{"id":"7b8360fc-08b5-4f3d-b9f4-350dbeb8b0a0","arxiv_id":"2504.12909","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Gaussian-splat avatars reach 166 fps with finer details by replacing a large CNN with spatially distributed pose-only MLPs whose interpolated coefficients drive per-Gaussian offset bases.","lead":"This paper presents a way to build and render 3D human avatars from multi-view video using many small neural networks placed on the body instead of one big network. The result is sharper pose-dependent detail at 166 frames per second, roughly 16 times faster than the leading prior method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's quality comparison is computed on the training frames themselves, so the claimed appearance advantage is not established on held-out data.","rationale":"The reader's weakest assumption was the dependence on the AnimatableGaussians template. While that is a plausible robustness concern, the most load-bearing issue is that the primary quantitative evidence for the quality claim is methodologically contaminated. The supplementary explicitly states that Table 1 and the ablations are evaluated on the first 500 frames of the training set. This means the reported PSNR/SSIM/LPIPS/FID improvements over AnimatableGaussians in Table 1 partly reflect overfitting rather than generalization. The only clean comparison, Table 2 on novel poses, shows a much smaller margin (PSNR +0.216 dB, LPIPS -0.0015). Without error bars or multiple sequences, that margin is not sufficient to establish the claimed 'better appearance quality with finer details.' The speed claim (166 fps vs 10 fps) is not affected by this issue and appears credible, but the central claim combines quality and speed. Since the quality half is currently unverified, a judgment on the manuscript's central claim cannot be made until a clean held-out evaluation is provided. This moves the verdict from CONDITIONAL to UNVERDICTED. The concern is concrete and fixable, so REJECT is not appropriate, but the current evidence is insufficient to conditionally accept the quality claim.","tokens_in":15029,"tokens_out":13311,"duration_ms":132747,"concrete_test":"Using the avatarrex_zzr sequence, retrain or reuse the released checkpoints and evaluate Ours and AnimatableGaussians on a held-out camera view not used in training (e.g., a 16th view) or on a disjoint temporal segment (e.g., frames 2001-2500 if available). Compute PSNR/SSIM/LPIPS/FID; if the PSNR margin over AnimatableGaussians drops to roughly 0.2 dB (as in Table 2) or the LPIPS gap reverses, the headline appearance-quality claim is unsupported.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The supplementary material, Section 2 (Experiment Details), states: 'We use the first 2000 frames for training and calculate metrics on the first 500 frames from \"22010710\" camera view.' Since the first 500 frames are a subset of the first 2000 training frames, Table 1 and the ablations in Tables 4-6 evaluate on training images. This is train-set leakage: PSNR/SSIM/LPIPS can be inflated by overfitting, and FID computed on training images is not a valid distribution distance. The held-out evaluation in Table 2 (novel poses) avoids this flaw, and there the margin over AnimatableGaussians shrinks to +0.216 dB PSNR and -0.0015 LPIPS. Thus the large quality gap in Table 1 (+1.45 dB PSNR) is not trustworthy evidence for the central claim of 'better appearance quality with finer details.' The same leakage affects the ablation tables, so the design choices are validated only on training fit. The central claim therefore rests on a single clean evaluation with a small margin, which is not robust support.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a real-time Gaussian human avatar representation in which a set of spatially distributed MLPs, located at anchor points on a canonical template mesh, take only the pose vector as input and output coefficients. Each Gaussian’s property offsets are obtained by combining these interpolated coefficients with per-Gaussian offset basis vectors, and the Gaussian position offsets are generated by interpolation from control points so that Gaussians remain on a surface layer. The method is evaluated on AvatarReX, THuman4.0, and ActorsHQ sequences against 3DGS-Avatar, MeshAvatar, AnimatableGaussians, and DEGAS, reporting better appearance quality and a large rendering speedup (166 fps versus 10 fps).","tokens_in":15263,"tokens_out":4477,"duration_ms":44345,"significance":"If the quantitative claims hold, the paper makes a meaningful contribution to real-time high-fidelity human avatar rendering: the speed-quality trade-off relative to AnimatableGaussians is notable, the idea of interpolating coefficients rather than high-frequency properties is well motivated and supported by ablations, and the control-point surface constraint is a practical solution to a real artifact. The authors ship code, provide a supplementary viewer, and include ablations on PCA components and Gaussian counts, which are strengths. However, the central quality claim is currently supported mainly by metrics computed on training frames, and the single held-out evaluation is narrow, so the headline result is not yet established to the standard expected for a high-impact venue.","major_comments":[{"comment":"The quantitative comparison in Table 1 and the ablations in