{"id":"1ba1b07b-2569-4e9e-8610-94b3b197aa95","arxiv_id":"2501.07104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RMAvatar couples mesh-embedded Gaussian splats with a pose-conditioned rectification MLP to reconstruct animatable clothed avatars from monocular video.","lead":"RMAvatar builds a 3D human avatar from a single monocular video by attaching Gaussian splats to an SMPL mesh and learning pose-dependent corrections for clothing and hair. It reports higher rendering quality than four prior methods on two public benchmarks, though several closely related systems are not compared.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baselines may not use the same Anim-NeRF-refined SMPL poses or ZJU-MoCap split, so the reported dB gains could be pose-input artifacts rather than representation quality.","rationale":"The reader's weakest assumption was pose accuracy and the lack of quantitative pose sensitivity analysis. I see that as a real issue, but the more load-bearing concern is that the headline comparison itself is not shown to be controlled: the paper commits to using Anim-NeRF-refined poses only for its own method, and does not state that the baseline numbers were produced under the same pose inputs, train/test split, and evaluation protocol. Because RMAvatar's performance depends strongly on mesh pose quality, uncontrolled pose inputs can confound the claimed improvements. The method itself appears internally coherent, the ablations are informative, and the limitations are stated, so I do not think the central contribution is invalid. The appropriate response is to require a controlled re-evaluation and code/pose release before the SOTA claim can be accepted; this matches the reader's CONDITIONAL verdict, so I leave the verdict unchanged.","tokens_in":15427,"tokens_out":6165,"duration_ms":62203,"concrete_test":"Rerun SplattingAvatar and GaussianAvatar under the authors' exact protocol: Anim-NeRF-refined SMPL poses fitted only on the training frames, the Table 3 train/test frames, and the same evaluation camera; also rerun them with the unrefined SMPL poses used in their original papers. If the PSNR/SSIM/LPIPS advantage over the retrained baselines drops below about 0.1 dB or reverses, the SOTA claim is unsupported. Additionally, evaluate a held-out ZJU-MoCap camera at training-time frames to verify true novel-view performance. If code is not released, this test requires the authors to supply runnable baselines and pose files for independent reproduction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec 4.3) is state-of-the-art novel-view synthesis on PeopleSnapshot and ZJU-MoCap, supported only by Tables 1 and 2. Section 4.1 says 'we use the provided poses optimized by Anim-NeRF [4] in our training and inference' and 'We use Anim-NeRF [40] to get the refined poses of ZJU-MoCap in our training and inference,' but it never states that the compared methods were retrained with these same refined poses, nor whether the ZJU-MoCap pose refinement includes test frames or is restricted to the training split. Since RMAvatar is a mesh-guided Gaussian method, its positions are heavily determined by SMPL pose accuracy, and the paper itself notes SplattingAvatar's quality 'is highly dependent on the motion accuracy of the underlying mesh' (Sec 4.3). If the baselines' published numbers were produced with raw SMPL fits or different pose optimization, part or all of the reported 0.3-1.9 dB improvements may come from better pose inputs rather than the Gaussian rectification module. Additionally, the ZJU-MoCap 'test' split in Table 3 is temporal (later frames of the same camera), so the experiment tests novel poses, not genuinely novel views; no held-out camera is used. Thus the quantitative SOTA claim is not yet established under a controlled comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RMAvatar, a hybrid representation for animatable human avatars from monocular video. 