ECCV 2026 Medical Imaging 3D Vision EEG decoding brain-computer interface facial expression synthesis 3D Gaussian Splatting dense position map

Mind-to-Face: Neural-Driven Photorealistic Avatar Synthesis via EEG Decoding¶

Conference: ECCV 2026
Paper: ECCV 2026
Area: Medical Imaging / 3D Vision
Keywords: EEG decoding, brain-computer interface, facial expression synthesis, 3D Gaussian Splatting, dense position map

TL;DR¶

Mind-to-Face is the first framework to decode non-invasive 16-channel EEG directly into dense 256×256 3D position maps (~65k vertices) that drive 3D Gaussian Splatting bound to mesh faces, rendering photorealistic facial avatars that preserve both subject identity and emotional intensity; enabling it required a purpose-built dual-modality capture system with frame-accurate synchronization between EEG and 16 multi-view high-speed cameras.

Background & Motivation¶

Today's expression and avatar systems are built almost entirely on visual cues — cameras, motion capture, speech analysis. They collapse as a whole in one overlooked setting: when the face is occluded by a head-mounted display, or when an emotion is experienced internally but never externally expressed. Yet emotion and cognition originate in neural activity that precedes facial movement. Meanwhile three research lines that ought to connect remain separate: vision models reconstruct external geometry, affective computing classifies EEG into coarse labels such as happy or sad, and brain-computer interfaces chase symbolic or motor signals. The link from neural state to visual expression is missing.

EEG is the most realistic entry point: non-invasive, wearable, and high in temporal resolution, with substantial evidence of stable correlations between EEG and emotional states as well as symbolic facial expressions. Even so, EEG is rarely used as a driver for visual synthesis, and existing learning-based EEG-to-expression work mostly reduces to discrete label classification. The bottleneck is not merely that labels are coarse — discrete categories simply cannot carry the continuous, fine-grained dynamics a photorealistic avatar needs. Direct high-fidelity reconstruction from EEG is also hard: few channels, low signal-to-noise ratio, and highly individualized neural responses, since identical stimuli produce distinct EEG signatures across subjects. That last fact is why scalable cross-subject modeling is not yet realistic here.

This paper therefore narrows the problem to subject-specific continuous expression reconstruction under a given stimulus, and addresses two concrete obstacles: EEG and facial geometry must be aligned at frame level, otherwise the supervision itself is wrong; and the mapping from low-dimensional EEG to high-dimensional facial geometry should not pass through an expression-coefficient bottleneck. Core idea: use a dual-modality rig synchronized by linear timecode to obtain paired (EEG window, multi-view face video) samples, take dense 3D position maps reconstructed by multi-view photogrammetry as the decoding target, encode EEG with a CNN-Transformer, regress the position maps with a randomly initialized VAE decoder, and render photorealistic avatars through a modified 3D Gaussian Splatting pipeline — turning neural-signal-to-visual-expression into an end-to-end, continuously supervised mapping.

Method¶

Overall Architecture¶

The pipeline takes a 16-channel, 125Hz raw EEG stream together with synchronized multi-view face video, and outputs a view-consistent photorealistic facial avatar carrying emotional expression. Four steps sit in between: the EEG is band-pass filtered at 4–40Hz, z-score normalized per channel, and cut into sliding windows of 375 samples (exactly 3 seconds at 125Hz); the multi-view video is reconstructed by photogrammetry into topologically consistent 3D face meshes, rigidly aligned into a canonical frame and converted into 256×256 UV position maps that serve as supervision; a CNN-Transformer encoder compresses each EEG window into a latent vector, and a VAE decoder expands it into a dense position map; finally, vertices are sampled from the position map to form a mesh, and every triangle drives a bound 3D Gaussian for differentiable rendering. Only EEG is an input at inference time; video and meshes appear only during training and evaluation.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["16-ch EEG<br/>+ multi-view video"] --> B["Dual-modality capture & paired supervision<br/>frame-level sync → position maps"]
    B --> C["CNN-Transformer EEG encoder<br/>spatiotemporal conv + self-attention"]
    C --> D["Dense position-map decoding<br/>VAE → dense position map"]
    D --> E["Gaussian binding & rendering<br/>mesh-driven 3D Gaussian Splatting"]
    E --> F["EEG-driven photorealistic avatar"]

