{"id":"40fcff92-7a43-4cce-8d82-40406d5e35a4","arxiv_id":"2607.23657","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A torso-rooted FLAME+SMPL-X portrait model plus coarse-to-fine graduated routing improves monocular mesh recovery and jaw–expression disentanglement over face-only and body-only baselines.","lead":"GRAPE recovers an animatable head–neck–shoulder mesh from one portrait photo by fusing FLAME with an SMPL-X torso and regressing pose before expression. It matters for talking avatars and 3D portrait tools that need a stable neck joint and a jaw that actually opens.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The flagship disentanglement metric (JSR) is computed from the same nose–chin signal the model is explicitly trained to match via Lnc, so the Table 2 gap may reflect training-on-the-metric rather than genuine jaw–expression disentanglement.","rationale":"The reader identified pseudo-label staging bias as the weakest assumption. I agree that is a real concern, but I locate the sharper soft spot one step downstream: even granting perfect pseudo-labels, the evaluation of the central disentanglement claim is confounded because Lnc (training) and JCR/JSR (evaluation) are the same monocular nose–chin signal. This is a correctness risk on the evidence, not an accusation of misconduct — the paper is unusually transparent: Table 5's per-loss ablation is exactly what lets a reader see that Lnc, not the router, produces the bulk of the JSR gain, and §D candidly acknowledges JCR/JSR is a weak proxy. I only partially agree with the reader's framing because the mechanism is metric–loss coupling rather than label-pipeline echo.\n\nI do not move to REJECT: the representation/engineering contributions (PPM canonical injection, torso anchoring) are supported by independent benchmarks (NoW, Stirling, landmark NME) where no such circularity exists; the router's marginal effect (0.686→0.799) and the MPJPE ablation for the torso anchor are real; and the downstream RGBAvatar body-region metrics (Table 6) provide partially independent signal. I do not move to ACCEPT because the paper's most novel scientific claim — structural jaw–expression disentanglement — currently lacks evidence decoupled from its own training signal. CONDITIONAL with a sharpened condition (independent skeletal/4D validation, or the counterfactual baseline test above) is the right call.","tokens_in":20172,"tokens_out":3053,"duration_ms":129824,"concrete_test":"Counterfactual fine-tune: take the released TEASER regressor and fine-tune it on the same in-the-wild data with only Lnc (Eq. 14) added, no router or PPM changes. Then (a) recompute its HDTF JSR, and (b) run a jaw-only re-animation probe — zero the expression vector, drive with predicted jaw pitch, and measure whether 2D mouth opening still tracks the reference. If TEASER+Lnc reaches JSR ≳ 0.7 while still failing the jaw-only probe, JSR is confounded by train-on-the-metric and Table 2's comparison is invalid; if its JSR stays low or the probe shows real kinematic coupling, GRAPE's claim is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim's most distinctive component — improved jaw–expression disentanglement — rests on JSR (Eq. 23): Pearson correlation between predicted jaw pitch and JCR, a normalized nose-tip–chin-tip distance from 2D detections. But the training objective includes Lnc (Eq. 14), which directly penalizes the mismatch between projected nose–chin distance and the same detected nose–chin distance. In FLAME/PPM, jaw pitch is the dominant skeletal DOF that moves the chin relative to the nose, while expression blendshapes act mostly on the lip region — so a network supervised with Lnc is essentially trained to make its jaw channel reproduce the JCR signal. High JSR is then near-mechanical, regardless of whether articulation and expression are truly disentangled, and baselines (TEASER 0.371, SPECTRE 0.564, PEAR 0.537) were never trained on this signal, making the Table 2 comparison asymmetric.\n\nInternal evidence reinforces this: in Table 5, adding Lnc alone (V1.1→V1.2, router still disabled at αact=1.0) jumps JSR from 0.566 to 0.686 — already above every baseline — while the Graduated-Mask Router (V1.2→V2.2) contributes only the remaining 0.686→0.799. SMPLest-X, with no disentanglement design and a 10-dim expression space, still scores 0.591. So most of the headline JSR margin is attributable to supervision aligned with the metric, not to the graduated-routing mechanism the paper credits. The geometric claims (Tables 1, 3, NoW/Stirling) are unaffected; the concern is scoped to the disentanglement claim and its downstream-animation motivation.