{"id":"fd7e56f5-5518-4e3b-8a1c-aff4e495f91d","arxiv_id":"2504.14967","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 3D Gaussian head avatar method uses triplanes for neutral appearance and per-blendshape 1D feature lines for dynamic opacity offsets, reaching 10MB storage and 300 FPS with competitive fidelity.","lead":"This paper introduces a compact head-avatar representation that stores static face appearance in triplanes and dynamic expression-dependent texture changes as 1D feature lines, decoded into opacity offsets. It reports 10MB avatars with 300 FPS rendering and improved PSNR over prior Gaussian-based head avatar methods on the Nersemble dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The opacity-offset design in Sec. 3.3 can only reweight fixed canonical colors; expression-driven color changes such as skin-tone shifts or specular changes are outside the model, and the worse LPIPS vs GA/GHA is consistent with this ceiling.","rationale":"I read the paper as a system contribution whose central claim is that compact tensorial representations, with neutral texture in triplanes and dynamic details as opacity offsets, achieve high-fidelity dynamic appearance at low storage and real-time speed. The weakest point in the argument is the opacity-only dynamic model in Sec. 3.3. The formal issue is not merely that real skin has color dynamics; it is that the parameterization has no mechanism to change the RGB color of a surface point independently of opacity, since all RGB comes from the canonical triplane. Therefore, the central claim requires the empirical assumption that all dynamic appearances in the tested setting are expressible as opacity reweighting. The paper provides no direct test of that assumption, and the supplementary comparison with geometric offsets does not address color. The LPIPS deficit against GA and GHA gives additional reason to treat the dynamic appearance claim cautiously. I agree with the reader that this is the load-bearing assumption, and I do not see a more serious internal inconsistency. The comparison details, missing hyperparameters, and lack of code all matter for reproducibility, but they are secondary to the representational ceiling. The proposed test would settle whether the concern is practical or merely theoretical on Nersemble. Since the reader already marked the paper CONDITIONAL, my read does not change the verdict; it reinforces the need for that condition.","tokens_in":12812,"tokens_out":8802,"duration_ms":65986,"concrete_test":"Train a minimal variant on the same data in which the feature-line decoder outputs a 3-channel RGB offset in addition to (or instead of) the scalar opacity offset, keeping all other architecture and regularization choices identical. Evaluate on the same held-out self-reenactment frames for all nine Nersemble subjects. If the color-offset variant improves LPIPS by more than about 0.005 or PSNR by more than about 0.5 dB, the opacity-only model is a binding representational limit; if there is no meaningful improvement, the concern is empirically refuted for this benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is in Sec. 3.3: final opacity is alpha = alpha_c + theta(l_b(p), l_r(p)), while RGB always comes from a fixed canonical triplane. In 3DGS, the final pixel color is an alpha-weighted composite of these fixed colors and background, so any expression-dependent color change at a surface point must be emulated by changing only per-splat opacity. That restricts the achievable pixel colors to a low-dimensional manifold determined by the fixed canonical colors. If an expression produces an intrinsic color change at a point without a corresponding change in the canonical color set, no choice of opacity offset can reproduce it. The paper does not test a color-offset variant; the supplementary ablation in Table 4 only compares opacity offsets against geometric position/rotation/scale offsets, not against color offsets. Furthermore, the limitations paragraph in Sec. 4.5 does not disclose this opacity-only restriction, listing only mesh dependence and material/lighting decoupling. The reported LPIPS results are consistent with this concern: on both novel-view synthesis and self-reenactment, the method is worse than GA and GHA on LPIPS (Table 1), so the broad claim of high-fidelity dynamic appearance is not yet established beyond the specific benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a compact 3D Gaussian head avatar representation that stores neutral appearance in a triplane and per-blendshape dynamic texture as 1D feature lines decoded into opacity offsets. The method binds Gaussians to a FLAME mesh, renders with 3DGS, and introduces an adaptive truncated opacity penalty and class-balanced sampling. On nine Nersemble subjects, it reports ~10MB storage, 300 FPS, and higher PSNR than GA, GHA, GBS, and INSTA on novel-view and self-reenactment tasks, while using less storage than all baselines.","tokens_in":13083,"tokens_out":8230,"duration_ms":69744,"significance":"If confirmed, the method is practically