Synthetic-Faces

Synthetic-Faces is a fully synthetic 3D face dataset designed to study identity preservation vs topology variation in geometric deep learning models.

The dataset was created as part of the WBES-FaceEmbedding project and is primarily intended for:

• latent space analysis
• topology invariance studies
• metric learning on 3D facial geometry
• evaluating robustness to remeshing, cropping, and noise

No real human data is included.

---

Motivation

Direct cross-topology datasets (e.g. BFM ↔ FLAME) are currently difficult to obtain in practice due to:

• unavailable or deprecated conversion pipelines
• inaccessible pretrained models
• unstable public tooling

To unblock research progress, this dataset simulates cross-topology conditions by generating multiple controlled variants of the same synthetic identity.

This allows systematic analysis of how topology, resolution, and surface perturbations affect:

• latent identity embeddings
• distance-based metrics
• encoder robustness

---

Dataset Structure

All files are stored flat in the repository root. This is intentional and acceptable for Hugging Face datasets.

Each identity is represented by multiple .npz files:

idXXXX_GTready_original.npz
idXXXX_GTready_remesh.npz
idXXXX_GTready_crop.npz
idXXXX_GTready_noisy.npz

Where XXXX is the subject ID.

Variants

Variant	Description
original	Reference mesh (canonical topology)
remesh	Same geometry, different mesh connectivity / resolution
crop	Z-axis cropped surface (region mismatch simulation)
noisy	Surface noise added (robustness stress test)

All variants share the same identity and differ only in representation.

---

File Format (.npz)

Each .npz file contains precomputed geometric operators compatible with DiffusionNet-style models:

import numpy as np
z = np.load("id0000_GTready_original.npz")
print(z.files)

Typical keys include:

• verts
• faces
• mass
• evals, evecs
• L_indices, L_values, L_shape
• gradX_indices, gradX_values, gradX_shape
• gradY_indices, gradY_values, gradY_shape

The dataset is ready-to-use for intrinsic surface learning pipelines.

---

Intended Use

This dataset is suited for:

• identity vs topology disentanglement
• encoder-only latent analysis
• metric learning (contrastive / ranking losses)
• correlation studies (Pearson, Spearman, WBES)
• robustness evaluation under domain shift

It is not intended for:

• texture modeling
• expression dynamics
• photorealistic rendering

---

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Pampaj/Synthetic-Faces",
    repo_type="dataset",
    local_dir="Synthetic-Faces"
)

---

License

This dataset is released under a permissive research license.

• Synthetic data only
• No biometric or personal information
• Free for academic and research use

(Exact license to be finalized if required for publication.)

---

Related Work

• WBES-FaceEmbedding (code): https://github.com/Pampaj7/WBES-FaceEmbedding
• DiffusionNet: https://github.com/nmwsharp/diffusion-net

---

Contact

For questions, collaborations, or extensions:

Leonardo Pampaloni GitHub: https://github.com/Pampaj7