Tables 4-6 are evaluated on the first 500 frames of the first 2000 training frames, as stated in Supplementary Section 2. This is train-set leakage: PSNR/SSIM/LPIPS/FID computed on training images can be inflated by overfitting, and FID on training frames is not a valid distribution distance. The only held-out evidence is Table 2, where the margin over AnimatableGaussians shrinks to +0.216 dB PSNR and -0.0015 LPIPS. Please rerun the main comparison and the ablations on held-out frames (e.g., the last 500 frames) and report both fitting quality and held-out quality separately; the central claim of better appearance quality requires held-out results.","section":"Sec. 4.1, Table 1; Supplementary Sec. 2"},{"comment":"The novel-pose evaluation uses a single sequence (subject00) and a single camera view (\"cam18\") with no error bars or multiple random seeds. Given that the margin over AnimatableGaussians on this held-out test is small, the current evidence is too narrow to support the general claim of \"better appearance quality with finer details\" under novel poses. Please add multiple sequences and multiple evaluation views, and report per-sequence results with variance.","section":"Sec. 4.1, Table 2"},{"comment":"The PCA projection of novel poses into the space of training poses is a test-time operation that moves test inputs toward the training distribution. The paper should state explicitly that the reported novel-pose results are obtained after this projection and discuss the implications for the generalization claim. It should also confirm that all comparison baselines are given exactly the same PCA treatment; if they are not, the comparison is not apples-to-apples, and if they are, the term \"novel pose\" should be qualified.","section":"Sec. 3.3, Testing"}],"minor_comments":[{"comment":"The first paragraph of the introduction contains a duplicated word: \"learning the the human appearance\" should read \"learning the human appearance.\"","section":"Sec. 1"},{"comment":"The phrase \"position-ware MLP\" should be \"position-aware MLP.\"","section":"Sec. 3.2"},{"comment":"The main paper states that each sequence contains 1000 to 2000 frames, but Supplementary Section 2 says the avatarrex_zzr sequence uses the first 2000 frames for training. If the actual frame counts are shorter, this is inconsistent; please clarify the exact training and evaluation splits for every sequence.","section":"Sec. 4; Supplementary Sec. 2"},{"comment":"The relationship between the appearance coefficients and the position-offset coefficients is not defined in the main text; the supplementary clarifies that the MLP outputs 2B coefficients, half used for Gaussians and half for control points. This detail should be moved to the main text for clarity.","section":"Sec. 3.2"},{"comment":"The pose vector dimensionality is described as excluding finger joints, but Supplementary Code 1 uses a 63-dimensional input. Please state the exact dimensionality of the pose vector used in the main text.","section":"Sec. 3.3; Supplementary Code 1"}],"recommendation":"major_revision","confidential_remarks":"The train-set evaluation is a serious flaw that undermines the headline quality claim, but it is fixable and the method itself is promising. The PCA test-time projection and the narrow single-view, single-sequence evaluation also need attention. I recommend major revision; after rerunning the evaluation on held-out frames and broadening the test suite, the paper could be a strong contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper and the supplement, and the stress-test note is right: the headline quality numbers in Table 1 are computed on training frames. The supplement says the metrics use the first 500 frames from one camera view, and the first 2000 frames are used for training. So Tables 1, 4, 5, and 6 are in-sample. That matters because the main claim of 'better appearance quality with finer details' is supported in Table 1 by a 1.45 dB PSNR margin over AnimatableGaussians. The only clean held-out evaluation is Table 2 (novel poses), where the margin shrinks to +0.216 dB PSNR and -0.0015 LPIPS, with FID essentially tied. So the quality advantage is real but much smaller than the abstract implies, and the design-choice ablations are validated only on training fit.\n\nThe speed claim, by contrast, is clean and believable. The architecture is plausible: spatially distributed MLPs run once per pose, coefficients are interpolated, and rasterization is standard. 166 fps vs 10 fps is a real order-of-magnitude improvement, and the training time cut (17.5h vs 100h) is also credible. The novel mechanism itself is worth credit: instead of interpolating Gaussian property offsets directly, the MLPs output coefficients that combine a freely learned per-Gaussian basis. That avoids the smoothing artifact in Fig. 6 and is a genuine conceptual step beyond the single-MLP and CNN baselines. The control-point surface constraint is more incremental, but the ablation shows it helps under novel poses.\n\nOther soft spots: metrics are from a single camera view per sequence, no error bars, and the test-time PCA projection (Sec. 3.3) pulls novel poses onto the training pose space, so even Table 2 is somewhat narrowed. The dependence on the AnimatableGaussians template is a fair concern, but since the baseline shares that template, it doesn't break the comparison. It does limit generality.