Key Designs¶

1. Dual-modality synchronized capture and paired supervision: aligning EEG with facial geometry at frame level

Supervised learning between EEG and facial geometry requires the two streams to be truly aligned in time, and this is the most easily underestimated engineering difficulty: the EEG headset and the camera array are driven by different hardware clocks, and a drift of a few tens of milliseconds is enough to match "this stretch of brain activity" to the wrong frame. The paper moves the whole acquisition into a Light Stage. The subject faces a display playing emotion-eliciting film clips, surrounded by 16 cinema-grade global-shutter cameras (up to 8K, 120fps) evenly spaced at 15° intervals for dense view coverage, with the cameras synchronized among themselves by genlock and timecode generators. EEG is recorded with OpenBCI's Cyton-Daisy at 16 channels and 125Hz, with electrodes placed on the international 10–20 system and conductive gel keeping impedance below 500kΩ. The critical step is a linear signal generated by an Arduino Teensy that pins the EEG headset and the camera array onto a single timeline, yielding frame-accurate dual-modality data. Stimuli are drawn from OpenLAV and the emotion-eliciting clips curated by Coan and Allen, grouped into five categories — neutral, disgust, funny, angry, sad — each forming one trial, with a 30-second rest in between to prevent emotional carryover; a separate mixed-emotion sequence from EMOSTIM never enters training and serves as a leave-one-trial-out cross-trial test. Each subject completes five emotional trials plus one test sequence (the introduction counts six trials in total), for roughly 25–30 minutes of recording per session.

The supervision side needs its own design: a 3D face mesh must be reconstructed for every video frame. The pipeline first solves for the subject's identity on a neutral frame using the 3DMM from ICT-FaceModel, fits the identity template mesh to the multi-view photogrammetry observations, maintains consistent topology via Laplacian deformation, and refines temporally with optical flow. Each mesh then undergoes rigid Procrustes alignment to remove head pose, bringing the whole sequence into one canonical coordinate frame — a worthwhile step, because it leaves only non-rigid expression deformation in the decoder's supervision target, effectively handing the network a simpler problem for free. The aligned mesh is converted into a position map \(P(t) \in \mathbb{R}^{256\times256\times3}\), where each pixel encodes the \((x,y,z)\) coordinates of the facial surface in UV space, and paired with the corresponding EEG window. Building this rig is justified because the closest public dataset, MAHNOB-HCI, offers only one color plus five monochrome camera streams, whose resolution and color format cannot support 3D face reconstruction and therefore cannot carry this supervision chain.

2. CNN-Transformer EEG encoder: extracting expression-synchronized representations from a low-dimensional noisy signal

EEG has only 16 channels at 125Hz and is contaminated by muscle activity, eye movement, and power-line noise, so regressing facial geometry from it directly is close to hopeless. The paper reshapes an EEG window \(E(t) \in \mathbb{R}^{W\times C}\) (\(W=375\), \(C=16\)) into a single-channel \(1\times C\times W\) tensor and treats it as a spatiotemporal image to be convolved: the first layer convolves along the temporal axis (40 output channels, kernel 40) to capture short-range waveform patterns, and the second convolves across all 16 electrodes (kernel 16) to model spatial correlations between channels — the standard recipe of CNN-based EEG decoders, replacing handcrafted band power or CSP features with learned convolutions. The convolutional branch is followed by batch normalization, ELU, and average pooling, which shrinks the temporal dimension from 375 to about 19 feature patches, discarding redundancy while preserving slowly varying emotion-relevant content; a 1×1 convolution then projects to an embedding dimension of \(E=40\), forming a token sequence of 19 tokens.

Six Transformer encoder layers follow: each uses pre-normalization, 10-head multi-head self-attention, and a feed-forward network that expands the embedding dimension by a factor of 4 with GELU and dropout (\(p=0.5\)), with residual connections on both sub-layers. Convolution handles local spatiotemporal structure while self-attention handles long-range dependencies across the entire 3-second window — neither suffices alone. Only by capturing both the instantaneous spatial layout across electrodes and the evolution of emotion within one window can the model extract the component that is synchronized with facial expression from a low-SNR signal. It should be noted that this encoder is a mature backbone inspired by EEG-Conformer; the contribution here is not the architecture itself but the fact that it is attached to a dense geometric target.

