EEGFaceSem

EEG Dataset for Semantic Visual Response

Paper · Project page

Accepted at NeurIPS 2026, Evaluations and Datasets Track.

Installation

git clone https://huggingface.co/datasets/yefllower/EEGFaceSem
cd EEGFaceSem
pip install -e .

Quick Start

import EEGFaceSem

# Auto-downloads data when loading
X, Y, ids = EEGFaceSem.load_data(task='female')
print(f"X: {X.shape}, Y: {Y.shape}")  # X: (n_trials, 32, 1101), Y: (n_trials,)

# Or download specific subjects only (~300MB per subject)
# EEGFaceSem.download(subjects=[1])

Image Generation

# Load latent vectors and generate face image
image_ids, latents, id_to_idx = EEGFaceSem.load_latent()
EEGFaceSem.generate(latents[0:1])[0].save("face.png")

Dataset Info

Metric	Value
Subjects	30
Total epochs	64,124
EEG channels	32
Sampling rate	1000 Hz
Epoch window	[-0.2, 0.9]s

8 Tasks

Task ID	Task Name
0	female
1	male
2	blond
3	darkhaired
4	smiles
5	nosmile
6	old
7	young

API Reference

Data Loading

# Download data
EEGFaceSem.download(data_type="processed")  # or "raw", "both"
EEGFaceSem.download(subjects=[1, 2, 3])     # specific subjects only

# Load data
X, Y, ids = EEGFaceSem.load_data(task='female')

Splitting

# Random split
(X_train, Y_train), (X_test, Y_test) = EEGFaceSem.split_random(X, Y, test_size=0.2)

# Leave-one-subject-out
(X_train, Y_train), (X_test, Y_test) = EEGFaceSem.split_by_subject(X, Y, ids, test_subject=1)

Benchmarking

EEGFaceSem.benchmark(
    model='LDA',        # LDA, LR, MLP, EEGNet, EEGPT
    task_id=0,          # 0-7 or -1 for all
    strategy='single_subject',
)

Image Generation

# Load latent vectors
image_ids, latents, id_to_idx = EEGFaceSem.load_latent()

# Generate from latent
images = EEGFaceSem.generate(latents[0:1])
images[0].save("face.png")

# Generate from specific image ID in data
img_id = int(ids[0, 4])
EEGFaceSem.generate(latents[id_to_idx[img_id]:id_to_idx[img_id]+1])[0].save(f"face_{img_id}.png")

License

The dataset is released under CC BY 4.0. The accompanying benchmark code is released under the Apache License 2.0.

Citation

@misc{ma2026eegfacesem,
  title = {EEGFaceSem: An EEG Benchmark with Paired Generative Latents for Semantic Visual Modeling},
  author = {Jun Ma and Michiel M. Spapé and Keith M. Davis III and Tuukka Ruotsalo},
  year = {2026},
  eprint = {2611.10781},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2611.10781}
}