{"id":"4a1a0e87-06be-4807-a8c8-373d52b04947","arxiv_id":"2507.18758","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A graph over SMPL mesh vertices aggregates per-frame 3D Gaussians from a video, yielding a reposable human avatar in a single feed-forward pass.","lead":"Researchers built a neural network that turns a short video of a person into a 3D avatar viewable from new angles and poseable in new motions, without needing per-person optimization. It connects frame-by-frame 3D Gaussian blobs to a standard body mesh so information from all frames is pooled.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The graph construction assumes LGM-predicted Gaussians and posed SMPL vertices share a common Euclidean frame; no alignment step is described, so Eq. 6's nearest-vertex assignment is undefined across the Wonder3D canonical space and the SMPL coordinate system.","rationale":"The reader's weakest_assumption correctly targets the SMPL-anchoring step, but the paper's own limitation statement and ablations show this is exactly where the method lives or dies. My concern sharpens this: beyond SMPL accuracy, the paper never specifies in what coordinate frame the LGM-predicted Gaussians are expressed when they are compared with posed SMPL vertices. In the single-view branch, Wonder3D's generated multi-views define a canonical object space; the SMPL parameters estimated from the original image are in the image camera's coordinate. Without an explicit registration, the Euclidean distances in Eq. 6 are not necessarily physical distances. This is not an internal inconsistency that falsifies the results, but it is an unforced gap between the text and the claimed construction: either the code contains an undocumented alignment (making the paper incomplete on a core step) or the method relies on an implicit assumption that is likely false for in-the-wild monocular inputs. The proposed test directly settles whether the graph is actually using metric proximity. The verdict remains CONDITIONAL because the empirical support and ablations are otherwise coherent and the claimed gains are plausible; the condition is that the authors release the code and either document the alignment or demonstrate insensitivity to coordinate perturbations.","tokens_in":13735,"tokens_out":6528,"duration_ms":70121,"concrete_test":"Check whether the released code contains any explicit coordinate registration between LGM-predicted Gaussian centers and posed SMPL vertices before Eq. 6. Then run the single-view pipeline twice: once with the SMPL mesh as-is, and once with a rigid transform (e.g., 5 cm translation, 5° rotation) applied to the SMPL vertices before graph construction, keeping everything else fixed. If the PSNR on the 10-scan evaluation changes by more than ~0.5 dB, the Euclidean assignment in Eq. 6 is sensitive to coordinate alignment and the missing alignment description is a real gap; if it is unchanged, the distance metric is not doing the work claimed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The method's graph (Section 3.2, Algorithm 1) builds Gaussian-mesh edges by computing Euclidean distances between per-frame Gaussian centers µ_t^m and posed SMPL vertices V_t^i in Eq. 6. This is only meaningful if both sets of 3D points live in the same coordinate frame. For the multi-view setting, MvHumanNet provides calibrated cameras and SMPL in a common world coordinate, so this may hold. For the single-view setting, however, the pipeline first runs a fine-tuned Wonder3D to synthesize multi-view images (Section 4.1, Supplementary §6), and LGM predicts Gaussians in the canonical coordinate system induced by those generated views. The SMPL pose p_t is estimated from the original monocular image, which is not obviously registered to that canonical system. The paper never describes a canonicalization or root-alignment step between the LGM output and the SMPL mesh before the nearest-vertex assignment. If the frames are misaligned by even a few centimeters, the hard assignment in Eq. 6 groups unrelated content, and the claimed temporal aggregation in the intra-node operation (Eqs. 8-9) mixes Gaussians from different body parts. The paper's own ablation attributes 2.1 dB of the novel-pose gain to the intra-node operation, so this assumption is load-bearing, not a minor implementation detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Human Gaussian Graph (HGG), a graph-based module that links per-frame Gaussian primitives predicted by a frozen feed-forward model (LGM) to shared SMPL mesh vertices. Via intra-node attention across frames and inter-node attention across mesh neighbors, the module produces SMPL-aligned Gaussians that can be re-posed for novel pose animation. The method operates in a single forward pass, supporting both multi-view and monocular video input. Experiments on MvHumanNet report state-of-the-art results in novel view synthesis and novel pose animation, with ablations showing contributions of each component.","tokens_in":14014,"tokens_out":4637,"duration_ms":44095,"significance":"If the results hold, the method provides a practical route to generalizable animatable avatars without per-instance optimization, which is an important step for real-time applications. The dual-layer graph formulation is a clean way to aggregate temporal information while maintaining mesh alignment, and the ablation study carefully isolates each module's contribution. The efficiency analysis in the supplementary is a useful contribution. The major caveats are the unaddressed coordinate-frame alignment