3. Dense position-map decoding: replacing 52 blendshape coefficients with ~65k vertices

The latent vector \(z(t)\) from the encoder is projected by a sequence of fully connected layers into a \(4\times8\times8\) latent feature map, which a VAE decoder expands into a coarse 3-channel output; two transposed convolution layers then upsample it into a \(256\times256\times3\) position map whose pixels likewise encode the surface \((x,y,z)\). This decoder is randomly initialized and trained jointly with the EEG encoder — the only training signal is the MSE and smoothing regularization on the position map, with no intermediate supervision from images or FLAME coefficients.

The authors give two concrete reasons for predicting position maps rather than expression coefficients. The first is expressiveness: mapping EEG to a fixed set of 52 blendshape coefficients stacks a second 52-dimensional bottleneck on top of an already narrow EEG bottleneck, biasing the model toward averaged-out predictions and smoothing away subtle expressions; the approximately 65k vertices of a position map (against roughly 5000 vertices in FLAME) can carry mid-frequency geometric detail such as wrinkles. The second is trainability: a position map is a regular 2D grid parameterized in UV space, so standard 2D convolutional networks and image decoders apply directly, with no need for operators specialized to irregular meshes. In addition, a binary mask restricts supervision to the inner-face region so the decoder focuses on expression deformation rather than rigid head motion, and a self-supervised Laplacian regularizer improves surface smoothness and continuity. The ablation (Tab. 1) confirms the practical benefit of this choice.

4. FLAME-free tracking with Gaussian binding: decoupling geometry from appearance

The rendering stage follows the 3D Gaussian framework of GaussianAvatars while replacing the source of geometry entirely. Standard GaussianAvatars relies on the VHAP library for landmark detection and derives face geometry from FLAME parameters; in this capture setting the subject must wear an EEG headcap, which confuses landmark detection, and the extreme expressions induced by emotional stimuli exceed what FLAME parameters can express. Together these cause tracking errors and occasionally total geometric collapse. The paper therefore uses the topologically consistent mesh produced by its own robust tracking pipeline as the geometric hub: at training time it no longer fits FLAME parameters but takes geometry directly from the tracked mesh, letting the Gaussians fit shapes that stay closer to real expressions.

The binding strategy is the heart of this rendering scheme. At initialization each triangular face is assigned one Gaussian, and the Gaussian's local parameters — rotation \(r\), position \(\mu\), and scale \(s\) relative to the face center, alongside third-order spherical-harmonic color — remain learnable quantities; during rendering they are transformed into global coordinates by the face's own rotation \(R\), translation \(T\), and scale \(k\):

\[r' = Rr,\quad \mu' = kR\mu + T,\quad s' = ks\]

In other words, the local Gaussian parameters are learned once in the canonical frame, and animation only updates the global transform of the triangle each Gaussian is bound to. Appearance and geometry are thereby decoupled: any new position map only needs to be sampled into a mesh, and the same set of Gaussians immediately follows it and renders, with no retraining and with view consistency guaranteed by construction. Each training iteration samples a triplet of (mesh, camera, image) uniformly at random, renders, and back-propagates; the objective augments L1 and D-SSIM with position and scale regularizers that keep Gaussians from drifting away from their bound face centers and from growing unbounded in scale.

A Worked Example¶

Consider one 3-second EEG segment at inference time. The raw signal passes through a 6th-order Butterworth band-pass filter (4–40Hz) that removes drift and high-frequency noise, then per-channel z-score normalization using training-set statistics, and is cut into a \(1\times16\times375\) window. A temporal convolution (40 channels, kernel 40) and a spatial convolution (kernel 16) apply in sequence; after batch normalization, ELU, and average pooling the temporal dimension becomes 19 patches, and a 1×1 convolution projects them into 40 dimensions, giving a 19-token sequence. Six Transformer layers (10 heads, FFN×4, dropout 0.5) output the 40-dimensional \(z(t)\). Fully connected layers project \(z(t)\) into a \(4\times8\times8\) latent feature map, the VAE decoder emits a coarse 3-channel map, and two transposed convolutions upsample it to \(256\times256\times3\) — the predicted position map \(\hat P(t)\). Vertex coordinates are sampled from it according to the template UV layout, yielding a mesh \(\hat M(t)\) of about 65k vertices; each triangle on that mesh carries its own Gaussian, which is globally transformed and splatted to produce a photorealistic, view-consistent facial image for that instant. The EEG windows overlap, so adjacent position maps transition smoothly and the sequence is temporally coherent without any additional smoothing.