3D Gaussian splats are embedded on the triangular faces of an SMPL mesh; the mesh drives their rigid motion via LBS, and a pose-conditioned 5-layer MLP predicts offsets to position, rotation, and scale to capture non-rigid deformations such as cloth and hair. Training uses RGB/SSIM/LPIPS losses plus position, scaling, and offset regularizers. Experiments on PeopleSnapshot and ZJU-MoCap report PSNR/SSIM/LPIPS improvements over Anim-NeRF, InstantAvatar, GaussianAvatar, and SplattingAvatar, and a qualitative animation comparison.","tokens_in":108,"tokens_out":8597,"duration_ms":125676,"significance":"If the results are taken at face value, the representation is a sensible marriage of explicit mesh guidance (robust under novel poses) and Gaussian splatting appearance, with the rectification MLP addressing the known LBS limitation for non-rigid deformations. The method is described in sufficient detail to be re-implemented, and the ablation isolates the contribution of the rectification module. The paper does not ship code or machine-checked proofs; its main deliverable is the empirical claim. That claim is currently weakened by (i) an uncontrolled comparison protocol with respect to SMPL pose refinement, (ii) a ZJU-MoCap evaluation that tests novel poses rather than held-out views, and (iii) missing comparisons against closely related Gaussian-on-mesh methods. With a controlled evaluation, the method could be a solid contribution.","major_comments":[{"comment":"The comparison is not controlled for SMPL pose inputs. Section 4.1 states 'For fair comparison, we use the provided poses optimized by Anim-NeRF [4] in our training and inference' (PeopleSnapshot) and 'We use Anim-NeRF [40] to get the refined poses of ZJU-MoCap in our training and inference,' but it never states whether Anim-NeRF, InstantAvatar, GaussianAvatar, or SplattingAvatar were retrained with these same refined poses or used the same ZJU frame selection. Since the paper itself notes in §4.3 that SplattingAvatar's quality 'is highly dependent on the motion accuracy of the underlying mesh,' the 0.3–3.0 dB advantages reported in Tables 1–2 may reflect better pose inputs rather than the rectification module. Please re-run all baselines under identical pose refinement and data splits, or provide evidence that the published baseline numbers were produced under the same protocol, and include a pose-sensitivity study (e.g., RMAvatar with raw vs. refined SMPL).","section":"§4.1, Tables 1–2"},{"comment":"The ZJU-MoCap evaluation does not demonstrate novel-view synthesis. Table 3 shows training on frames 1–456 and testing on frames 457–end for each subject, which is temporal extrapolation at the same camera viewpoint, not a held-out view. The claim in §4.3 of 'novel view synthesis' is therefore overstated. Either add a held-out camera evaluation (train on one camera, test on another) or explicitly re-label the experiment as novel-pose synthesis and adjust the conclusion accordingly.","section":"Table 3, §4.1–4.3"},{"comment":"Three closely related Gaussian-avatar methods discussed in the related work—3DGS-Avatar [45], GoMAvatar [55], and GART [26]—are absent from the quantitative comparison. GoMAvatar in particular is a Gaussians-on-mesh method with the same design philosophy. Without these baselines (or a clear justification for their exclusion, such as incompatible pose estimation requirements), the claim of far exceeding the state of the art is not established. Please add these methods under the same protocol or narrow the claim to the compared set.","section":"§2.2–2.3, Tables 1–2"},{"comment":"The ablation support for the OffsetLoss and ScalingLoss components is weak. On female-3-casual, the full model achieves PSNR 32.51 versus 32.48 without OffsetLoss and 32.45 without ScalingLoss, with identical SSIM and LPIPS in both cases; no variance or multiple seeds are reported. This does not support the statement in §4.4 that the offset and scaling losses 'significantly improve' reconstruction quality. Please report mean and standard deviation over multiple runs or over additional subjects, and use a test that accounts for run-to-run variance.","section":"Table 4, §4.4"}],"minor_comments":[{"comment":"The sentence 'Note that L_rgb, L_ssim and L_lpips are L1-norm losses' is incorrect: SSIM and LPIPS losses are not L1 norms. Please replace it with proper definitions (e.g., 1−SSIM and the LPIPS distance).","section":"Eq. (12), §3.4"},{"comment":"The claim 'Our method far exceeds the SoTA methods' is too strong for gains of 0.3–0.6 dB on several ZJU subjects (e.g., subject 386: 30.61 vs 30.31 PSNR). Substantiate with statistical analysis or temper the language.","section":"§4.3"},{"comment":"Please specify how out-of-distribution poses were generated (e.g., novel SMPL poses from other sequences