{"id":"7e690c06-731d-4d84-afff-6368a6ed0469","arxiv_id":"2607.28039","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-stage pipeline (Gaussian reenactment plus geometry-anchored masked diffusion) transfers cross-identity tongue dynamics, roughly doubling tongue-specific metrics over prior reenactment baselines.","lead":"TongueReenact is the first face-reenactment system that transfers tongue motion from a driving video onto a different person’s face. It matters for talking-head avatars and speech animation, where missing tongues make mouths look empty or wrong.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Abstract’s “>2× on every tongue-specific metric” is not supported by Table 1; Full-model numbers also disagree with Table 2 without a stated protocol change.","rationale":"The paper still reads as a real new_method contribution: tongue dynamics are a genuine blind spot across parametric, warping, and diffusion reenactment (§1, Fig. 1), and the two-stage design (GaS reenactment + every-step geometry-anchored masked diffusion with adaptive dilation) is coherent, with ablations moving in the expected direction (Table 2). I do not push REJECT.\n\nThe reader correctly flags the VLM (100 expert frames, own ablations only, then ~100% preference) as weak independent perceptual evidence. That concern is valid but not the most load-bearing piece of the strongest claim: Table 1’s tongue metrics are what underwrite “>2× on every metric,” and they already fail that sentence on LPIPS and Presence while Full scores disagree with Table 2. That is a primary-result reporting/stability issue, not only a side perceptual protocol issue.\n\nVerdict stays CONDITIONAL for the same practical reasons the reader gave (unify metric protocol, release code/data, add non-self-referential human study). My stress pass does not upgrade or downgrade that verdict; it reframes the weakest link toward the overstated and internally inconsistent quantitative headline rather than solely the VLM. agreement_with_reader is partial because we both treat evaluation as the soft spot, but disagree on which evaluation pillar is load-bearing for the central empirical sentence.","tokens_in":13391,"tokens_out":978,"duration_ms":67598,"concrete_test":"Freeze BiSeNet* and the exact mask→metric pipeline; re-score the identical 9,000 VFHQ pairs for (a) all Table 1 methods and (b) Table 2 Full/ablations. Publish per-metric ratios vs each baseline and the Table1–Table2 delta for Full. The abstract claim fails if LPIPS or Presence is <2× vs every baseline, or if Full IoU/Presence shift >10% between the two table settings without a documented subset change.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest empirical claim is not merely “tongue transfer works,” but that the method “improve[s] … more than two times over all baselines on every tongue-specific metric” (abstract; mirrored in the reader’s strongest_claim). Table 1 does not support that wording. Relative to the best baseline in each row: LPIPS-tongue only moves 0.2591→0.2306 (~1.12×, lower better); Tongue Presence 0.7367→0.7891 (~1.07×); Tongue IoU 0.1918→0.3582 (~1.87×); only Tongue Area Similarity is clearly >2× (0.1330→0.4338). So “every” and “more than two times” are jointly false on the reported SOTA table.\n\nCompounding this, the same “Full (Ours)” system is scored differently in the ablation table with no disclosed subset or protocol change (Table 1 vs Table 2: LPIPS 0.2306 vs 0.2089; Presence 0.7891 vs 0.8851; IoU 0.3582 vs 0.3956; Area 0.4338 vs 0.4779). Tongue IoU/Presence/Area (and the tongue crop for LPIPS) necessarily depend on automatic tongue masks—almost certainly the authors’ BiSeNet*, which also drives training and inference (§4.1–4.2). If Full scores drift this much across tables, the multiplicative “2× vs all baselines” headline is not a stable, protocol-locked finding. The VLM study (Table 4; Supp. §2) is a secondary, self-referential preference check and cannot repair an overstated primary metric claim.