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. Constructed entirely with pure classical CV (dlib + OpenCV Poisson + Reinhard LAB) — no neural networks, no generative AI used to create any pixel.

Release	Paradigm	Identities/Templates	Images	Size
v1.0 (Thatcher)	Thompson 1980 eye-only	26,317 identities	105,268 PNG	134 GB
v1.2 (Part-Whole)	Tanaka 2004 template-based	955 templates × 9 targets	103,140 PNG	72.7 GB
v1.1 (Composite)	Young 1987	—	planned	—

Both subdatasets share the same FFHQ-1024 source (Karras 2019) and the same classical-CV construction philosophy. They are organized as separate config_names under one HF repo.

---

Quick start

from datasets import load_dataset

# v1.0 (Thatcher)
thatcher = load_dataset("Enhui-1/HoloFaceIllusion-Bench-EEG",
                         name="thatcher_v1.0", split="train", streaming=True)

# v1.2 (Part-Whole) — different schema (sample = template × target × feature)
partwhole = load_dataset("Enhui-1/HoloFaceIllusion-Bench-EEG",
                          name="partwhole_v1.2", split="train", streaming=True)

For per-config details see:

• v1.0 (Thatcher): docstring at the top of the 00000.tar directory + this README's Thatcher section below
• v1.2 (Part-Whole): v1.2/README.md for full schema, paradigm, citations

---

v1.0 — Thatcher illusion (Thompson 1980)

Paradigm: For each face, rotate eyes + mouth by 180° in place (eye-only variant: brows stay still per Thompson canonical). Upright thatcherized face appears bizarre; inverted version appears almost normal.

Identities	26,317 (FFHQ that pass dlib filter — 37.6% pass rate)
Conditions per identity	4 (V1 upright_normal, V2 upright_thatched, V3 inverted_normal, V4 inverted_thatched)
Tar shards	70 (~1.9 GB each) named 00000.tar ... 69000.tar
Pixel-baseline ISI	1.000 (V3 = rot180(V1), V4 = rot180(V2) exactly)

Tar layout (webdataset, one sample per identity):

ffhq_NNNNN/V1_upright_normal.png
ffhq_NNNNN/V2_upright_thatched.png
ffhq_NNNNN/V3_inverted_normal.png
ffhq_NNNNN/V4_inverted_thatched.png
ffhq_NNNNN/landmarks.json

Why ~1,300× larger than prior Thatcher sets:

Source	N identities	Released?
Jacob et al. 2021 Nat Comm	20	Yes (OSF)
Dobs et al. 2023 PNAS	40-100	Yes (OSF)
HoloFaceIllusion-Bench-EEG v1.0	26,317	Yes (this)

Construction: dlib HOG+SVM detection + 68-pt Kazemi-Sullivan landmark + centroid-symmetric bbox rotation + ghost-union mask + Poisson NORMAL_CLONE.

---

v1.2 — Part-Whole illusion (Tanaka 2004 template-based)

Paradigm (per Tanaka & Sengco 1997; Crookes et al. 2013): "a single face outline template ... features (taken from three different individuals) pasted into the template. Foils ... swap one feature with another target's same feature."

Templates	955 (k-means demographic clustering of 24,374 adult FFHQ faces)
Targets per template	9 (each = 3 distinct donors' features)
Features tested	3 (eyes / nose / mouth)
Conditions per (target, feature)	4 (V1 whole_target, V2 whole_foil, V3 part_target, V4 part_foil)
Total samples	25,785 (template, target, feature) triples
Total images	103,140
Tar shards	20 (v1.2/partwhole_00.tar ... partwhole_19.tar, ~3.8 GB each)

Tar layout (webdataset, one sample = template × target × feature):

ffhq_NNNNN/tNN_feature/V1.png       (whole_target)
ffhq_NNNNN/tNN_feature/V2.png       (whole_foil)
ffhq_NNNNN/tNN_feature/V3.png       (part_target on gray)
ffhq_NNNNN/tNN_feature/V4.png       (part_foil on gray)
ffhq_NNNNN/tNN_feature/meta.json    (target_donor, foil_donor, ssim_v1v2)

Why template-based, not 26k pairwise: random cross-identity feature swap on FFHQ pairs produces unnatural results because inter-person feature variation (eye size, nose length, face shape) is too large for any substitution to look natural. Tanaka et al. (2004) solved this by using a shared face outline template per demographic — features paste into identical surrounding context. We replicate that design at scale (397 templates vs Tanaka's 4).

Construction: 2-point similarity transform per feature (preserves donor shape) + mask UNION + Reinhard LAB recolor + MIXED_CLONE Poisson + Laplacian detail-preservation layer + infant filter via lwr/upr face proportion < 3.85.

Full details: v1.2/README.md.

---

Why this dataset exists

Holistic face processing — the brain's tendency to process faces as integrated wholes — is the most robust behavioural signature of human face perception. Three canonical illusions probe it:

Illusion	Test of	Status here
Thatcher (Thompson 1980)	sensitivity to local feature orientation in face context	✅ v1.0
Composite face (Young 1987)	half-face integration	📋 v1.1 planned
Part-Whole (Tanaka & Farah 1993)	feature discrimination depends on whole-face context	✅ v1.2

This benchmark provides controlled paired stimuli (V1/V2 illusion contrasts, V3/V4 baseline controls) at a scale 100-1300× larger than published psychophysical stimulus sets — enough for representational similarity analysis, neural-net layer probing, and EEG-decoder pair-wise pairing.

Intended uses

1. Holistic-processing benchmark across model checkpoints (vision transformers, CNNs, vision-language models): measure ISI / PWI per layer / per checkpoint.
2. EEG paired stimulus design: V1/V2 pairs for N170/N250 ERP differences; V3/V4 for early visual baselines.
3. Representational similarity analysis: paired V1-V4 RDMs vs model RDMs vs brain RDMs.
4. Fine-tuning bio-inspired face models on holistic processing.

Intentionally NOT for

• Training general-purpose face recognition from scratch (use VGGFace2 / MS1MV2 / WebFace260M)
• Commercial applications (CC BY-NC-SA 4.0 — non-commercial only)
• Demographic-fairness-critical applications without further audit (FFHQ has Western-Caucasian bias; we inherit)

License

CC BY-NC-SA 4.0 — inherited from FFHQ. Academic / non-commercial use; share-alike if you redistribute modified versions.

Citation

Main dataset:

@dataset{holofaceillusionbench_eeg_2026,
  title  = {HoloFaceIllusion-Bench-EEG (v1.0 Thatcher + v1.2 Part-Whole)},
  author = {En Hui Li (HKUST)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Enhui-1/HoloFaceIllusion-Bench-EEG},
  note   = {CC BY-NC-SA 4.0; derivative of FFHQ (Karras 2019)},
}

Required upstream citations:

@inproceedings{karras2019stylebased,
  title={A Style-Based Generator Architecture for GANs},
  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{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{crookes2013holistic,
  title={Holistic Processing for Other-Race Faces in Chinese Participants},
  author={Crookes, Kate and Favelle, Simone and Hayward, William G.},
  journal={Frontiers in Psychology}, volume={4}, pages={29}, year={2013},
}

Changelog

• v1.2 (2026-05-26): Added Part-Whole subdataset (template-based Tanaka 2004 design, 397 templates × 6 targets × 3 features × 4 conditions = 28,584 PNGs).
• v1.0 (2026-05-25): Initial release — Thatcher illusion, 26,317 identities × 4 conditions.

Contact

GitHub Issues — feature requests, bug reports, or paradigm extension proposals.