{"id":"e27ded81-401c-4612-8f08-2283d228463c","arxiv_id":"2505.15385","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EVA reconstructs an actor-specific avatar from multi-view video that renders in real time and allows independent control of body motion, hand gestures, and facial expressions through a disentangled two-layer Gaussian representation.","lead":"EVA builds a photorealistic, real-time digital human from roughly 120 synchronized cameras, then animates that copy with separate controls for the body, hands, and face. A generalist might care because this is a concrete step toward lifelike avatars for VR meetings, games, and film where expressions and motions are edited independently.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The expressive-control claim rests on unvalidated FLAME expression tracking; without an independent check of psi, the reported gains do not demonstrate faithful expression control.","rationale":"The central claim is that EVA achieves real-time, photorealistic rendering with independent control of body, hands, and face, and outperforms DDC and ASH. For that claim to hold, the expression parameter channel psi must faithfully represent facial expressions. The paper's tracking pipeline optimizes psi against 2D landmarks and Chamfer fits, but never compares the recovered psi to an independent ground truth. This is exactly the reader's weakest_assumption, and it is load-bearing: all expression-control demonstrations, the 'novel pose/expression' evaluation, and the expressive superiority over ASH/DDC depend on psi being a reliable control signal. If psi is inaccurate, the appearance networks could memorize mappings from a specific noisy psi distribution to training images, and the claimed generalization to unseen expressions would be unsupported. The concern is a verification gap rather than a demonstrated flaw: the method is internally consistent, the ablations are informative, and the real-time numbers are plausible. I therefore agree with the reader's CONDITIONAL verdict and do not recommend changing it.","tokens_in":27961,"tokens_out":5291,"duration_ms":53298,"concrete_test":"On held-out test frames, independently recover FLAME expression parameters from the 20 close-up face cameras using an established multi-view face tracker or manual landmarks on high-resolution crops, and compare with EVA's tracked psi. Then render EVA with both psi sets and measure face-crop LPIPS and landmark distance to ground-truth close-up images. If the independent psi disagrees strongly with EVA's psi, or renders no closer to ground truth than EVA's own psi, the closed tracking loop is unvalidated and the expressive-control claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central capability is independent, faithful control of facial expressions via FLAME parameters psi (Sec. 3.2, Eq. 3), and its quantitative case for 'novel pose/expression synthesis' (Tab. 1) is evaluated by driving EVA with psi produced by the same tracking pipeline used to train the appearance model. Nowhere is psi validated against an independent source: no face mocap, no manual annotation on held-out frames, no comparison with a high-quality multi-view FLAME fit from the 20 close-up cameras, and no quantitative check that a user-specified psi produces the intended expression. The tracking objective is 2D landmark reprojection (MediaPipe/FAN) plus Chamfer distance to per-frame reconstructions (Sec. D.2). Sparse 2D landmarks are ambiguous for FLAME's high-dimensional expression space, especially for jaw/lip and eyelid coefficients; eye gaze is deliberately zeroed. If the tracked psi is biased or noisy, the U-Nets can learn a mapping from that specific noisy psi distribution to the training images, which will not transfer to novel expression coefficients from other sources such as audio, retargeting, or user sliders. The rendering-quality gains in Tab. 1 would then partly reflect training-set memorization rather than a generalizable control signal, and the qualitative 'independent expression change' demos would not establish faithful control. This is a load-bearing verification gap, not an observed failure; the architecture is plausible and the ablations are internally consistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"EVA proposes an actor-specific, real-time animatable avatar system that separates the human representation into an expressive template geometry layer (a DDC-style deformable body model stitched to a personalized FLAME-based head avatar) and a disentangled 3D Gaussian appearance layer in which separate U-Nets predict body and face Gaussian parameters. The method takes skeletal motion and FLAME expression parameters as inputs, aiming for independent control of body, hands, and face. The authors introduce a new multi-view dataset of four subjects with strong facial expressions, evaluate against DDC, ASH, AnimatableGaussians, and a reimplemented GaussianAvatars, and report quantitative improvements in LPIPS/PSNR/SSIM plus qualitative demonstrations of novel-view and novel-expression synthesis.","tokens_in":28213,"tokens_out":3612,"duration_ms":34158,"significance":"If the expressive-control claim is validated, EVA would be a practically significant contribution: it combines a deformable, expressive template with Gaussian splatting in a two-module design, supports direct FLAME-expression driving (enabling audio-driven avatars), and reaches real-time 2K rendering. The paper is unusually thorough in documenting implementation details, including the head-stitching procedure, the optimization losses, and several ablations. The main load-bearing weakness is that the facial