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes TongueReenact, claimed as the first cross-identity tongue-dynamics transfer framework for face reenactment. A foundation-model-assisted bootstrapping pipeline (SAM + face parsing + iterative human refinement) trains a dedicated BiSeNet* tongue segmenter for in-the-wild frames. A two-stage system then (i) produces a tongue-absent, geometry-grounded render via FLAME-conditioned Gaussian splatting and (ii) synthesizes tongue appearance with a latent masked diffusion model that recomposes the reference latent outside an adaptively dilated driving tongue mask at every DDIM step, dual-conditioned on source geometry and driving mouth appearance. Experiments on VFHQ compare against five reenactment baselines on four tongue-specific metrics, with ablations, temporal-consistency checks, mouth-masked ArcFace identity checks, and a Qwen3-VL perceptual protocol fine-tuned on 100 expert frames.","tokens_in":13802,"tokens_out":1458,"duration_ms":30789,"significance":"If the technical claims hold, the work fills a genuine and previously ignored gap: existing parametric, warping, and diffusion reenactment systems leave the oral interior unconstrained or empty, which matters for speech and extreme expression. The bootstrapped segmenter, geometry-anchored masked diffusion with per-step recomposition, and adaptive dilation are concrete, reusable components, and the diffusion stage is explicitly designed to couple to other geometry-faithful backbones. Table 1 shows consistent gains over strong baselines on tongue IoU and area similarity; ablations move in the expected direction; temporal and masked-identity checks address two natural failure modes. These are real contributions for face animation and avatar pipelines, even after the overstated multiplicative headline is corrected.","major_comments":[{"comment":"Abstract and §5.3 claim “improvements of more than two times over all baselines on every tongue-specific metric.” Table 1 does not support that wording. Relative to the best baseline per row: LPIPS-tongue improves only 0.2591→0.2306 (~1.12×, lower better); Tongue Presence 0.7367→0.7891 (~1.07×); Tongue IoU 0.1918→0.3582 (~1.87×); only Tongue Area Similarity is clearly >2× (0.1330→0.4338). “Every” and “more than two times” are jointly false on the reported SOTA table. The abstract and corresponding discussion must be rewritten to match the actual ratios (or report absolute deltas with clear baselines).","section":"Abstract; §5.3; Table 1"},{"comment":"The same “Full (Ours)” system is scored differently in Table 1 vs Table 2 with no disclosed subset, seed, or protocol change (LPIPS 0.2306 vs 0.2089; Presence 0.7891 vs 0.8851; IoU 0.3582 vs 0.3956; Area 0.4338 vs 0.4779). Tongue Presence/IoU/Area (and the tongue crop for LPIPS) depend on automatic masks—almost certainly BiSeNet*, which also drives training and inference (§4.1–4.2). Unexplained drift undermines the stability of the primary quantitative claim. Please lock a single evaluation protocol (split, mask source, post-processing), report Full under that protocol in both tables, and state whether metrics use BiSeNet* or an independent tongue detector.","section":"Table 1 vs Table 2; §5.2–5.4"},{"comment":"The VLM protocol (Table 4; Supp. §2) fine-tunes Qwen3-VL on 100 expert frames that compare the authors’ Full model side-by-side against their own ablations, then applies it to 1,000 further frames, yielding 99.9–100% “best tongue/overall.” This is a useful scaled preference check but is not an independent human study: the labeled set and forced four-way comparison can bias the model toward Full, and there is no held-out human agreement on the 1,000-frame pool or comparison against external baselines (e.g., X-NeMo). Either (i) report inter-rater / VLM–human agreement on a held-out expert set and include SOTA baselines in the VLM ballot, or (ii) clearly demote Table 4 to an ablation-preference diagnostic and not “replicat[ing] expert annotation at scale” as primary perceptual evidence.","section":"§5.6; Table 4; Supplementary §2"}],"minor_comments":[{"comment":"Figure numbering jumps from Fig. 3 to Fig. 5 in the main text; Fig. 4 appears missing or mislabeled.","section":"§5 / Figures"},{"comment":"Algorithm 1 Phase 1 line “V_ext += GaS(f, FLAME(t))” is underspecified: it is unclear how reenacted frames enter the segmentation bootstrap and whether this couples segmenter training to the reenactment backbone.","section":"Algorithm 1; §4.1"},{"comment":"Stratification thresholds τ1, τ2 and dilation hyperparameters ρ, r_min are named but not given numeric values in the main text (only ρ=0.3, r_min=5 at inference in §5.1). A short hyperparameter table would aid reproducibility.","section":"§4.1–4.2; §5.1"},{"comment":"Related work on oral/tongue modeling in speech animation and medical tongue segmentation is thin beyond TongueSAM; a brief pointer to talking-head / viseme work would better situate the “first framework” claim.","section":"§2.4"},{"comment":"Identity-preservation paragraph in §5.4 reports ArcFace gaps but no table; adding a small numeric row (full vs mouth-masked) would make the claim easier to verify.","section":"§5.4"}],"recommendation":"major_revision","confidential_remarks":"The core engineering idea is plausible and likely useful; I would not reject on novelty grounds. The main risk is overclaim in the abstract plus self-referential evaluation (BiSeNet*-dependent tongue metrics and a VLM trained only on Full vs internal ablations). If the authors correct the 2× wording, reconcile Table 1/2, and either strengthen or demote the VLM study, this is a reasonable accept after revision. Scope fit for a CV venue is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a legitimate first cut at cross-identity tongue transfer in reenactment, and the engineering story is clear. The abstract’s strongest number claim is not supported by their own SOTA table.\n\nWhat is actually new is the end-to-end framing, not the parts. FLAME, GAGAvatar-style 3DGS, SAM pseudo-labels, BiSeNet, and every-step latent mask recomposition are standard. Nobody in the cited reenactment line (GPAvatar, Portrait4D, LivePortrait, X-NeMo, X-Portrait) treats tongue dynamics as a dedicated transfer target, and Figure 1 is fair about that failure mode. The bootstrapped in-the-wild tongue segmenter, dual guidance (source geometry + driving mouth crop), and adaptive dilation are a sensible package. Table 1 beats five baselines on all four tongue metrics; ablations move the right way; mouth-masked ArcFace and temporal mouth metrics are the right checks for identity leak and flicker.\n\nSoft spots, in proportion. The abstract says “more than two times over all baselines on every tongue-specific metric.” Against the best baseline per row in Table 1 that is false for LPIPS-tongue (~1.12×) and Presence (~1.07×), borderline for IoU (~1.87×), and only clearly true for Area Similarity. That is overclaim, not a math error. Full (Ours) also disagrees between Table 1 and Table 2 with no stated subset or protocol change (IoU 0.358 vs 0.396, Presence 0.79 vs 0.89). Absolute IoU stays modest. Tongue metrics almost certainly depend on their own BiSeNet*, so evaluation is partly self-coupled. The VLM protocol (100 expert frames, then 1k auto, near-100% preference vs own ablations) is a useful sketch, not independent human evidence. No code/data, no error bars.\n\nWho it is for: people building talking-head and avatar stacks who care about oral interiors. Citation pattern looks normal for the area. Serious thinking, honest gap, overstated headline. I would send it to referees; I would not bring it to reading group unless someone is deep in reenactment. Engage if you work on mouths or speech-driven faces; fix the metric wording and unify Full protocols before trusting the 2× line.","headline":"Real gap, working pipeline, but the abstract’s “>2× on every tongue metric” does not match Table 1, and Full scores drift between tables.","tokens_in":14533,"tokens_out":596,"would_cite":true,"duration_ms":15521,"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":"Face reenactment can finally transfer tongue motion from one person to another by anchoring diffusion synthesis to driving-frame geometry.","keywords":["face reenactment","tongue synthesis","latent diffusion","Gaussian splatting","tongue segmentation","geometry-anchored inpainting","cross-identity transfer"],"falsifier":"Run the same VLM protocol (or real expert raters) on held-out pairs where an independent tongue segmenter and human labels disagree with the paper’s full model, or swap in a different reenactment backbone and check whether Tongue IoU, area similarity, and best-tongue rates still more than double the listed baselines.","tokens_in":14180,"feed_emoji":"👅","tokens_out":902,"duration_ms":18591,"temperature":0.7,"pith_summary":"Existing face reenactment systems move pose and expression well but leave the mouth interior empty or wrong, because the tongue is missing from their geometric models and cannot be invented by warping source pixels. This paper claims the first end-to-end way to copy tongue dynamics across identities: bootstrap a tongue segmenter without hand-labeled data, reenact the face with a geometry backbone that produces a tongue-free render, then fill only the mouth with a masked latent diffusion model guided by both that render and the driving tongue appearance. Adaptive mask dilation