in the monocular pipeline and the lack of statistical rigor in the evaluation, both of which need to be resolved before the quantitative claims are fully convincing.","major_comments":[{"comment":"The graph construction as written assumes that the Gaussian centers µ_t^m and the posed SMPL vertices v_t^i lie in the same Euclidean coordinate frame. In the multi-view setting this is plausible because MvHumanNet provides calibrated cameras and SMPL in a common world frame. In the single-view setting, however, the pipeline first synthesizes multi-view images with a fine-tuned Wonder3D and then runs LGM, whose Gaussian predictions live in the canonical coordinate frame induced by those generated views, while the SMPL pose p_t is estimated from the original monocular image. The paper never describes a canonicalization or root-alignment step between these two coordinate systems before Eq. (6) is applied. If the two frames are misaligned, the nearest-vertex assignment groups unrelated content and the intra-node operation (Eqs. 8–9) mixes Gaussians from different body parts; the ablation attributes 2.1 dB of the novel-pose gain to the intra-node operation, so this is load-bearing. Please specify the alignment procedure or add an explicit coordinate normalization step.","section":"Section 3.2, Eq. (6)"},{"comment":"All quantitative comparisons are reported on a single set of 10 held-out MvHumanNet scans, with no error bars, per-scan variance, or cross-seed statistics. Given the small evaluation set, the claimed improvements (e.g., 1.6 dB over ExAvatar for single-view novel view, 2.3 dB for novel pose) may not be statistically reliable. Please report per-scan results or error bars, and state the number of random seeds used for the evaluation split selection.","section":"Section 4.2, Table 1"}],"minor_comments":[{"comment":"There are several typos: 'fisrt' should be 'first' and 'closet' should be 'closest' in the description of Eq. (6).","section":"Section 3.2"},{"comment":"B_n(V) is defined as the 'neighbor Gaussian group' but actually contains mesh vertices; rename it to 'neighbor vertex group' to avoid confusion.","section":"Section 3.3, Eq. (14)"},{"comment":"The threshold d0 for mesh-level edges is not specified in the implementation details; please provide its value or a sensitivity study.","section":"Section 3.2, Eq. (7)"},{"comment":"The input line lists 'SMPL poses {p_t}_{i=1}^{N}' but the index should run t=1,...,T; please correct the notation.","section":"Algorithm 1"},{"comment":"IDOL and AniGS are mentioned without citations or clear definitions; please add references or a brief description.","section":"Supplementary Section 7.3"},{"comment":"The 'Performance comparison' panel is referenced but the axes and units are not described in the caption; please clarify what is plotted.","section":"Figure 1(b)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript includes a GitHub link but no code or checkpoints are provided, which limits reproducibility of the reported numbers. The evaluation set of 10 scans is small; I encourage the authors to release the evaluation split details already provided in the supplementary, and to report variance or per-scan results. The claim of being 'the first work to achieve animatable Gaussian avatars within inference time' is strong and should be carefully scoped with respect to concurrent work (e.g., IDOL is mentioned only in the supplementary)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental contribution to feed-forward animatable human avatars. The dual-layer Gaussian-mesh graph is a sensible way to pool information across frames, and the ablations show each piece pulls its weight. But the evidence is thinner than the claims: 10 held-out scans, no error bars, no released artifacts, and a coordinate-frame ambiguity in the single-view pipeline that the paper never explicitly resolves.\n\nWhat's genuinely new: the cross-frame temporal aggregation anchored to SMPL vertices. Previous feed-forward methods (LGM, HumanSplat, GPS-Gaussian) either process frames independently or don't produce pose-driven output. The graph construction and the intra/inter-node attention are straightforward but effective, and the complexity reduction from quadratic to linear in frame count is a nice practical point. The authors also do fair comparisons, fine-tuning LGM on their split and letting optimization baselines run long.\n\nThe soft spots are real but not disqualifying. The evaluation set is very small, and without per-scan numbers, error bars, or cross-seed statistics, the reported 1.6 dB over ExAvatar and 2.1 dB over the baseline could be within noise. The GitHub link appears, but no code or checkpoints are described in the paper, so none of this is independently checkable. The 'first animatable feed-forward avatar' claim is overreaching: the supplementary mentions AniGS without a citation, and that's a claim that should be hedged or removed. Finally, the stress-test note about coordinate frames is worth taking seriously. For single-view input, the Gaussians come from Wonder3D's canonical space while the SMPL pose comes from a monocular estimator; the paper never describes how the two are aligned before the nearest-vertex assignment in Eq. 6. This isn't fatal for the multi-view setting, where calibration provides a common frame, but it's load-bearing for the single-view numbers and for the intra-node ablation that gains 2.1 dB. The authors should either explain the implicit canonicalization or add an alignment module.