The Great Dataset Pivot (Why Clustering Failed)

Research is mostly about realizing you were wrong.

Remember back in April when I used K-Means to cluster 30,000 synthetic faces into 200 “identities”? It turns out, clustering visual embeddings doesn’t create a realistic person. It creates a group of visually similar people.

When you ask an algorithm to unlearn a cluster, it’s not unlearning a single person under different lighting conditions, angles, or expressions. It’s unlearning a highly specific, narrow visual aesthetic. This makes the unlearning task artificially easy and fundamentally flawed.

My results are shaky. The foundation is cracked.

I have made the painful decision to halt the core unlearning evaluation and throw away the clustered dataset entirely. I need explicitly synthesized Virtual Identities, not clustered approximations.

It’s time to build a massive generative pipeline from scratch.