expression parameters used both for training targets and for test-time control signals come from the same unvalidated tracking pipeline, so the reported gains do not yet demonstrate faithful, generalizable expression control. The quantitative comparison also lacks variance analysis and is confined to the authors' dataset, which tempers the strength of the 'surpasses state-of-the-art' claim.","major_comments":[{"comment":"The expression parameters psi are recovered from 2D landmarks (MediaPipe/FAN) and Chamfer fitting to per-frame 3D reconstructions, with no independent validation. The same tracking pipeline is used to create the training targets and the test-time control signals, so the 'Novel Pose/Expr.' numbers in Tab. 1 may reflect the model learning the specific noisy distribution of this tracker rather than a generalizable map from psi to appearance. Please add an independent check of psi, for example manual landmark annotation on held-out frames, a high-quality multi-view FLAME fit using the 20 close-up cameras, or a cross-driving experiment where psi comes from an external source (audio model, retargeted motion, or user sliders) and is evaluated for expression fidelity.","section":"Sec. 3.2, Sec. D.2, Eq. (3)"},{"comment":"The quantitative comparison aggregates metrics over four subjects without error bars, per-subject breakdowns, or significance tests. Several reported differences are small relative to the likely subject-to-subject variation (for instance, DDC actually leads PSNR in the Novel View row, 46.87 vs. 46.39), yet the text concludes that EVA 'surpasses' the baselines. Please report per-subject values and variance; if the advantage is not consistent, soften the claim to match the evidence.","section":"Tab. 1 and Sec. 4.2"},{"comment":"All quantitative evaluations are on the authors' new dataset, and the public DynaCap benchmark is excluded because its actors show limited facial expressions. Since the body-pose rendering quality is a central part of the claim, the absence of any public-benchmark body/pose comparison makes it hard to assess generalization. Please either report body-only metrics on DynaCap or a similar public sequence, or explicitly scope the claims to the new dataset.","section":"Sec. 4, Dataset and Evaluation Protocol"},{"comment":"The comparison against GaussianAvatars is based on the authors' own reimplementation (the text states the original method cannot faithfully track the full-body motions, so they implemented a version with constant head Gaussians). Because a reimplementation may not reflect the original method's performance, this comparison is not a verifiable substitute for running the published code. Please label this row as an ablation of constant head Gaussians rather than as a direct comparison, or provide a more faithful baseline.","section":"Sec. K.3, Tab. 4"}],"minor_comments":[{"comment":"The real-time claim is '35 fps on two A40 GPUs with one frame delay'; the abstract says 'over 30 fps' without this qualification. Please state the GPU count and latency explicitly in the abstract and in the runtime paragraph.","section":"Sec. 4, Implementation and Runtime"},{"comment":"The eyelid postprocessing relies on empirically chosen hyperparameters zeta = 0.75 and omega (the downscale factor). Please add a sensitivity analysis or a short ablation to show the results are not sensitive to these choices.","section":"Sec. 3.2 and Sec. F"},{"comment":"The main text says the system comprises '100 full-frame cameras' while the supplementary (Sec. H) says 'approximately 120 synced 25 fps cameras'; these numbers should be reconciled.","section":"Sec. 4 vs. Sec. H"},{"comment":"The explanation that DDC's higher novel-view PSNR is 'likely due to training on all frames rather than sampled ones' is speculative; if this is the cause, it can be tested by training EVA with a comparable frame schedule, and the result should be reported.","section":"Sec. 4.2"},{"comment":"The title and headers repeatedly display 'EV A' with a space; this should be 'EVA' consistently.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central idea is promising and the technical description is unusually detailed, but the missing validation of the expression tracking is a load-bearing gap: the paper's headline 'expressive control' claim depends on it. The lack of error bars and the reliance on the authors' own dataset further weaken the comparative claims. I would support a major revision requiring either independent psi validation or a clear reframing of the contribution as a controllable rendering system whose expression accuracy is demonstrated qualitatively, with the quantitative claims limited accordingly. A faithful public benchmark comparison would also materially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"EVA is a solid systems paper from a strong group, and the new bit is real: a two-layer avatar that combines a deformable DDC-style body template with a personalized FLAME head and two separate U-Nets that predict body and head Gaussians from skeletal pose and expression parameters independently. That disentanglement is not just cosmetic; the ablations show the body-conditioned and expression-conditioned variants behave differently, and the qualitative demos of changing expression without altering body appearance are convincing. The paper also ships a new multi-view dataset with deliberate facial expression coverage, which addresses a real gap. It runs at roughly 35 fps at 2K, and the detailed runtime breakdown is useful. The main soft spot is exactly what your stress-test flags: the expression parameters psi are tracked with a 2D-landmark and Chamfer