Loss & Training¶

The position-map loss combines a supervised term and a smoothing term, where \(\mathcal{L}_{\mathrm{rec}}\) is the per-pixel MSE inside the mask and \(\mathcal{L}_{\mathrm{smooth}}\) is the Laplacian smoothing loss applied to the predicted position map:

\[\mathcal{L}_{\mathrm{pm}} = \lambda_{\mathrm{rec}}\mathcal{L}_{\mathrm{rec}} + \lambda_{\mathrm{smooth}}\mathcal{L}_{\mathrm{smooth}}\]

⚠️ The formula is corrupted in the extracted PDF text; it is reconstructed here as a weighted sum of the two terms as described in the prose, and the exact coefficients should be checked against the original paper.

The 3DGS training loss augments the reconstruction terms with position and scale regularizers, where \(\mathcal{L}_{\mathrm{pos}} = \|\max(\mu, \epsilon_{\mathrm{pos}})\|^2\) constrains each Gaussian's offset from its bound face center and \(\mathcal{L}_{\mathrm{scale}} = \|\max(s, \epsilon_{\mathrm{scale}})\|^2\) constrains its scale:

\[\mathcal{L} = (1-\lambda)\mathcal{L}_1 + \lambda\mathcal{L}_{\mathrm{D\text{-}SSIM}} + \lambda_{\mathrm{pos}}\mathcal{L}_{\mathrm{pos}} + \lambda_{\mathrm{scale}}\mathcal{L}_{\mathrm{scale}}\]

⚠️ The attribution of \(\lambda\) versus \((1-\lambda)\) is unclear in the extracted text; refer to the original paper. The remaining 3DGS hyper-parameters follow GaussianAvatars. For data splitting, the final 15% of frames in each emotional trial are held out for testing and the rest used for training (following the standard signal-processing practice of minimizing statistical leakage), while the EMOSTIM sequence never enters training and is used solely to evaluate cross-trial generalization.

Key Experimental Results¶

Main Results¶

The authors state that this is the first work to decode EEG into dense facial geometry, so no directly comparable baseline exists; the quantitative evaluation in the main text therefore centers on a comparison of representations: predicting position maps versus FLAME blendshapes within the same EEG decoding framework, measured by point-to-surface distance of the driven avatar geometry against photogrammetric ground truth, over the masked facial region only, in millimeters.

Trial	PosMap·S1 mean	PosMap·S1 <1mm	PosMap·S2 mean	PosMap·S2 <1mm	Blendshape·S1 mean	Blendshape·S1 <1mm	Blendshape·S2 mean	Blendshape·S2 <1mm
Angry	0.4982	87.37%	0.5534	84.28%	1.0792	60.76%	1.6492	42.79%
Disgust	1.0238	61.89%	0.6347	80.57%	1.7680	42.03%	1.3410	60.35%
Funny	0.9614	69.12%	0.7432	78.24%	1.3385	56.56%	1.6195	48.93%
Neutral	0.4028	92.85%	0.3245	95.84%	0.3270	94.99%	1.3065	52.91%
Sad	0.3571	94.60%	0.3112	96.60%	0.5391	84.95%	1.3134	56.22%
EMOSTIM	0.5158	86.76%	0.7077	78.88%	0.6838	78.78%	1.5078	52.22%

Means are in millimeters. The EMOSTIM row is the leave-one-trial-out cross-trial generalization result: neither its stimulus content nor its emotion category participated in training. The original paper also reports the percentage of vertices below 3 mm (generally ≥92.8% for the position-map-driven avatar), omitted here; this is the only quantitative table in the main text, with image-level metrics (FID, identity similarity, etc.) deferred to the supplementary materials.

Ablation Study¶

Because the paper's ablations are predominantly qualitative, the table below gathers both of them: the numeric row is aggregated over the trials of the table above by this note (the original paper does not report an overall average), and the non-numeric rows are honestly marked as qualitative evidence.