or retargeting) and provide quantitative metrics for animation frames if possible.","section":"Figure 4, §4.3"},{"comment":"For ZJU-MoCap, specify which camera viewpoint was used per subject and confirm that the same camera and frame split were used for all compared methods.","section":"§4.1"},{"comment":"The sentence attributing SplattingAvatar's artifacts to 'Walking on mesh strategy' is speculative; either provide a citation or evidence, or soften it.","section":"§4.3"},{"comment":"The position regularization L_pos = ||max(μ−ε_pos,0)||2 constrains only positive deviations from the parent triangle; if the intent is to constrain both directions, the absolute value should be used inside the max. Please clarify.","section":"Eq. (13), §3.4"},{"comment":"There are numerous typos and spacing artifacts (e.g., 'threshhold' near Eq. (14), 'recitification' in §3.3, 'farexceeds' in §4.3). A careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about pose refinement is valid and should be the primary request in the revision. I do not see internal inconsistency in the method itself; the problem is external validity of the comparison. If the authors can rerun baselines with identical pose inputs and clearly describe the ZJU protocol, the paper may be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Reading this paper, I side with the conditional verdict. The core idea is a natural hybrid: embed 3D Gaussians on an SMPL mesh, then add a pose-conditioned MLP that predicts small offsets to Gaussian position, rotation, and scale to capture non-rigid cloth and hair deformation. Neither ingredient is new on its own, but the combination, with the offset regularization to keep corrections small and local, is new and clearly motivated. The writing is readable, and the internal evaluation is honest: the 'w/o GauRec' ablation gains about 0.6 dB PSNR on PeopleSnapshot, which supports the rectification module's utility, and the position/scaling regularizers plausibly target the artifacts shown.\n\nThe soft spots are real but addressable. The most serious is pose input control. The paper states it uses Anim-NeRF refined poses for its own training and inference, but it never says whether SplattingAvatar and GaussianAvatar numbers were produced with those same refined poses. For a mesh-guided representation, the Gaussian positions are heavily determined by the SMPL fit, so some fraction of the reported 1-2 dB advantage could be pose accuracy rather than the rectification module. That is a genuine confound, and the stress-test note lands. The SOTA claim is also thinner than it looks: 3DGS-Avatar, GoMAvatar, and GART are cited but not compared, there are no error bars, and the ZJU-MoCap 'test' split is later frames from the same camera—novel pose, not truly novel view. The animation results are qualitative only. None of this kills the method; it does mean the 'state-of-the-art' claim outruns the evidence.\n\nThis paper is for people working on monocular avatar reconstruction with Gaussian splatting. It is a legitimate incremental contribution, timely and coherent, with limitations honestly acknowledged in the conclusion. It deserves a serious referee. My recommendation: send it to peer review and require a controlled comparison—report the pose inputs for every baseline or retrain them under identical conditions, add error bars, and soften the SOTA claim to 'best among compared methods under our protocol.'","headline":"A sensible hybrid avatar method whose internal ablations support the rectification module, but whose state-of-the-art claim rests on uncontrolled pose inputs and a too-thin baseline set.","tokens_in":16264,"tokens_out":3379,"would_cite":true,"duration_ms":32493,"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":"RMAvatar reconstructs photorealistic animatable avatars from monocular video by binding Gaussian splats to a SMPL mesh and rectifying them with a pose-conditioned network, reaching state-of-the-art rendering and animation quality on…","keywords":["human avatar reconstruction","3D Gaussian splatting","mesh-embedded Gaussians","monocular video","non-rigid deformation","SMPL","novel view synthesis","animatable avatar"],"falsifier":"Render the same ZJU-MoCap subjects from held-out viewpoints using ground-truth SMPL fits from the multi-view capture instead of Anim-NeRF-refined poses; if RMAvatar's margin over