\n\nOverall, the core idea is sound and the paper deserves serious peer review. I'd engage with it and ask for code, error bars, and a precise statement about coordinate alignment before acceptance.","headline":"Solid incremental architecture for feed-forward animatable avatars, but the evidence is thin and the single-view alignment is under-specified.","tokens_in":14566,"tokens_out":3957,"would_cite":true,"duration_ms":37911,"reading_group":"yes","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 dual-layer graph linking per-frame 3D Gaussians to SMPL mesh vertices yields a generalizable, animatable human avatar from a video in a single feed-forward pass, the first such inference-time animatable Gaussian…","keywords":["animatable human avatar","3D Gaussian splatting","feed-forward reconstruction","temporal aggregation","SMPL","graph attention","novel pose animation","generalizable Gaussian model"],"falsifier":"Render a synthetic person with known ground-truth SMPL from motion capture, run the method with the true SMPL pose and with progressively perturbed SMPL estimates, and compare novel-pose PSNR; if perturbing the mesh by a few centimeters does not degrade output more than the per-frame baseline, the nearest-vertex edge construction is not doing the claimed work. Alternatively, test on a subject in loose flowing clothing: if temporal aggregation visibly merges cloth motion with body-surface Gaussians, the assignment assumption fails.","tokens_in":13528,"feed_emoji":"👤","tokens_out":9513,"duration_ms":90003,"temperature":0.7,"pith_summary":"Traditional animatable avatars require minutes to hours of per-person optimization, while recent feed-forward Gaussian predictors handle one frame at a time and cannot be reposed. The paper claims that a video's frames contain useful complementary evidence: attach every predicted 3D Gaussian to the nearest vertex of a shared human body mesh, aggregate information across frames at each vertex, and let neighboring vertices exchange messages, and the output is an SMPL-aligned avatar that can be driven into new poses in a single inference pass. On held-out multi-view scans the reported quality is higher than both per-frame feed-forward baselines and per-instance optimization methods, with new-view rendering above 120 frames per second. This would make poseable avatars from ordinary video practical for interactive applications.","feed_headline":"One graph turns a human video into a poseable avatar in seconds","feed_subtitle":"Per-frame 3D Gaussians wire to a shared body mesh, so new poses render in one pass with no per-person training.","key_machinery":"The central object is the Human Gaussian Graph (HGG), a dual-layer graph where the first-layer nodes are the 3D Gaussians predicted for every frame of the video and the second-layer nodes are the vertices of an SMPL mesh, a parametric human body model. Edges are defined in two ways: each Gaussian is connected to the SMPL vertex closest to it in the posed frame, and each mesh vertex is connected to its topological neighbors on the mesh. Information flows through two transformer operations: an intra-node operation in which a learnable query at each vertex attends to all Gaussians attached to that vertex across timesteps, and an inter-node operation in which queries at neighboring vertices exchange messages; the operations are stacked to widen the receptive field. The updated vertex queries are then used to refine the Gaussians of a chosen frame, yielding SMPL-aligned Gaussians that are reposed via linear blend skinning and rendered with Gaussian rasterization. This graph is what carries the argument: it converts temporal grouping into spatial grouping, so that cross-frame evidence is aggregated without expensive all-pairs attention between Gaussians.","core_discovery":"On the paper's own terms, the discovery is that temporal information in a video can be folded into a poseable avatar without any optimization by using a Human Gaussian Graph with two layers: Gaussian primitives from all frames and SMPL mesh vertices fixed across the timeline. Each Gaussian is attached to its nearest posed SMPL vertex; the graph then runs a cross-attention intra-node operation that lets a per-vertex query collect all Gaussians assigned to that vertex over the video, and an inter-node operation that passes messages between neighboring mesh vertices. The resulting vertex queries, applied to one frame's Gaussians, upgrade them into SMPL-aligned Gaussians that can be driven by new poses through linear blend skinning and rendered by Gaussian splatting. The paper reports that this yields the first animatable Gaussian avatars produced within inference time, surpassing per-frame feed-forward methods and optimization-based baselines on novel view synthesis and novel pose animation, with inference at roughly 0.9 seconds per frame for multi-view input and 9.7 seconds for monocular input.","pith_inferences":["The two-layer graph construction is not specific to humans; any deformable category with a parametric template or rigged mesh could adopt the same cross-frame aggregation, so the idea may transfer to hands, faces, or animals.","A likely improvement path is to replace hard nearest-vertex assignment with a learned soft assignment or additional per-vertex features, which would let the graph handle hair and loose clothing instead of leaving those regions to rely on nearby body vertices.","The bottleneck for monocular input is the quality of the generated pseudo-views; if a publicly available human-specific multi-view diffusion model emerges, the monocular and multi-view