Ablation	Config	Geometric error (mean over 6 trials, mm)	<1mm vertex ratio (mean)	Evidence / Note
Geometric representation	Dense position map (S1 / S2)	0.627 / 0.546	82.1% / 85.7%	Aggregated from the table above (computed in this note)
Geometric representation	FLAME blendshape (S1 / S2)	0.956 / 1.456	69.7% / 52.2%	Aggregated from the table above (computed in this note)
Tracking pipeline	Ours (robust tracking)	—	—	Fig. 6, qualitative: stable under headcap occlusion and extreme expressions
Tracking pipeline	GaussianAvatars' VHAP + FLAME	—	—	Fig. 6, qualitative: wrong expressions, even collapsed geometry (red cross in the paper)
Mouth enhancement	GFP-GAN post-processing (mouth interior only)	—	—	Authors state it touches only the mouth interior and does not alter the decoded emotion
Supervision region	Inner-face mask + Laplacian smoothing	—	—	Described in the paper (focus on deformation, smoother surfaces); not separately quantified

Cross-trial generalization can be read as a third ablation: on the never-seen mixed-emotion EMOSTIM sequence the position-map-driven avatar still reaches 0.5158 / 0.7077 mm mean error, indicating that performance does not come from memorizing the emotion categories of the training clips.

Key Findings¶

The dense position map beats blendshapes almost everywhere: apart from Subject 1's Neutral trial (0.4028 vs 0.3270, marginally favoring blendshapes), the position-map-driven geometry has lower error in every trial, and for Subject 2 it is better across all six. Aggregated mean error drops from 0.956 / 1.456 mm to 0.627 / 0.546 mm, and the mean share of vertices below 1 mm rises from 69.7% / 52.2% to 82.1% / 85.7%.
The gap concentrates in strongly expressive trials: on Disgust, Subject 2 shows 0.6347 vs 1.3410 mm with 80.57% vs 60.35% of vertices below 1 mm, and Subject 1's Disgust is 1.0238 vs 1.7680 mm. This matches the authors' explanation that the 52-dimensional blendshape basis has limited expressive range and visibly shrinks amplitude on large, complex deformations; the fact that Subject 1's Disgust error exceeds their Angry and Funny errors also suggests that trial carried either stronger deformation or a harder signal to decode.
Smooth, low-amplitude expressions are easiest: Neutral and Sad yield the lowest errors for both subjects (0.31–0.40 mm). The weaker the expression and the closer to static, the simpler the EEG-to-geometry correspondence.
The tracking pipeline is an invisible prerequisite for feasibility: the authors note that GaussianAvatars' default tracking renders comparably to theirs in occlusion-free scenes, but once the setting involves an EEG headcap plus extreme expressions it mis-tracks or collapses outright. The robust tracking is thus not there to chase a metric — it is what makes the task viable under this capture setup.
Qualitative results (Fig. 5) show the EEG-driven avatar reproducing emotions (a frown for sad clips, smiles for funny clips, brow and nose tightening for disgust) while preserving subject-specific intensity — for instance, how broadly each subject smiles — which indicates subject-customized expression rather than coarse categorization.
The mouth interior is filled in by pretrained GFP-GAN post-processing, an acknowledgement of a known weakness of 3DGS avatars in modeling the oral cavity; the authors stress that the enhancement is strictly confined to the mouth interior and does not tamper with the decoded emotion.

Highlights & Insights¶

Replacing "EEG to discrete labels" with "EEG to UV-parameterized dense position maps": a position map is simultaneously continuous geometry and a regular 2D grid, so it can reuse standard image decoders (VAE plus transposed convolutions) and a mature 3DGS renderer, sidestepping the two-stage dimensionality reduction of "EEG → expression coefficients → geometry." This interface design is what makes the two ends connect at all, and it is the most transferable idea in the paper.
Procrustes alignment as a free difficulty reduction: removing rigid head motion from the supervision signal leaves only non-rigid deformation for the decoder to learn. For an input with an intrinsically poor signal-to-noise ratio, pre-processing the target space this way buys more than adding capacity to the network.
A binding strategy that decouples geometry from appearance: local Gaussian parameters are learned once in the canonical frame, and animation only updates the global transform of the bound face, so one set of Gaussians can follow any new position map, with view consistency guaranteed and no per-frame retraining.
The dataset and its synchronization scheme are themselves a contribution: EEG frame-aligned with 16 8K cameras in a layout that supports photogrammetric reconstruction is a configuration that did not previously exist, and it turns "neural signal ↔ high-fidelity facial geometry" paired supervision into something obtainable. The authors promise to release it.
Transferable directions: the chain of "low-dimensional noisy biosignal → dense canonical-space representation → differentiable rendering" can move wholesale to other channels, for example EMG, eye tracking, or fNIRS driving the same kind of position map, or a SMPL-X dense position map for neural-signal-driven full-body pose and gesture synthesis. The position map as a general interface should be reusable across cross-modal generation tasks.