pipeline, then used both as training input for the head appearance network and as the test-time control signal. There is no independent check on those parameters, no face mocap, no manual annotation on held-out frames, no comparison against a high-quality offline face fit from the 20 close-up cameras. If the tracked psi is biased, the network learns a mapping from that particular distribution rather than a generalizable expression control signal. This does not invalidate the headline rendering results, EVA beats ASH and DDC on LPIPS on their own dataset, but the strong claim of faithful expression control is not proven. It is a verification gap, not an observed failure, and the paper would be materially stronger with one held-out expression source or a user study. Other weaknesses are minor but worth naming: no code or data released, no error bars or per-subject variance in Tables 1/2/4, evaluation only on the authors' own capture with DynaCap excluded for a plausible reason, and full-frame metrics that dilute facial errors. The DEGAS comparison is skipped because code is unavailable; the supplementary does add AnimatableGaussians and a GaussianAvatars reimplementation, which is more effort than most papers make. Overall, the central claim that EVA outperforms ASH and DDC in rendering quality and expressiveness is plausible and internally supported. The architecture is clean, the ablation coverage is thorough, and the limitations section is honest, even naming the neck seam and cross-region lighting issues. The serious concerns are about verification boundaries, not about the core method being wrong. Recommendation: send it to peer review. A competent referee should push for independent validation of the expression tracking and for per-subject error bars, but the paper deserves their time and would likely come out solid after a revision. I would not cite it as proof of faithful expression control until that gap is closed, but I would cite it as a strong system integration with a clean disentanglement design.","headline":"Solid real-time full-body avatar system with a genuinely disentangled face/body Gaussian model, but the expressive-control claim hinges on an unvalidated closed-loop expression tracker.","tokens_in":840,"tokens_out":977,"would_cite":true,"duration_ms":21298,"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 two-layer avatar model separates facial expressions from body and hand control, and renders the result in real time.","keywords":["expressive virtual avatars","3D Gaussian splatting","multi-view reconstruction","human avatar animation","facial expression tracking","disentangled body-face control","real-time rendering","deformable human template"],"falsifier":"Render the same trained avatar from the same multi-view test frames while using FLAME parameters obtained from an independent, high-accuracy face measurement (for example, a calibrated facial capture rig or manually annotated landmarks) and compare the rendered face region to the recorded images. If the appearance network's outputs degrade when the independently measured parameters replace the paper's tracked ones, that indicates the learned face mapping has absorbed errors from the closed tracking loop rather than truly grounding expression control.","tokens_in":27737,"feed_emoji":"🎭","tokens_out":8218,"duration_ms":65895,"temperature":0.7,"pith_summary":"EVA is an actor-specific avatar system that takes multi-view video of a person and produces a drivable full-body model in which body motion, hand gestures, and facial expressions are controlled independently. The paper argues that this independence is what existing full-body avatar methods lack: they either tie facial appearance to body pose or require image-based facial drivers, so they cannot faithfully re-enact new expressions on a moving body. EVA's answer is to split the avatar into an expressive deformable template mesh and a decoupled 3D Gaussian appearance layer, with separate networks predicting body and face appearance. The authors report that this design beats the DDC and ASH baselines on perceptual quality (lower LPIPS) in novel-view, novel-pose, and novel-expression settings, while rendering at over 30 frames per second.","feed_headline":"Decoupling face and body yields real-time expressive avatar control","feed_subtitle":"Multi-view video trains a drivable human whose expressions and poses edit independently, beating DDC and ASH baselines.","key_machinery":"The load-bearing object is the expressive template mesh $\\boldsymbol{\\Phi}_{\\mathrm{mesh}}(\\boldsymbol{\\theta}^{\\triangleright}, \\boldsymbol{\\psi}) = C(\\boldsymbol{\\theta}^{\\triangleright}) \\bowtie H(\\boldsymbol{\\psi})$, a watertight union of a learned deformable body character $C$ and a personalized FLAME head $H$, stitched at the neck. This mesh is the shared geometric proxy for everything downstream: it is tracked to multi-view video via 2D facial landmarks and Chamfer losses, and it is rendered into UV-space normal and position maps that feed the two disentangled Gaussian predictors. The second central object is the appearance function $\\Phi_{\\mathrm{app}} = \\Phi^{\\mathrm{body}}_{\\mathrm{app}}(C^{\\triangleright}) \\cup \\Phi^{\\mathrm{head}}_{\\mathrm{app}}(H^{\\triangleright}(\\boldsymbol{\\psi}), \\hat{\\boldsymbol{\\vartheta}}^{\\triangleright})$, which emits body Gaussians and face Gaussians separately so that pose and expression can be edited without mutual interference.","core_discovery":"On its own terms, the paper claims that photorealistic, real-time avatars with full expression control are achievable by disentangling geometry and appearance along