Datasets:
The dataset viewer is not available for this split.
Error code: JobManagerCrashedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
HoloFaceIllusion-Bench-EEG
A large-scale benchmark of holistic-face illusion stimuli for testing human-vs-DNN alignment on configural face processing and for paired EEG-decoder evaluation. Built entirely with classical CV (dlib landmarks + MediaPipe Face Mesh + InsightFace gender/age + OpenCV Poisson cloning + Reinhard LAB colour transfer) — no neural networks or generative AI are used to create any pixel.
Three paradigms are included:
| Paradigm | Cases | Conditions per case | Total images | Reference |
|---|---|---|---|---|
| Thatcher | 26,317 identities | 4 (V1/V2/V3/V4) | 105,268 | Thompson (1980) |
| Part-Whole | 188,988 (template × target × feature) | 4 (V1/V2/V3/V4) | 755,952 | Tanaka & Farah (1993); Tanaka et al. (2004) |
| Composite | 61,688 (template × donor) | 4 (V1/V2/V3/V4) | 246,752 | Young, Hellawell & Hay (1987); McKone et al. (2013) |
| Total | 276,993 cases | 1,107,972 images |
What a "case" means in each paradigm:
- Thatcher case = one FFHQ identity. Its 4 conditions are: V1 upright_normal, V2 upright_thatched (eyes + mouth rotated 180° in place), V3 inverted_normal, V4 inverted_thatched. (The Thatcher effect: V2 looks bizarre upright, but V4 looks nearly normal.)
- Part-Whole case = one (template identity, target identity, swapped feature) triple. Each template face is paired with multiple target identities, and one of three local features (eye / nose / mouth) is swapped in. The 4 conditions are: V1 whole-face target, V2 whole-face with foil's feature swapped, V3 feature alone on grey background, V4 foil's feature alone.
- Composite case = one (template, donor) pair. Template's top half is kept; donor's bottom half replaces template's bottom half. The 4 conditions are: V1 aligned upright (the canonical Young 1987 illusion — two halves fuse into a novel identity), V2 misaligned upright, V3 aligned inverted, V4 misaligned inverted. V2/V3/V4 are control conditions required by McKone et al. (2013) and Murphy, Gray & Cook (2017) to interpret V1 as a pure measure of holistic processing.
Examples
Thatcher illusion
Each identity is shown in 4 conditions: V1 upright_normal, V2 upright_thatched (eyes + mouth rotated 180° in place), V3 inverted_normal, V4 inverted_thatched. The Thatcher effect: V2 looks bizarre upright, but V4 looks nearly normal.
Part-Whole illusion
For each template face, one local feature (eye / nose / mouth) is replaced by a different identity's feature. V1 is the whole target face, V2 swaps in the foil's feature, V3/V4 isolate the same feature on a grey background. The Part-Whole effect: telling V1 from V2 (whole-face context) is easier than telling V3 from V4 (feature alone).
Composite illusion
For each template face T, the top half is kept and the bottom half is replaced by a different identity D's bottom half (similarity-aligned, cheek-width matched, L-channel ring-shifted, Poisson NORMAL_CLONE seam). V1 aligned upright is the canonical Young 1987 illusion (the two halves fuse into a percept of a novel identity); V2 misaligned upright + V3 aligned inverted + V4 misaligned inverted are control conditions required to interpret V1 as a pure measure of holistic processing (McKone et al. 2013, Murphy, Gray & Cook 2017). Donor pool was filtered for: HARD same-gender (InsightFace), age within 15 yrs, skin-LAB Δ ≤ (10, 8, 8), dlib face-embedding L2 ∈ [0.55, 0.85], mouth on midline (≤ 3% face-width offset), cheek width at cut ± 10%, cut-to-chin distance ± 10%, head yaw |°| ≤ 15.
