic Context: Inertia-aware 3D Human Modeling with Pose Sequence","cited_arxiv_id":"2403.19160","evidence_quote":"Supplies the pose-sequence temporal baseline and the I3D-Human dataset splits; the coarse skeleton-motion condition extends its pose-residual idea."},{"cited_title":"Gauhuman: Articu- lated gaussian splatting from monocular human videos","cited_arxiv_id":null,"evidence_quote":"Pose-conditioned non-rigid deformation baseline whose LBS weight-offset MLP, mask loss, and evaluation protocol the method adopts and extends."},{"cited_title":"3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting","cited_arxiv_id":null,"evidence_quote":"The pose-conditioned 3DGS avatar baseline it must beat; also provides ZJU-MoCap splits and the as-isometric-as-possible losses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SMPL-X template used for DNA-Rendering experiments, providing the 10475-vertex Gaussian initialization."}],"review_version":1}