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript introduces GRAPE for monocular portrait mesh estimation. It constructs a Portrait Parametric Model (PPM) by inserting a FLAME2023 head into an SMPL-X torso in canonical space, with a torso-rooted spine–neck–head–jaw kinematic chain. A Progressive Anatomical Alignment network uses a frozen Sapiens-2 encoder, three anatomically specialized experts, and a Graduated-Mask Router that stochastically removes supervision from finer experts during training. Training combines offline staged pseudo-labels, feature distillation, landmarks, a projected nose–chin distance loss, foreground overflow constraints, and relative torso geometry. The paper reports improvements in image landmark NME, HDTF jaw–skeletal consistency, Nersemble/NoW/Stirling reconstruction, structural and loss ablations, and two downstream avatar applications.","tokens_in":20731,"tokens_out":10084,"duration_ms":85796,"significance":"If the central claims are substantiated, the work addresses a real gap between face-centric models with weak head–torso kinematics and body-centric models with limited facial fidelity. The explicit PPM construction and canonical injection are useful contributions, and the paper provides unusually broad empirical coverage: four image benchmarks, talking-video analysis, three 3D protocols, structural and training-objective ablations, and downstream tests with DiffPoseTalk and RGBAvatar. These are meaningful strengths. The standard NoW/Stirling gains are modest but consistent, while the larger Nersemble and JSR gains currently depend on insufficiently specified or partially self-referential evaluation. The work is potentially valuable for animatable portrait reconstruction, but its disentanglement and downstream conclusions need stronger independent evidence.","major_comments":[{"comment":"JSR is not independent of the training signal. Eq. (14) directly fits the projected nose-tip–chin-tip distance, while Eqs. (22)–(23) evaluate the Pearson correlation of jaw pitch with a per-clip normalization of that same detected distance. Moreover, Lord regresses pseudo-labels from App. B Step 2.2, where mouth-related expression is zeroed and mouth opening is assigned to the jaw before expression fitting. Table 2 can therefore measure consistency with the imposed proxy/staging rather than anatomical disentanglement. The stress-test’s row attribution is imprecise—V1.2 adds Lfd and V1.3 adds Lnc—but the overlap remains. Add validation against independent 3D/4D jaw motion and a no-Lnc/no-jaw-staged-label control, or narrow the claim.","section":"§3.2.3 Eq. (14); §C.2 Eqs. (22)–(23); App. B"},{"comment":"The router ablations appear internally inconsistent. Table 4 reports JSR 0.411 when stochastic graduated masking is disabled, whereas Table 5 V1.5 reports 0.701 with all losses and alpha_act=1.0, which Appendix A describes as the same always-active-mask setting. Table 5 also has alpha_act=0.25 scoring 0.801, slightly above the selected alpha_act=0.50 at 0.799. Please reconcile the configurations, define alpha_act unambiguously, state whether JSR is pooled or averaged over the 20 HDTF clips, and report per-clip confidence intervals and multiple training seeds. As written, the magnitude and even the source of the router effect are uncertain.","section":"Tables 4–5; Appendix A"},{"comment":"The PPM forward pass appears to apply jaw articulation twice. Algorithm 1 first evaluates FLAME with theta_jaw, while the final unified LBS is driven by a pose vector whose hierarchy again includes the jaw; Appendix A.1 repeats this description. Unless the implementation removes or reparameterizes jaw rotation in one of these stages, this double transformation would undermine the claimed animatable semantics. Eq. (1) also introduces a head scale s, but the canonical-injection algorithm only specifies eye-center translation. Please provide the exact forward equations, including which transform is baked into the template, which is applied by unified LBS, and where scale/rotation alignment occurs.","section":"Algorithm 1; §3.1.2; Appendix A.1"},{"comment":"The evaluation protocols are not sufficiently specified for the main reconstruction claims. §3.3 says the model is finetuned on LS3DW, CelebA, LaPa, and LFW, while §4.1 evaluates on test sets of the same four benchmarks without giving exact splits or exclusion rules; any image overlap would inflate Table 1. For Table 3, Nersemble compares method-specific mesh topologies and may mix FLAME2020 and FLAME2023 after GRAPE was pretrained on NersembleV2. Define the ground-truth registration, vertex correspondence, scale and pose alignment, subject/scene split, and baseline domain treatment, and release the evaluation scripts and IDs. This matters because the NoW/Stirling margins are small, while the opaque Nersemble margin is large.","section":"§3.3, §4.1, Tables 1 and 3; §C.3"},{"comment":"The MPJPE result is load-bearing for the torso-anchor claim, but its protocol is missing. The metric definition says only “head–shoulder joint error in millimeters”; no dataset, joint set, ground-truth source, coordinate frame, alignment procedure, or uncertainty estimate is given. If the reference joints come from Sapiens-2 or the same pseudo-label pipeline used for training, the 89.1-to-69.4 improvement is not independent validation of head–torso kinematics. Please specify the protocol and preferably evaluate against multi-view or motion-capture joint trajectories.","section":"Table 4; §4.3; §C.3"},{"comment":"The downstream evidence does not yet support the broad claim that GRAPE benefits talking-head generation. The DiffPoseTalk result consists of training-loss curves, but the compared parameterizations