significant: it demonstrates that a very compact tensorial representation can achieve real-time, low-storage head avatars with better PSNR than established baselines. The strengths are the clear empirical validation over nine subjects with multiple baselines, a plausible compression rationale, and ablations of the proposed components. The main weakness is that dynamic appearance is limited to opacity modulation, which the paper does not disclose as a limitation, and the reported LPIPS is consistently worse than Gaussian Avatars, so the perceptual-quality claim is not fully established.","major_comments":[{"comment":"Since the RGB color of each splat is obtained from the canonical triplane and is independent of expression, the only expression-dependent appearance channel is the opacity offset. This restricts dynamic appearance to reweighting of fixed colors; any expression-induced color change (e.g., skin-tone changes, specular variations) is unrepresentable. The paper's title and abstract claim 'expressive dynamic appearances,' and the limitations paragraph in Sec. 4.5 does not mention this restriction. Please either add a color-offset variant to the ablations (e.g., a 1D feature line decoded into an RGB offset) to test whether the opacity-only design is sufficient, or explicitly acknowledge this as a limitation and temper the claims accordingly.","section":"Sec. 3.3 (final-opacity formula)"},{"comment":"The reported LPIPS is worse than GA on both tasks (0.0594 vs 0.0514 novel view; 0.0772 vs 0.0667 self-reenactment) and worse than GHA on novel view. The text only highlights PSNR improvements and does not discuss this perceptual-metric regression. Because LPIPS is a standard perceptual measure, the claim of 'high-fidelity dynamic appearances' needs qualification. Please provide statistical significance tests, qualitative analysis of failure modes, or an explanation of the trade-off.","section":"Table 1 and Sec. 4.3"},{"comment":"The ablation table is formatted ambiguously; the checkmark columns for trip/fl/penalty/resample are not properly aligned in the text, and the verbal description ('the first two rows', 'the first and third rows') does not clearly map to the table rows. This makes it impossible to verify which component combinations are compared and the contributions claimed in Sec. 4.4. Please reformat the table so each row explicitly lists the components present.","section":"Table 2"},{"comment":"The rendering resolution used for the FPS measurement is not specified for GA, GBS, or the proposed method; the supplementary says GHA is tested at 1024x1024, but if other methods are tested at 802x550, the comparison is not at matched resolution. Please report FPS at identical resolutions (or state the resolution for each method) to support the real-time claim.","section":"Table 1 FPS column"}],"minor_comments":[{"comment":"The threshold tau in the adaptive truncated opacity penalty is never given a value in the text or supplementary; please report it and ideally a sensitivity analysis.","section":"Eq. (2)"},{"comment":"The statement 'Using only the leading expression coefficients achieves similar results' is unsupported; either cite an ablation or remove it.","section":"Sec. 3.3"},{"comment":"'10M storage' should be '10MB storage' for consistency with Table 1.","section":"Sec. 2.1"},{"comment":"The 'Others' variant (position/rotation/scale offsets) achieves higher PSNR on novel-view synthesis (36.39 vs 35.16), suggesting the opacity-only choice may not dominate in all conditions; please discuss this trade-off.","section":"Supplementary Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a computer vision/graphics journal. The main technical contribution is sound in its narrow scope, but the opacity-only dynamic appearance and the unaddressed LPIPS regression need to be resolved before acceptance. The ablation table formatting is also a serious readability issue that obstructs verification of the component-wise claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the headline claims are real—10MB per subject, 300 FPS on a 4090, and PSNR above GA, GHA, and GBS on Nersemble. The design is a clean compression: triplanes for neutral appearance, 1D feature lines per blendshape for opacity offsets, plus class-balanced sampling and an adaptive penalty. That combination is new, and the ablations show each component buying a measurable PSNR gain. This is a genuine step toward deployable head avatars.\n\nWhere it gets soft: LPIPS is worse than GA on both main tasks (0.059 vs 0.051 novel view; 0.077 vs 0.067 self-reenactment). So the broad \"high-fidelity dynamic appearance\" claim is too strong. The likely reason is the opacity-only dynamic model: canonical RGB from the triplane is fixed per splat, and expression change only reweights splats via opacity. That cannot synthesize an intrinsic color change at a surface point. The stress-test note is right, and the paper's own limitation section should say so. It also never tests a color-offset variant; Table 4 only compares opacity offsets to position/rotation/scale offsets. A color-offset ablation would either rescue the claim or expose the ceiling.