The SFHQ-VirtualID Release (And the Final Push)

It’s finally done. The generative array jobs finished, and I spent the last week wrangling the massive output.

I wrote scripts to clean up the 123 rejected/unreadable generations, fix the manifest paths, convert the raw images to JPEG q95 for storage efficiency, and lay out the identity-level splits: 675 retain / 75 forget identities on a 15-step deletion protocol, with MUFAC-aligned train/holdout subsets inside every identity.

pie title "Dataset Split (Identities)"
    "Retain" : 675
    "Forget (deletion targets)" : 75

I formalized the architecture into two concrete releases: the Raw release (all 75,000 images) and the Bench release (curated specifically for iterative unlearning experiments).

Releasing a pristine, ethically-safe facial identity benchmark feels incredible. The SFHQ-VirtualID dataset is officially live. And now, finally, I can rerun my AdaptiForget iterative evaluations on a dataset that actually represents reality.

The final push to the dissertation starts now.