numbers should converge."],"forward_implications":["From any short video of a person, a poseable avatar can be produced in a single forward pass, without per-person optimization, at roughly 0.9 seconds per frame for multi-view and 9.7 seconds for monocular input.","Because the avatar is anchored to a body mesh, it can be reposed by any pose parameter and rendered at over 120 FPS, enabling real-time interactive use.","The graph is a separate module on top of the single-frame Gaussian predictor; replacing that predictor with a stronger one improves results, so the method compounds with progress in per-frame reconstruction.","Novel-pose quality is limited by the accuracy of the estimated body pose and shape; correcting those estimates should narrow the reported novel-view/novel-pose gap."],"supporting_citations":[{"why":"Defines the SMPL parametric body model whose vertices form the second graph layer and drive novel poses through linear blend skinning.","marker":"[24]"},{"why":"Frozen single-frame Gaussian predictor that produces the initial per-frame Gaussian sets the graph refines.","marker":"[43]"},{"why":"Multi-view diffusion model fine-tuned to generate the pseudo-views that let the monocular pipeline use the same graph.","marker":"[23]"},{"why":"Optimization-based Gaussian avatar baseline that binds Gaussians to SMPL, setting the comparison for inference-time animatable avatars.","marker":"[20]"},{"why":"Closest feed-forward human Gaussian baseline; the paper contrasts its per-frame prediction and non-animatable output with the graph approach.","marker":"[29]"},{"why":"Per-frame generalizable human Gaussian method used as a multi-view baseline that the graph must improve upon.","marker":"[55]"},{"why":"Dataset of multi-view human videos with SMPL estimates that provides the training and evaluation scans for the reported numbers.","marker":"[49]"},{"why":"3D Gaussian splatting rendering procedure used to rasterize the SMPL-aligned Gaussians into images.","marker":"[18]"}],"fun_headline_variants":["Two-layer graph makes video avatars pose-ready in seconds","Human Gaussian Graph: video to animatable avatar without training","One forward pass: video to poseable avatar via Gaussian graph","Temporal Gaussians bound to SMPL mesh for fast avatar animation","Graph-wired Gaussians give instant poseable avatars from video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes each predicted 3D Gaussian sits close enough to the estimated surface of a parametric human body mesh that attaching it to the nearest mesh vertex groups genuinely related content across frames; if the estimated body shape or pose is wrong, or the Gaussian belongs to loose clothing or hair, the edges misassign and temporal aggregation mixes unrelated evidence.","fun_headline_variants_meta":{"raw":{"variants":["Two-layer graph makes video avatars pose-ready in seconds","Human Gaussian Graph: video to animatable avatar without training","One forward pass: video to poseable avatar via Gaussian graph","Temporal Gaussians bound to SMPL mesh for fast avatar animation","Graph-wired Gaussians give instant poseable avatars from video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000667,"raw_usage":{"total_tokens":3031,"prompt_tokens":922,"completion_tokens":2109,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":2023}},"tokens_in":538,"tokens_out":2109,"duration_ms":14978,"temperature":1.0,"reasoning_tokens":2023,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:08:48.709169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic person with known ground-truth SMPL from motion capture, run the method with the true SMPL pose and with progressively perturbed SMPL estimates, and compare novel-pose PSNR; if perturbing the mesh by a few centimeters does not degrade output more than the per-frame baseline, the nearest-vertex edge construction is not doing the claimed work. Alternatively, test on a subject in loose flowing clothing: if temporal aggregation visibly merges cloth motion with body-surface Gaussians, the assignment assumption fails.","supporting_citations":[{"cited_title":"Smpl: A skinned multi- person linear model.ACM transactions on graphics (TOG),","cited_arxiv_id":null,"evidence_quote":"Defines the SMPL parametric body model whose vertices form the second graph layer and drive novel poses through linear blend skinning."},{"cited_title":"Lgm: Large multi-view gaussian model for high-resolution 3d content creation.ECCV, 2024","cited_arxiv_id":null,"evidence_quote":"Frozen single-frame Gaussian predictor that produces the initial per-frame Gaussian sets the graph refines."},{"cited_title":"Gart: Gaussian articulated template mod- els","cited_arxiv_id":null,"evidence_quote":"Optimization-based Gaussian avatar baseline that binds Gaussians to SMPL, setting the comparison for inference-time animatable avatars."},{"cited_title":"Gps- gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis","cited_arxiv_id":null,"evidence_quote":"Per-frame generalizable human Gaussian method used as a multi-view baseline that the graph must improve upon."},{"cited_title":"Mvhumannet: A large- scale dataset of multi-view daily dressing human captures","cited_arxiv_id":null,"evidence_quote":"Dataset of multi-view human videos with SMPL estimates that provides the training and evaluation scans for the reported numbers."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023","cited_arxiv_id":null,"evidence_quote":"3D Gaussian splatting rendering procedure used to rasterize the SMPL-aligned Gaussians into images."}],"review_version":2}