\n\nMinor but real: the FPS comparison isn't apples-to-apples (GHA is measured at 1024x1024, the method at 802x550, per the supplement). The threshold τ in Eq. 2 is never given a value. The main ablation is one subject with no error bars. No code is released. None of these are fatal, but together they make the quantitative story less clean than the abstract suggests.\n\nBottom line: this deserves a serious referee. The core method is sensible, the benchmark is standard, and the results are a real improvement in cost and speed. A referee should ask for a color-offset comparison, matching FPS measurements, and a frank statement of the opacity-only limitation. I'd be happy to cite it once those are addressed.","headline":"The 10MB/300FPS head avatar is a real practical win, but the opacity-only dynamic model explains the LPIPS gap and the paper should own that limitation.","tokens_in":13676,"tokens_out":2990,"would_cite":true,"duration_ms":27614,"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":"A compact tensorial representation stores neutral face appearance in triplanes and expression-dependent wrinkles as 1D feature lines decoded into opacity offsets, enabling 3D Gaussian head avatars that render at 300 FPS with about 10MB…","keywords":["3D Gaussian Splatting","head avatar","dynamic texture","triplane","1D feature lines","opacity offset","blendshape","tensorial representation"],"falsifier":"Train the avatar on a subject who exhibits strong expression-dependent color changes (e.g., skin reddening, glossy highlights) and test on a held-out expression; if the rendered images fail to show those color shifts while geometry matches, the opacity-only assumption is falsified.","tokens_in":12594,"feed_emoji":"👤","tokens_out":6377,"duration_ms":51881,"temperature":0.7,"pith_summary":"This paper aims to show that a 3D Gaussian head avatar can capture expression-dependent texture changes without paying a large cost in speed or storage. The proposed representation uses tri-planes to store the neutral face's appearance and lightweight 1D feature lines, one per expression blendshape, to store dynamic details; these lines are interpolated by the current expression coefficients and decoded into small opacity offsets applied on top of the neutral opacity. On the Nersemble dataset, the method improves PSNR over the GA, GHA, and GBS baselines on novel-view synthesis and self-reenactment while rendering at 300 FPS and storing roughly 10MB per subject. If the result holds, photorealistic animatable avatars become light enough for real-time video conferencing and mobile deployment.","feed_headline":"Head avatars animate at 300 FPS in just 10MB","feed_subtitle":"Triplanes hold the neutral face, 1D feature lines encode wrinkles as opacity offsets, beating larger baselines in PSNR.","key_machinery":"The central object is the compact tensorial representation: a triplane $T=\\{T_{xy},T_{xz},T_{yz}\\}$ in canonical space for neutral appearance, and per-blendshape 1D feature lines $L_i=(L_x,L_y,L_z)\\in\\mathbb{R}^{3\\times n_{d2}\\times n_s}$ for dynamic texture. Feature lines are interpolated by blendshape coefficients $\\beta$ and jaw rotation bases via inverse distance weighting, projected onto axes and interpolated to give features $l(p)$, then decoded by a tiny MLP $\\theta$ into an opacity offset $\\Delta\\alpha=\\theta(l_b(p),l_r(p))$, with final opacity $\\alpha=\\alpha_c+\\Delta\\alpha$. This carries the argument by separating static appearance from expression-dependent dynamics in a way that compresses spatial redundancy (triplane) and expression redundancy (1D lines), enabling real-time rendering and low storage.","core_discovery":"The paper's central claim is that dynamic facial appearance can be factored into a static canonical part and a low-rank expression-dependent part: a triplane stores the view-dependent color of the neutral face, and per-blendshape 1D feature lines, interpolated by expression coefficients and jaw rotation bases, are decoded by a small MLP into an opacity offset added to the canonical opacity. This factorization lets the avatar reproduce wrinkles and other expression-specific details while keeping the model to about 10MB per subject and rendering at 300 FPS. Combined with an adaptive truncated opacity penalty that forces static triangles to have zero offset and a class-balanced sampling that oversamples rare large expressions, the method improves PSNR over GA, GHA, and GBS on novel views and self-reenactment on the Nersemble dataset.","pith_inferences":["A natural extension the paper leaves implicit is to allow the decoder to output RGB offsets rather than only opacity offsets; this would capture expression-dependent color changes such as blushing or specular shifts that the current model cannot represent.","The same triplane-plus-feature-lines pattern could transfer to other blendshape-driven articulated objects, such as hands or bodies, wherever a low-dimensional basis spans the