smooths the join with lips and teeth. On cross-identity pairs the method more than doubles every tongue-specific automatic score versus strong baselines, and a vision-language model trained on expert labels rates the full system as best almost every time. A sympathetic reader cares because speech and extreme expression look anatomically broken without the tongue, and the pipeline is designed to plug into other geometry-faithful reenactment backbones.","feed_headline":"Reenactment finally copies the tongue across faces","feed_subtitle":"Geometry-locked diffusion more than doubles tongue-transfer scores over leading baselines","key_machinery":"Geometry-anchored latent masked diffusion: at every DDIM step the denoised latent is forced back to the reenacted reference outside a driving-frame tongue mask (adaptively dilated), while dual guidance encoders inject source facial structure and driving tongue appearance so synthesis is spatially locked rather than free hallucination.","core_discovery":"Cross-identity tongue dynamics can be transferred in face reenactment by decoupling geometric alignment from generative fill: a Gaussian-splatting stage driven by FLAME parameters yields a source-identity, tongue-absent render whose mouth region is then synthesized by a latent masked diffusion model that is recomposed against the reference latent at every denoising step, conditioned on both source geometry and the driving tongue crop, with the active region set by a bootstrapped tongue mask that is adaptively dilated.","pith_inferences":["The same geometry-anchor-plus-masked-diffusion pattern could extend to other transient oral structures (teeth gaps, saliva, extreme jaw interiors) that parametric meshes leave empty.","If the VLM judge is validated against a larger multi-annotator panel, automated oral-region QA could become standard in portrait-animation benchmarks.","Failure modes will concentrate where the driving tongue is barely visible or the reenactment backbone misplaces the mouth, because the mask and reference both inherit that geometry."],"forward_implications":["Reenacted talking heads can show visible, driving-matched tongue articulation instead of empty or blurred mouths during speech and extreme expressions.","Any geometry-faithful reenactment backbone can be paired with the same masked diffusion stage without redesigning the whole avatar pipeline.","Tongue-specific metrics (presence, IoU, area similarity, regional LPIPS) become usable benchmarks for oral interior fidelity beyond FID-style global scores.","A bootstrapped segmenter removes the need for large curated tongue masks before training reenactment systems on in-the-wild video."],"fun_headline_variants":["Geometry-anchored diffusion transfers tongue motion across faces","Cross-identity tongue dynamics via masked latent fill","FLAME-guided render plus diffusion synthesizes reenacted tongues","Bootstrapped masks let diffusion double tongue-transfer scores","Source geometry locks tongue synthesis in face reenactment"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a vision-language model fine-tuned on only about a hundred expert side-by-side labels can stand in for human judgment of tongue quality on a thousand more frames.","fun_headline_variants_meta":{"raw":{"variants":["Geometry-anchored diffusion transfers tongue motion across faces","Cross-identity tongue dynamics via masked latent fill","FLAME-guided render plus diffusion synthesizes reenacted tongues","Bootstrapped masks let diffusion double tongue-transfer scores","Source geometry locks tongue synthesis in face reenactment"]},"model":"grok-4.5","effort":"low","cost_usd":0.003631,"raw_usage":{"total_tokens":1141,"prompt_tokens":695,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":36308000,"prompt_tokens_details":{"text_tokens":695,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":384,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":695,"tokens_out":62,"duration_ms":8408,"temperature":1.0,"reasoning_tokens":384,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T19:42:11.366968+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same VLM protocol (or real expert raters) on held-out pairs where an independent tongue segmenter and human labels disagree with the paper’s full model, or swap in a different reenactment backbone and check whether Tongue IoU, area similarity, and best-tongue rates still more than double the listed baselines.","supporting_citations":[],"review_version":1}