body/face lines rather than by improving a single unified model. Concretely, a FLAME-based personalized head is stitched onto a learned deformable character model, so one template mesh responds to both skeletal motion windows and facial expression parameters. The same split is repeated in appearance: two independent U-Nets predict 3D Gaussian parameters in UV space, one conditioned on body motion and one on facial expression, and their outputs are combined at splatting time. The empirical claim is that this explicit separation yields better rendering quality than DDC and ASH, and the ablations show that removing the separation or feeding body features into the head network hurts quality, indicating the model is learning the intended independent control.","pith_inferences":["A natural next test the paper does not run: re-train on a subject whose expressions are measured by an independent rig, then compare the recovered FLAME parameters; a mismatch would reveal how much of the apparent expressiveness comes from the appearance network absorbing tracking error.","The same disentanglement could be turned into a quantitative controllability metric, measuring how much changing expression parameters changes body-region pixels and vice versa, which the paper reports only qualitatively.","Because the dataset combines full-body motion, hand gestures, detailed expressions, and close-up face views, it could serve as a shared benchmark for future expressive-avatar work, though the paper does not frame it that way."],"forward_implications":["An actor's expression can be changed without retouching the body, because the head and body Gaussian parameters are produced by separate networks.","The same avatar can be driven by any source of FLAME expression parameters, including audio-based expression synthesis, rather than requiring front-facing images at inference.","Loose clothing remains dynamic: the body appearance network consumes motion windows of the deformed character mesh, so garment motion is not baked into a static template.","The reported runtimes put the full pipeline at roughly 35 fps for the appearance layer and 41 fps for the geometry layer, i.e., real-time 2K rendering on the tested hardware."],"supporting_citations":[{"why":"Supplies the deformable character model whose learned embedded-graph and displacement networks provide body and hand control and loose-cloth deformations; EVA extends it with a personalized head and refined motion.","marker":"Habermann et al. [2021]"},{"why":"Is the ASH baseline and the source of the UV-space 3D Gaussian prediction and warmup training that EVA adapts into its disentangled body and face networks.","marker":"Pang et al. [2024]"},{"why":"Provides the FLAME parametric head model that EVA fits to scans and tracks to obtain the facial expression parameters driving the head appearance.","marker":"Li et al. [2017]"},{"why":"Defines 3D Gaussian splatting, the rendering primitive whose parameters the appearance layer predicts in UV space.","marker":"Kerbl et al. [2023]"},{"why":"Supplies MediaPipe 2D facial landmarks that, together with FAN, are the observations used to refine head pose and track eyelids and expressions.","marker":"Lugaresi et al. [2019]"},{"why":"Supplies FAN 2D facial landmarks with confidence scores, used in motion optimization and expression tracking.","marker":"Bulat and Tzimiropoulos [2017]"},{"why":"Provides one captured subject and the practice of adding 3D-scan geometric supervision to the deformable character, which EVA adopts.","marker":"Shetty et al. [2024]"},{"why":"Provides the per-frame 3D surface reconstructions used inside the bidirectional Chamfer losses for motion optimization and character-deformation training.","marker":"Wang et al. [2023]"}],"fun_headline_variants":["EVA: Decouple face and body for real-time expressive avatars","Two-layer avatar model gives independent face and body control","Expressive avatars: split face and body, splat in real time","EVA: Multi-view video trains drivable avatar with distinct face and body","Disentangled face and body appearance for lifelike avatars"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the FLAME expression parameters recovered from 2D landmarks and Chamfer fitting are accurate enough to serve as the true control signal, with no independent face-rig or manual validation to catch systematic tracking errors.","fun_headline_variants_meta":{"raw":{"variants":["EVA: Decouple face and body for real-time expressive avatars","Two-layer avatar model gives independent face and body control","Expressive avatars: split face and body, splat in real time","EVA: Multi-view video trains drivable avatar with distinct face and body","Disentangled face and body appearance for lifelike avatars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1465,"prompt_tokens":993,"completion_tokens":472,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":609,"tokens_out":472,"duration_ms":4555,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:18:31.555650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the same trained avatar from the same multi-view test frames while using FLAME parameters obtained from an independent, high-accuracy face measurement (for example, a calibrated facial capture rig or manually annotated landmarks) and compare the rendered face region to the recorded images. If the appearance network's outputs degrade when the independently measured parameters replace the paper's tracked ones, that indicates the learned face mapping has absorbed errors from the closed tracking loop rather than truly grounding expression control.","supporting_citations":[],"review_version":1}