and reconstruction targets differ, so lower or smoother loss is not a direct measure of generated-video quality. RGBAvatar provides held-out metrics, but the improvements are small (PSNR 21.08 to 21.69; PSNR-b 17.90 to 18.64) on only ten clips with no variance. Add held-out perceptual, identity, temporal, and lip-motion metrics under a common output representation over more sequences, or temper the abstract’s downstream claim.","section":"§4.5, Figure 8, Table 6"}],"minor_comments":[{"comment":"The checkmark columns make it difficult to determine which component each V1.x row adds. Replace some checkmarks with explicit component names and align the row descriptions with Table 4.","section":"Table 5"},{"comment":"Use p and alpha_act consistently. State explicitly whether the value is the probability of activating an expert’s supervision or of dropping it, and give the test-time setting.","section":"§3.2.2 and Appendix A"},{"comment":"“Omit the weight of each loss term without loss of generality” is not correct: the relative weights determine the graduated trade-offs. List the final weights or provide a sensitivity analysis.","section":"Eq. (9)"},{"comment":"Define d_t before using it in the normalized expression, give epsilon, and state whether mu_d and sigma_d are computed per subject, per clip, or over all HDTF frames.","section":"Eq. (22)"},{"comment":"Several equations, labels, and highlighted regions are difficult to read at print size. Increase font sizes and make the deep-red/deep-blue encodings distinguishable in grayscale.","section":"Figures 3–6"},{"comment":"Fix typographical errors such as “SPECTR” in Figure 5 and “PMMFitting” in Figure 11.","section":"Figures 5 and 11"},{"comment":"Provide the random seeds and exact IDs for the 20 HDTF evaluation videos and the 10 RGBAvatar clips to make the small-sample comparisons auditable.","section":"§4.1 and §C.4"},{"comment":"Define J/SMPL_T, W_unified, pi, and all parameter blocks in the algorithm itself, and repair the formatting so that the variable names and comments are legible.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"Several important baselines, including TEASER and PEAR, share authors with this submission. I do not view that overlap as improper, but because the headline JSR benchmark is custom and the evaluation clips are randomly selected, release of evaluation code, model checkpoints, clip IDs, and derived labels would materially improve auditability."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core here is practical, not theoretical: a FLAME head fused to an SMPL-X torso via canonical eye-center injection, plus a coarse-to-fine expert stack that actually keeps neck and shoulders in the monocular portrait problem. That is the right diagnosis of floating-head methods, and the mesh overlays and NME/NoW-style numbers back a real improvement over face-only regressors and weak-face body models.\n\nWhat is new is the combination, not the parts. Canonical injection, graduated masking of expert losses, Pixel3DMM distillation, and Sapiens keypoints are each familiar. Stitched together they form a coherent portrait pipeline, with ablations (Tables 4–5) and two downstream checks (DiffPoseTalk training stability, RGBAvatar neck–shoulder continuity) that make the engineering claim credible for avatar work.\n\nThe soft spot is scoped, not fatal. The flagship jaw story leans on JSR: correlation of jaw pitch with a normalized nose–chin distance. Training also has Lnc on that same projected distance. Table 5 shows adding Lnc alone lifts JSR from ~0.57 to ~0.69 before the router does much; baselines were not trained on that cue, so the Table 2 gap is partly metric-aligned supervision, not pure proof of anatomical routing. Pseudo-labels from a staged TEASER/ProHMR fitter add a milder self-consistency loop. Geometric claims (landmarks, Nersemble MVE/LVE, NoW/Stirling) do not depend on that loop and still look fine. Missing code/labels and some AI-generated figure inputs are ordinary CV friction, not a reason to dismiss the work.\n\nWho it is for: people building talking heads and upper-body avatars who need controllable neck/jaw parameters, not pure face identity papers. Math is standard LBS + losses; citations are appropriate. I would send it to referees. Engage if you care about portrait kinematics; treat JSR as supportive, not decisive, until someone checks jaw against stronger 4D or dental supervision.","headline":"Solid portrait-mesh engineering with a real torso-rooted model; the jaw-disentanglement headline is softer than the geometric gains because JSR is partly trained via Lnc.","tokens_in":21518,"tokens_out":523,"would_cite":true,"duration_ms":22648,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Portrait meshes recover better when pose, jaw, and expression are estimated in anatomical order rather than all at once.","keywords":["portrait mesh estimation","talking head","anatomical disentanglement","3D morphable models","jaw-expression separation","kinematic chain","monocular reconstruction"],"falsifier":"On talking clips with stable head pose, if predicted jaw pitch stayed weakly correlated with normalized nose-to-chin distance while