Limitations & Future Work¶

Limitations the authors admit: data comes from only two subjects with six trials each (five emotional trials plus one test sequence) and about 25–30 minutes of recording per session, and the work explicitly avoids cross-subject modeling in favor of subject-specific expression control. Future work plans more subjects and more diverse stimuli, plus stimulus context during decoding to disentangle perceptual from affective signals.
No direct baseline: the authors state there is nothing comparable to benchmark against, and the only quantitative comparison in the paper is the internal position-map-versus-blendshape ablation. That leaves open how much the specific encoder-plus-VAE chain buys over a simpler EEG regression baseline, for example regressing blendshapes directly with EEG-Conformer.
Narrow evaluation: the quantitative results report only geometric point-to-surface error, with image-level quality metrics (FID, identity similarity, LPIPS, landmark error) pushed to the supplementary materials — yet for a system that renders photorealistic avatars, image quality is exactly what readers care about most.
Weak emotional conditioning: each trial corresponds to a single emotion and there are no per-frame emotion annotations, so the model could be learning "the average expressive tendency of this trial" rather than frame-by-frame emotional dynamics. The EMOSTIM cross-trial test demonstrates generalization exists but cannot rule this risk out.
Physical limits at the input: consumer-grade hardware with only 16 channels at 125Hz, and a 4–40Hz band-pass that excludes the gamma band, where emotion-relevant high-frequency content is often thought to live. In addition, the position maps used as supervision come from photogrammetric reconstruction, whose own error becomes a lower bound on the whole system.
The mouth interior needs an external GFP-GAN pass, showing that the geometry-plus-3DGS route remains unreliable inside the oral cavity — which is precisely the content most needed in the headset-occlusion scenario that motivates the work.
Concrete improvements: add subject embeddings or lightweight adapters for cross-subject transfer; use contrastive learning to explicitly align the EEG representation with facial action units, giving the latent space an interpretable middle layer; add an auxiliary emotion-classification loss to verify that affective information genuinely enters the predicted geometry.

vs EEG-based expression classification (e.g. CNN+GA decoding for a facial-expression BCI): they predict discrete expression categories from EEG, so the output is a label; this paper outputs continuous dense 3D geometry that is subsequently rendered into photorealistic images. The difference is not model size but the target representation — a label space cannot carry the fine-grained dynamics a photorealistic avatar needs, which is precisely the starting point of this work.
vs multi-modal affective datasets such as MAHNOB-HCI: they also record facial video synchronized with EEG, but MAHNOB-HCI is one color plus five monochrome streams at low resolution, supporting only 2D analysis and not photogrammetric 3D facial reconstruction. This paper's capture configuration is the first to deliver high-fidelity 3D geometry synchronized with EEG.
vs GaussianAvatars: they drive Gaussians bound to a mesh with FLAME expression coefficients and are optimal in the common case of mild expressions and an unoccluded face; this paper swaps the geometric hub for its own densely tracked position map, giving up the editable semantics of a parametric model in exchange for stability under headcap occlusion and extreme expressions.
vs NeRF-family head avatars (dynamic NeRF, IMavatar, Neural Head Avatars): these also reconstruct photorealistic heads, but 3DGS is chosen here because its explicit scene representation makes geometry controllable and rendering efficient — well suited to a task where geometry is driven by an external signal and precise control is required.
vs speech- or audio-driven avatars: speech is an external behavioral signal and remains an indirect observation of facial motion; EEG is an internal neural signal, and when the face is fully occluded by a headset and external cues fail entirely, it is the only potentially available channel. That is exactly the value of this paper's setting.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ The first work to decode non-invasive EEG into dense 3D facial geometry and render photorealistic avatars; both the problem setting and the interface design are new.
Experimental Thoroughness: ⭐⭐ Only two subjects, no comparable baseline, a single quantitative geometric metric, and all image-quality metrics deferred to the supplementary materials.
Writing Quality: ⭐⭐⭐⭐ The pipeline and the rationale for each design are explained clearly, and the subject-specific framing plus the absence of baselines are handled honestly; the formula extraction in the available version is corrupted and the quantitative table is thin.
Value: ⭐⭐⭐⭐ The synchronized capture pipeline and the "neural signal → dense position map → 3DGS" route have lasting value, and the promised dataset would be a scarce resource for this direction; but the system remains far from a usable occlusion-scenario avatar driver.