SplattingAvatar shrinks, its gain is mostly correcting pose error rather than modeling authentic non-rigid deformation.","tokens_in":15239,"feed_emoji":"🧍","tokens_out":7932,"duration_ms":64804,"temperature":0.7,"pith_summary":"RMAvatar tries to reconstruct a photorealistic, animatable clothed human avatar from a single monocular video. The paper argues that a hybrid representation—3D Gaussian splats bound to the faces of a posed SMPL mesh—captures gross motion and shape, while a pose-conditioned rectification network adds the fine non-rigid detail (cloth wrinkles, hair, hand motion) that linear blend skinning misses. If the claims hold, the method would give higher novel-view synthesis and animation quality than prior NeRF- and Gaussian-based avatar methods on the PeopleSnapshot and ZJU-MoCap benchmarks, and do so with the speed of Gaussian rasterization. The paper reports state-of-the-art PSNR, SSIM, and LPIPS scores on both datasets, with ablation studies showing the rectification module is the main contributor.","feed_headline":"RMAvatar beats prior art on avatar rendering quality from one video","feed_subtitle":"A pose-conditioned rectifier on mesh-bound Gaussians captures cloth and hair detail from a single video.","key_machinery":"The central representation is the mesh-embedded Gaussian: each 3D Gaussian is attached to a parent triangle of the posed SMPL mesh, so its local position, rotation, and scale are transformed to world space by the triangle's frame, then refined by a Gaussian rectification module—a five-layer MLP mapping the encoded Gaussian position and the current pose to small offsets $(\\delta\\mu, \\delta r, \\delta s)$. The mesh guides low-frequency motion; the rectifier adds pose-dependent non-rigid detail; three regularization terms keep Gaussians near their parent triangles, prevent overly large scaling, and keep the rectifier's offsets small.","core_discovery":"On its own terms, the paper's central claim is that binding Gaussian splats to mesh triangles and correcting them with a pose-dependent MLP yields the best reported novel-view and animation quality for monocular human avatar reconstruction. The mesh supplies a compact, topology-complete scaffold that places Gaussians correctly under LBS-driven motion, and the rectification module predicts per-Gaussian offsets in position, rotation, and scale as a function of pose and Gaussian location, compensating for the non-rigid deformations LBS cannot express. The paper states that this decomposition of motion into rigid mesh guidance plus learned non-rigid rectification surpasses both pure-Gaussian methods like GaussianAvatar and the mesh-embedded SplattingAvatar baseline on PSNR, SSIM, and LPIPS, and produces cleaner geometry for loose clothing and hair.","pith_inferences":["An unstated corollary is that the rectifier may be compensating for pose-estimation error as much as for true non-rigid deformation; if so, its benefit should shrink when trained with ground-truth poses, which is directly testable on multi-view datasets.","The same binding-plus-rectification pattern could transfer to other parametric templates such as FLAME for heads, where expression-dependent wrinkles require non-rigid correction beyond mesh motion.","The offset regularization caps how far Gaussians can leave their parent triangles, so very loose garments or large cloth folds that require big displacements may be under-modelled; testing on a video with a long skirt or wide sleeves would probe this boundary.","Because the rectifier is a single MLP over pose and position, its capacity is fixed; training on longer sequences with more pose variety should push generalization further, but the paper does not quantify this scaling behavior."],"forward_implications":["A direct corollary is that the same hybrid scaffold can be dropped into real-time rendering pipelines, since rasterization uses 3D Gaussian splatting rather than volumetric ray marching.","If the pose-conditioned rectifier generalizes, avatars animated to out-of-distribution poses should stay coherent on loose clothing and hair, where pure LBS mesh warping fails.","The method's sensitivity to pose accuracy is reduced relative to mesh-only baselines because the rectifier can absorb small pose errors, which matters for monocular capture where fitted SMPL is noisy.","The ablation results imply that the rectification