\n\nWho's this for: people working on real-time animatable avatars and Gaussian splatting. It deserves a serious referee and a conditional accept path, but the authors need to re-evaluate on held-out frames, report per-sequence and multi-view numbers, and soften the 'finer details' claim to match what the data actually show.","headline":"Genuinely new basis-interpolation idea and an honest speed win, but the quality comparison is mostly in-sample; the single held-out result is much closer.","tokens_in":15797,"tokens_out":2864,"would_cite":true,"duration_ms":28379,"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":"The paper claims that a Gaussian human avatar representation based on pose-only spatially distributed MLPs and per-Gaussian offset bases reconstructs high-fidelity pose-dependent appearance and renders at 166 fps, about 16 times faster…","keywords":["3D Gaussian Splatting","human avatar reconstruction","pose-dependent appearance","spatially distributed MLPs","Gaussian offset basis","inverse-distance interpolation","control points","real-time rendering"],"falsifier":"Render a sequence with the canonical template deliberately perturbed (for example, shift or rotate the mesh by a few centimeters before anchor points, control points, and neutral Gaussian positions are sampled), then compare PSNR and LPIPS with the unperturbed run; a sharp drop would show the fixed-template assumption is load-bearing.","tokens_in":14830,"feed_emoji":"👤","tokens_out":11058,"duration_ms":105942,"temperature":0.7,"pith_summary":"This paper aims to establish that a 3D Gaussian human avatar can capture pose-dependent appearance details that currently require a heavy convolutional network, while still rendering in real time. The proposed representation sets about 300 small MLPs at fixed anchor points on a canonical template; each MLP takes only the pose as input and emits coefficient vectors, and each Gaussian obtains its property offsets by inverse-distance interpolation of the nearest three anchor outputs combined with its own learned offset basis. The paper argues that because the basis vectors are learned freely while coefficients vary smoothly, the representation can produce sharply different Gaussian properties in neighboring regions without paying a per-Gaussian network cost. It reports better quantitative quality than the compared methods on the tested multi-view sequences, with rendering at 166 fps versus 10 fps for AnimatableGaussians.","feed_headline":"Gaussian avatars hit 166 fps while keeping fine pose-dependent detail","feed_subtitle":"Small pose-driven MLPs plus per-Gaussian offset bases replace a heavy network, improving quality and speed.","key_machinery":"The load-bearing object is the pair formed by the anchor coefficients $w^j_a = \\mathcal{E}^j(\\theta)$ output by $F=300$ spatially distributed MLPs located at anchor points, and the per-Gaussian offset bases $\\{\\delta\\Lambda_k\\}_{k=1}^{B}$, with $B=15$, covering rotation, scale, opacity, and SH coefficients. These are joined by inverse-distance interpolation: each pose passes through the MLPs once, and every Gaussian's coefficient vector is the distance-weighted combination $w_g = \\frac{\\sum_j \\gamma(x_0,x^j_a) w^j_a}{\\sum_j \\gamma(x_0,x^j_a)}$ of its three nearest anchors, with $\\gamma(x,y)=1/\\|x-y\\|^2$; the property offset is then $\\delta\\Lambda=\\sum_k w_g[k]\\,\\delta\\Lambda_k$ added to the neutral properties. Smoothly interpolated coefficients are deliberately combined with freely learned basis vectors, which is what lets the representation output sharply different Gaussian offsets in neighboring regions, and the control-point mechanism of Sec. 3.2 applies the same interpolation idea to position offsets to keep Gaussians on a surface layer.","core_discovery":"The central claim is that pose-dependent Gaussian appearance decomposes into a neutral mean, per-Gaussian learned offset bases $\\{\\delta\\Lambda_k\\}$, and pose-driven coefficients that are spatially interpolated from a small set of anchor MLPs, and that this decomposition recovers detail a single MLP misses while avoiding the per-frame cost of a large CNN. Each Gaussian's coefficient vector is $w_g = \\frac{\\sum_j \\gamma(x_0,x^j_a)\\,w^j_a}{\\sum_j \\gamma(x_0,x^j_a)}$ over the three nearest anchors, with $\\gamma(x,y)=1/\\|x-y\\|^2$, and its property offset is $\\delta\\Lambda = \\sum_{k=1}^{B} w_g[k]\\,\\delta\\Lambda_k$ added to the neutral properties. The same coefficient mechanism, routed through control points, also produces Gaussian position offsets and keeps Gaussians on a surface layer during novel-pose animation. On the sequences tested, the paper reports better appearance metrics than the compared methods and a rendering speed of 166 fps versus 10 fps for AnimatableGaussians.","pith_inferences":["An implicit consequence is that each Gaussian's offset basis defines a small linear subspace of appearance change; one could share basis sets across identities and learn only coefficients, which the paper does not explore.","The same anchor-and-control-point interpolation machinery should transfer to any articulated object with a template mesh and skinning weights, since nothing in the representation is human-specific beyond the pose vector.","Because each MLP is evaluated once per pose and the coefficients are then interpolated, the per-frame