Download
from datasets import load_dataset
# Thatcher — 26,317 identities × 4 conditions = 105,268 images
thatcher = load_dataset(
"Enhui-1/HoloFaceIllusion-Bench-EEG",
name="thatcher",
split="train",
streaming=True,
)
# Part-Whole — 188,988 (template, target, feature) cases × 4 conditions = 755,952 images
partwhole = load_dataset(
"Enhui-1/HoloFaceIllusion-Bench-EEG",
name="partwhole",
split="train",
streaming=True,
)
# Composite — 61,688 (template, donor) cases × 4 conditions = 246,752 images
composite = load_dataset(
"Enhui-1/HoloFaceIllusion-Bench-EEG",
name="composite",
split="train",
streaming=True,
)
For full local copies:
hf download Enhui-1/HoloFaceIllusion-Bench-EEG --repo-type dataset --local-dir ./HoloFaceIllusion-Bench-EEG
Tar layout (webdataset):
- Thatcher (
v1.1/NNNNN.tar):<id>.v{1,2,3,4}_*.png+<id>.landmarks.json, sample key =<id> - Part-Whole (
v1.4/partwhole_NNN.tar):<id>__t<NN>_{eye,nose,mouth}.v{1,2,3,4}.png+.json, sample key =<id>__t<NN>_<feature> - Composite (
v2.0/composite_NNN.tar):<id>__d<MMMMM>.v{1,2,3,4}.png+.json, sample key =<id>__d<MMMMM>
Per-case indexes (one row per case):
- Thatcher:
manifest.csv - Part-Whole:
v1.4/manifest.csv(188,988 rows) - Composite:
v2.0/manifest.csv(61,688 rows; columns: template, donor, sample_key, tar, y_cut, face_w, mis_dx_px, ssim_v1v2, seam_disc, lab_delta_cut, template_yaw_deg, donor_yaw_deg)
License: CC BY-NC-SA 4.0 (inherited from FFHQ; academic / non-commercial use).
Version notes
- v1.1 (Thatcher, 2026-05-27 repack): replaces the original v1.0 root-level tars.
The v1.0 layout used
NNNNN/V{1,2,3,4}_*.pngsubdirectories per identity, which made the HuggingFace data viewer fail (SplitsNotFoundError: webdataset requires files sharing the same sample key with different extensions). v1.1 flattens the layout toNNNNN.v{1,2,3,4}_*.pngso all 5 files per identity (4 PNG + 1 JSON) group correctly into one sample. Pixel content is identical.
Citation
@dataset{li2026holofaceillusionbench,
title = {HoloFaceIllusion-Bench-EEG: a classical-CV benchmark of Thatcher,
Part-Whole, and Composite face illusion stimuli},
author = {En Hui Li},
year = {2026},
howpublished = {Hugging Face Datasets},
url = {https://huggingface.co/datasets/Enhui-1/HoloFaceIllusion-Bench-EEG},
note = {XJTU AI EXP.LAB. CC BY-NC-SA 4.0; derived from FFHQ (Karras et al. 2019).},
}
Please also cite the upstream paradigm and dataset papers:
@inproceedings{karras2019stylebased,
title = {A Style-Based Generator Architecture for Generative Adversarial Networks},
author = {Karras, Tero and Laine, Samuli and Aila, Timo},
booktitle = {CVPR},
year = {2019},
}
@article{thompson1980thatcher,
title = {Margaret {T}hatcher: a new illusion},
author = {Thompson, Peter},
journal = {Perception},
volume = {9}, pages = {483--484}, year = {1980},
}
@article{young1987configurational,
title = {Configurational information in face perception},
author = {Young, Andrew W. and Hellawell, Deborah and Hay, Dennis C.},
journal = {Perception},
volume = {16}, pages = {747--759}, year = {1987},
}
@article{tanaka1993parts,
title = {Parts and wholes in face recognition},
author = {Tanaka, James W. and Farah, Martha J.},
journal = {Quarterly Journal of Experimental Psychology},
volume = {46}, pages = {225--245}, year = {1993},
}
@article{tanaka2004holistic,
title = {A holistic account of the own-race effect in face recognition:
evidence from a cross-cultural study},
author = {Tanaka, James W. and Kiefer, Markus and Bukach, Cindy M.},
journal = {Cognition},
volume = {93}, pages = {B1--B9}, year = {2004},
}
@article{mckone2013importance,
title = {Importance of the inverted control in measuring holistic face
processing with the composite effect and part-whole effect},
author = {McKone, Elinor and Davies, Anne Aimola and others},
journal = {Frontiers in Psychology},
volume = {4}, pages = {33}, year = {2013},
}
@article{murphy2017composite,
title = {The composite face illusion},
author = {Murphy, Jennifer and Gray, Katie L. H. and Cook, Richard},
journal = {Psychonomic Bulletin {\&} Review},
volume = {24}, pages = {245--261}, year = {2017},
}
- Downloads last month
- 99





