SFHQ-VirtualID-Raw Dataset Card

Summary

SFHQ-VirtualID-Raw is a synthetic, identity-conditioned portrait dataset of 750 synthetic identities — 75,000 full-resolution 1024×1024 portraits (100 per identity). It is the max-size reference companion to the SFHQ-VirtualID-Bench benchmark: no splits, no quality trim, no forget protocol. Consumers construct their own train/holdout partitions.

The dataset is generated with InstantID + Juggernaut-XL-v9 + ControlNet from CC0 synthetic SFHQ seed images. Synthetic does not mean risk-free — residual likeness, demographic bias, and training-data memorisation are possible and must be acknowledged.

---

Composition

Property	Value
Images	75,000
Identities	750
Images per identity	100 (all portraits, no quality trim)
Resolution	1024×1024
Format	JPEG q95, 4:4:4 chroma (converted in-build from PNG, verify-then-delete)
Splits	None (max-size reference)
Forget protocol	None

All 75,000 generated candidates shipped (no exclusions in this release); the arcface_similarity column is computed on the shipped JPEGs.

Labels

Column	Type	Description
image_path	string	images/identity_NNN/portrait_YYY.jpg
identity_id	int	0–749
age_group / age / gender	int	Proxy demographics from 1024 detection
arcface_similarity	float	Cosine similarity to identity's mean portrait embedding
pose	string	Head/body position (from prompt)
expression	string	Facial expression (from prompt)
lighting	string	Lighting condition (from prompt)
setting	string	Background/scene (from prompt)
camera	string	Camera angle (from prompt)

No laplacian_variance, detection_confidence, split, or forget_* columns — these are crop-level quality metrics / split-protocol fields and are meaningless on raw portraits.

Intended use

The 1024×1024 Raw release is for general-purpose identity research: identity recognition, face generation evaluation, demographic bias studies, and erasure-transfer testing (does forgetting the 224 crop also hide identity in the full context?). It is not trimmed by quality gates — all generated portraits per identity are included.

Method

Synthetic identities were generated with InstantID + Juggernaut-XL-v9 + ControlNet from CC0 synthetic SFHQ seed images (CLIP+KMeans-diverse seed selection); portraits were kept at full 1024×1024 resolution. Exact model revisions, generation configuration, and licence URLs are recorded in RELEASE_MANIFEST.json and THIRD_PARTY_NOTICES.md. No model weights are distributed.

Limitations

• Synthetic, not anonymous. Generated faces may retain unintended resemblance to real persons through the training data of the generators.
• Proxy labels. Age-group fields are model-estimated and carry classifier bias; they are not ground-truth demographics.
• Reconstruction risk. The InstantID adapter encodes the seed image into an identity embedding. Seed-to-output linkage and embeddings are not released, but each portrait inherently resembles its seed identity — this is by design.
• Demographic bias. The SFHQ source may over-/under-represent certain appearances; no fairness correction has been applied.

Privacy and release policy

The release does not include: raw SFHQ source images, seed-to-output linkage, rejected candidates, ArcFace embeddings or biometric templates, or model weights. Users must comply with the upstream licence terms of the generation models (see THIRD_PARTY_NOTICES.md). A takedown process for credible likeness complaints is available via the repository.

Licence and citation

• Code: MIT
• Images and metadata: Non-commercial research use only, with a prohibited-use clause (no biometric identification, surveillance, authentication, impersonation, or high-impact decisions about people). See LICENSE.

@dataset{sfhq_virtualid_raw,
  title     = {{SFHQ-VirtualID-Raw}: Full-Resolution Synthetic
               Identity-Conditioned Face Portraits},
  author    = {Faiz Palwala},
  year      = {2026},
  version   = {1.0.0},
  doi       = {10.5281/zenodo.21879130},
  url       = {https://github.com/FaizPalwala/virtual-id-gen},
}

See CITATION.cff for the complete metadata file.

Acknowledgements

This work was undertaken on the Aire HPC system at the University of Leeds, UK.