cost scales with the number of anchors, not the number of Gaussians; a compiled or quantized version of the MLP ensemble could push rendering further toward interactive VR budgets, though the paper does not test that.","A direct extension would vary the interpolation kernel (for example, learned weights or Gaussian falloff) and measure whether the reported detail gain persists; the paper fixes the inverse-distance kernel."],"forward_implications":["High-fidelity pose-dependent avatars no longer need a heavy per-frame CNN: a single pose-vector pass through 300 small MLPs produces all Gaussian properties, which is why the reported speed reaches 166 fps.","The number of MLP evaluations per frame becomes independent of the number of Gaussians, so raising the Gaussian count from 200K upward carries a much smaller speed penalty than in network-per-Gaussian designs.","The proposed basis combination permits high-frequency detail without per-Gaussian network inference, a key difference from single-MLP designs that the ablation study supports.","Tying position offsets to control points makes novel-pose animation more robust, since Gaussians move with the surface rather than drifting freely inside the body.","The use of PCA projection at test time anchors the method to the training pose distribution, so novel poses are interpreted within that space, which the paper reports improves generalization."],"supporting_citations":[{"why":"Supplies the canonical template mesh, skinning weights, SMPL-X registrations, and the primary high-quality baseline the method must beat.","marker":"[28]"},{"why":"Provides the 3D Gaussian splatting rasterizer, the Gaussian property parameterization, and the L1 loss used in training.","marker":"[20]"},{"why":"Introduces spatially distributed MLPs for local appearance; the anchor-MLP design extends it by taking only pose as input.","marker":"[58]"},{"why":"A single-MLP baseline whose limited capacity motivates the spatially distributed design.","marker":"[42]"},{"why":"A single-MLP-with-learnable-code baseline that the paper ablates as 'Predict properties'.","marker":"[54]"},{"why":"The StyleUNet architecture behind AnimatableGaussians, which the paper identifies as the computational bottleneck it avoids.","marker":"[47]"},{"why":"The LPIPS perceptual loss used both as a training term and as an evaluation metric.","marker":"[56]"},{"why":"Supplies the AvatarRex dataset and the 16-view capture setup used in the main evaluation.","marker":"[59]"}],"fun_headline_variants":["Real-time 166 fps avatars with spatially distributed MLPs","High-fidelity Gaussian avatars at 166 fps via MLP interpolation","Pose-dependent detail at 166 fps without heavy networks","Spatial MLP interpolation beats slow networks for avatar detail","Faster avatars: 166 fps with better fine detail"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the canonical template mesh and skinning weights taken from AnimatableGaussians are accurate enough that the anchor points, control points, and neutral Gaussian positions sampled once on that mesh stay valid across all training and novel poses.","fun_headline_variants_meta":{"raw":{"variants":["Real-time 166 fps avatars with spatially distributed MLPs","High-fidelity Gaussian avatars at 166 fps via MLP interpolation","Pose-dependent detail at 166 fps without heavy networks","Spatial MLP interpolation beats slow networks for avatar detail","Faster avatars: 166 fps with better fine detail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1694,"prompt_tokens":1068,"completion_tokens":626,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":539}},"tokens_in":684,"tokens_out":626,"duration_ms":6203,"temperature":1.0,"reasoning_tokens":539,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:19:52.372287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a sequence with the canonical template deliberately perturbed (for example, shift or rotate the mesh by a few centimeters before anchor points, control points, and neutral Gaussian positions are sampled), then compare PSNR and LPIPS with the unperturbed run; a sharp drop would show the fixed-template assumption is load-bearing.","supporting_citations":[{"cited_title":"Ani- matable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling","cited_arxiv_id":null,"evidence_quote":"Supplies the canonical template mesh, skinning weights, SMPL-X registrations, and the primary high-quality baseline the method must beat."},{"cited_title":"Structured local radiance fields for human avatar modeling","cited_arxiv_id":null,"evidence_quote":"Introduces spatially distributed MLPs for local appearance; the anchor-MLP design extends it by taking only pose as input."},{"cited_title":"3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting","cited_arxiv_id":null,"evidence_quote":"A single-MLP baseline whose limited capacity motivates the spatially distributed design."},{"cited_title":"Stylea- vatar: Real-time photo-realistic portrait avatar from a single 10 video","cited_arxiv_id":null,"evidence_quote":"The StyleUNet architecture behind AnimatableGaussians, which the paper identifies as the computational bottleneck it avoids."},{"cited_title":"Avatarrex: Real-time expressive full- body avatars","cited_arxiv_id":null,"evidence_quote":"Supplies the AvatarRex dataset and the 16-view capture setup used in the main evaluation."}],"review_version":1}