cff-version: 1.2.0
title: "SFHQ-VirtualID: Synthetic Identity-Conditioned Face Datasets for Machine Unlearning"
message: "If you use either release in your work, please cite the corresponding dataset entry."
type: dataset
authors:
  - family-names: Palwala
    given-names: Faiz
    orcid: 0009-0000-1426-9471
version: 1.0.0
date-released: 2026-08-15
doi: 10.5281/zenodo.21877893
identifiers:
  - type: doi
    value: 10.5281/zenodo.21877893
    description: "SFHQ-VirtualID-Bench (primary benchmark release)"
  - type: doi
    value: 10.5281/zenodo.21879130
    description: "SFHQ-VirtualID-Raw (full-resolution portraits)"
url: "https://github.com/FaizPalwala/virtual-id-gen"
repository-code: "https://github.com/FaizPalwala/virtual-id-gen"
abstract: >
  A family of synthetic, identity-conditioned face datasets generated via
  InstantID + Juggernaut-XL-v9 from CC0 synthetic SFHQ seed images.  750
  synthetic identities with identity-level split isolation (retain 675 /
  forget 75) and MUFAC-aligned per-image train + holdout subsets, following
  a sequential 15-step forgetting protocol.  Two releases:
  SFHQ-VirtualID-Bench (224x224 aligned crops — 67,500 balanced + 36,064
  imbalanced with a ratio-based 82:41:16 train gradient, both schedules
  uniform and seeded-Poisson) and SFHQ-VirtualID-Raw (1024x1024 portraits,
  75,000 images, JPEG q95 4:4:4, max-size reference).
keywords:
  - synthetic face dataset
  - machine unlearning
  - identity-conditioned generation
  - InstantID
  - Juggernaut-XL-v9
  - SDXL
  - long-tail distribution
license: other