deformation space.","The reported 300 FPS is measured on a high-end RTX 4090; a testable extension is to profile the same model on edge or mobile GPUs to see whether the real-time claim survives on consumer hardware.","The class-balanced sampling based on vertex-displacement clustering could be reused for any dataset with skewed expression distributions, independent of the tensorial representation."],"forward_implications":["If the representation is correct, photorealistic head avatars can be reduced to roughly 10MB per subject, making them practical for fast network transmission and real-time rendering in mobile video conferencing.","The opacity-offset factorization suggests that dynamic details for novel expressions can be generated by interpolating in feature-line space, so unseen blendshape combinations should render without per-subject retraining.","Using only the leading FLAME PCA blendshapes and a small number of jaw bases keeps storage and computation low, so adding more expressions to the model has modest cost.","The method's reliance on a tracked mesh means it inherits FLAME's coverage: complex hairstyles and mouth interiors remain outside the model's scope, and relighting is not supported."],"supporting_citations":[{"why":"Supplies the mesh-bound 3D Gaussian geometry and the GA baseline that the method extends and compares against.","marker":"[29]"},{"why":"Provides the 3D Gaussian Splatting rasterizer used for rendering.","marker":"[18]"},{"why":"Defines the FLAME face model whose blendshape coefficients and jaw rotation drive the dynamic feature lines.","marker":"[21]"},{"why":"Baseline method storing a Gaussian set per blendshape; the paper compares storage and fidelity against it.","marker":"[24]"},{"why":"Baseline GHA method modeling dynamic textures with MLPs; reference for the quality-versus-speed tradeoff.","marker":"[37]"},{"why":"Provides the multi-view Nersemble dataset used for training and evaluation.","marker":"[19]"},{"why":"Supplies the technique to extract linear jaw rotation bases for unifying jaw dynamics with linear blendshapes.","marker":"[22]"},{"why":"Provides the quaternion distance metric used to interpolate jaw feature lines via inverse distance weighting.","marker":"[16]"}],"fun_headline_variants":["Compact tensor head avatars: real-time 300 FPS, 10MB store","Triplane + 1D features: dynamic head avatars in 10MB at 300 FPS","Head avatars shrink to 10MB, keep 300 FPS dynamic texture","Express dynamic face details with light tensors at 300 FPS","Wrinkles in 1D: compact head avatars store 10MB, render 300 FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes every expression-dependent appearance change can be modeled as an opacity offset added to a fixed neutral triplane color, so if an expression changes surface color, specularity, or translucency, the model cannot capture it.","fun_headline_variants_meta":{"raw":{"variants":["Compact tensor head avatars: real-time 300 FPS, 10MB store","Triplane + 1D features: dynamic head avatars in 10MB at 300 FPS","Head avatars shrink to 10MB, keep 300 FPS dynamic texture","Express dynamic face details with light tensors at 300 FPS","Wrinkles in 1D: compact head avatars store 10MB, render 300 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2480,"prompt_tokens":881,"completion_tokens":1599,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":1484}},"tokens_in":497,"tokens_out":1599,"duration_ms":9434,"temperature":1.0,"reasoning_tokens":1484,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:35:56.625600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the avatar on a subject who exhibits strong expression-dependent color changes (e.g., skin reddening, glossy highlights) and test on a held-out expression; if the rendered images fail to show those color shifts while geometry matches, the opacity-only assumption is falsified.","supporting_citations":[{"cited_title":"3d gaussian blendshapes for head avatar animation","cited_arxiv_id":null,"evidence_quote":"Baseline method storing a Gaussian set per blendshape; the paper compares storage and fidelity against it."},{"cited_title":"Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians","cited_arxiv_id":null,"evidence_quote":"Baseline GHA method modeling dynamic textures with MLPs; reference for the quality-versus-speed tradeoff."},{"cited_title":"Nersemble: Multi-view radi- ance field reconstruction of human heads.ACM Transactions on Graphics (TOG), 42(4):1–14, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the multi-view Nersemble dataset used for training and evaluation."},{"cited_title":"Posevocab: Learning joint-structured pose embeddings for human avatar modeling","cited_arxiv_id":null,"evidence_quote":"Supplies the technique to extract linear jaw rotation bases for unifying jaw dynamics with linear blendshapes."},{"cited_title":"Metrics for 3d rotations: Comparison and anal- ysis","cited_arxiv_id":null,"evidence_quote":"Provides the quaternion distance metric used to interpolate jaw feature lines via inverse distance weighting."}],"review_version":1}