expression still drove mouth opening—matching or worse than strong face regressors—and head–shoulder joint error did not drop when the torso anchor was present, the central disentanglement claim would fail.","tokens_in":21126,"feed_emoji":"🗣️","tokens_out":842,"duration_ms":21234,"temperature":0.7,"pith_summary":"Monocular portrait recovery fails in three linked ways: face-only models float the head without a neck, body models lack expressive faces, and networks open the mouth with expression blendshapes instead of the jaw. GRAPE treats the problem as progressive anatomical alignment. It builds a torso-rooted Portrait Parametric Model by attaching a high-fidelity face head to a body torso in canonical space, then regresses parameters with coarse-to-fine experts whose deeper losses are randomly masked so the network cannot shortcut jaw motion through expression. Multi-source cues—keypoints, feature distillation, foreground masks, and relative geometry—supervise the weakly labeled neck and shoulders. The claim is that this graduated routing yields tighter image alignment, more coherent head–torso pose, and jaw motion that tracks skeletal mouth opening, which in turn makes talking-head and avatar training cleaner.","feed_headline":"Portrait meshes that put the jaw before the smile","feed_subtitle":"Routing pose then expression fixes floating heads and dead jaws in monocular recovery","key_machinery":"Progressive Anatomical Alignment with a Graduated-Mask Router: anatomy-aware experts predict global shape/camera, then skeletal pose (including jaw), then expression, while Bernoulli masks randomly drop deeper supervision so expression cannot absorb jaw errors during training.","core_discovery":"GRAPE establishes that monocular articulated portrait mesh estimation improves when representation, routing, and supervision all follow the torso→neck→head→jaw→expression hierarchy: a torso-rooted hybrid parametric model plus progressive experts gated by a Graduated-Mask Router disentangles camera and head pose and forces jaw articulation before expression, outperforming flat face- and body-centric regressors on alignment, kinematics, and downstream animation.","pith_inferences":["Any monocular body part with a rigid-joint vs soft-deformation tradeoff (hands, spine) may benefit from the same coarse-to-fine loss masking pattern.","If pseudo-label staging is the real source of jaw purity, end-to-end methods without staged labels may need an equivalent hard architectural prior to match JSR-style metrics.","Combining the explicit kinematic scaffold with hair and clothing representations is the natural next stress test of whether the torso anchor still helps when the silhouette is no longer skin-tight."],"forward_implications":["Portrait avatars can be driven with a real neck joint instead of centrifuge-like head spins around the face center.","Talking-head trainers get parameter tracks where jaw, not blendshapes, carries skeletal mouth opening, stabilizing audio-driven animation learning.","Animatable 3D avatar pipelines gain continuous head–neck–shoulder surfaces instead of floating heads over mismatched torsos.","Face and body reconstruction lines can share one hybrid mesh rather than trading facial fidelity against kinematic anchors."],"fun_headline_variants":["GRAPE routes torso to jaw before expression in portrait meshes","Torso-rooted PPM ends floating heads in monocular portraits","Graduated-Mask Router puts jaw articulation ahead of blendshapes","Progressive experts follow neck-head-jaw hierarchy for mesh recovery","Canonical FLAME-SMPL-X merge disentangles pose from expression"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The offline staged pseudo-labels used as training targets are assumed to already separate jaw from expression and to place shoulders well enough that the network learns real anatomy rather than the label pipeline’s staging habits.","fun_headline_variants_meta":{"raw":{"variants":["GRAPE routes torso to jaw before expression in portrait meshes","Torso-rooted PPM ends floating heads in monocular portraits","Graduated-Mask Router puts jaw articulation ahead of blendshapes","Progressive experts follow neck-head-jaw hierarchy for mesh recovery","Canonical FLAME-SMPL-X merge disentangles pose from expression"]},"model":"grok-4.5","effort":"low","cost_usd":0.003764,"raw_usage":{"total_tokens":1213,"prompt_tokens":824,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":37644000,"prompt_tokens_details":{"text_tokens":824,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":299,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":824,"tokens_out":90,"duration_ms":6459,"temperature":1.0,"reasoning_tokens":299,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T16:30:20.041569+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On talking clips with stable head pose, if predicted jaw pitch stayed weakly correlated with normalized nose-to-chin distance while expression still drove mouth opening—matching or worse than strong face regressors—and head–shoulder joint error did not drop when the torso anchor was present, the central disentanglement claim would fail.","supporting_citations":[],"review_version":1}