I’ve just released SFHQ-VirtualID, a synthetic face dataset family built for identity-level machine unlearning. Everything is generated, so there are no real faces in it, and both releases have DOIs.<\/p>\n
The problem I kept running into: in most face datasets used for unlearning, a person’s images are spread across splits, so “forgot the person” and “forgot some images” end up confounded. Here, each identity_id maps to exactly one split and contributes both train and holdout images (675 retain \/ 75 forget, 15-step protocol, MUFAC-aligned holdouts).<\/p>\n
What ships:<\/p>\n
– Bench: 67,500 balanced + 36,064 imbalanced 224\u00d7224 aligned crops. Uniform and seeded-Poisson forget schedules, plus a 5:1 long-tail popularity gradient for long-tail forgetting tests.<\/p>\n
– Raw: 75,000 1024\u00b2 portraits (100 per identity) with per-candidate prompt metadata for pose, expression, lighting, setting and camera.<\/p>\n
I did not filter candidates on ArcFace identity similarity. The 0.40\/0.45 thresholds are config defaults that ship as recorded columns (arcface_similarity, laplacian_variance, detection_confidence) rather than enforced filters, because silently dropping borderline-similar candidates hides exactly the confound that could explain a model’s apparent unlearning. Only the quality gate is enforced (detection plus Laplacian sharpness \u2265 80), and the 123 rejected candidates are documented in the manifest.<\/p>\n
Reproducibility: 15-shard Slurm array on UoL’s Aire HPC (~120 GPU-hours), InstantID + Juggernaut-XL-v9 + ControlNet, pinned environment (torch 2.6\/cu124, diffusers 0.39.0, insightface 1.0.1, pinned antelopev2 revision), RELEASE_MANIFEST.json and SHA-256 checksums.<\/p>\n
Links:<\/p>\n
Code<\/a> \u00b7 Hugging Face repos [Raw<\/a>] [Bench<\/a>] \u00b7 Project Overview<\/a> \u00b7 Bench DOI 10.5281\/zenodo.21877893 \u00b7 Raw DOI 10.5281\/zenodo.21879130<\/p>\n Caveats: the dataset is synthetic, so it’s a proxy for real-face benchmarks; demographic balance is inherited from the seed selection; and the two releases are related (Bench crops derive from the Raw candidates), so they aren’t independent test sets.<\/p>\n Happy to answer questions, and I’d genuinely like the split design stress-tested, since that’s the part I’d most want criticised.<\/p>\n<\/div>\n submitted by \/u\/faizpalwala <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