module, not densification, is the main source of the quality gain, so future avatar methods need not grow Gaussian counts to improve fidelity."],"supporting_citations":[{"why":"Supplies the differentiable Gaussian splatting rasterizer and the Gaussian primitive representation on which the avatar is built.","marker":"[21]"},{"why":"Provides the SMPL template mesh and LBS deformation that anchors Gaussians to body motion.","marker":"[32]"},{"why":"The mesh-embedded Gaussian baseline whose hybrid representation RMAvatar extends and which it outperforms in the quantitative comparisons.","marker":"[48]"},{"why":"Contributes the binding-inheritance idea and the position regularization (Eq. 13) used to keep Gaussians near their parent triangles.","marker":"[44]"},{"why":"A pure-Gaussian baseline that motivates the need for mesh guidance and is outperformed in the experiments.","marker":"[13]"},{"why":"Supplies the refined SMPL poses used for training and inference on both datasets, a load-bearing input for the pose-conditioned rectifier.","marker":"[4]"},{"why":"One of the two evaluation datasets; the state-of-the-art claim on novel views rests on this benchmark.","marker":"[2]"},{"why":"The other evaluation dataset, providing monocular videos with large motions and pose errors that stress the rectification module.","marker":"[41]"}],"fun_headline_variants":["RMAvatar: mesh-bound splats with pose rectification win","Rectified mesh-Gaussians beat prior for single-video avatars","RMAvatar: lifelike human avatars from a single video","Pose-corrected Gaussians on mesh: new avatar rendering bar"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the SMPL poses used for training and animation are accurate enough for the mesh to provide reliable guidance, and that the pose-conditioned rectifier can learn non-rigid deformation without overfitting to pose-specific errors or failing on out-of-distribution poses.","fun_headline_variants_meta":{"raw":{"variants":["RMAvatar: mesh-bound splats with pose rectification win","Rectified mesh-Gaussians beat prior for single-video avatars","RMAvatar: lifelike human avatars from a single video","Pose-corrected Gaussians on mesh: new avatar rendering bar"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1493,"prompt_tokens":885,"completion_tokens":608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":529}},"tokens_in":501,"tokens_out":608,"duration_ms":5937,"temperature":1.0,"reasoning_tokens":529,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:17.316319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the same ZJU-MoCap subjects from held-out viewpoints using ground-truth SMPL fits from the multi-view capture instead of Anim-NeRF-refined poses; if RMAvatar's margin over SplattingAvatar shrinks, its gain is mostly correcting pose error rather than modeling authentic non-rigid deformation.","supporting_citations":[{"cited_title":"3d gaussiansplattingforreal-timeradiancefieldrendering","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable Gaussian splatting rasterizer and the Gaussian primitive representation on which the avatar is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SMPL template mesh and LBS deformation that anchors Gaussians to body motion."},{"cited_title":"Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting, in: Proceedings of theIEEE/CVFConferenceonComputerVisionandPatternRecogni- tion, pp","cited_arxiv_id":null,"evidence_quote":"The mesh-embedded Gaussian baseline whose hybrid representation RMAvatar extends and which it outperforms in the quantitative comparisons."},{"cited_title":"20299–20309","cited_arxiv_id":null,"evidence_quote":"Contributes the binding-inheritance idea and the position regularization (Eq. 13) used to keep Gaussians near their parent triangles."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A pure-Gaussian baseline that motivates the need for mesh guidance and is outperformed in the experiments."},{"cited_title":"Video based reconstruction of 3d people models, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"One of the two evaluation datasets; the state-of-the-art claim on novel views rests on this benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The other evaluation dataset, providing monocular videos with large motions and pose errors that stress the rectification module."}],"review_version":1}