Playing God with InstantID (Generating Virtual Identities)

To fix the dataset, I turned to InstantID and Stable Diffusion. Instead of clustering random faces, I am now explicitly synthesizing 750 unique “Virtual Identities”. For each identity, I generate exactly 100 candidate images under various prompts, lighting, and angles.

This turned from a machine learning problem into a massive data engineering problem.

I wrote a completely new codebase (virtual-id-gen) and transitioned from local scripts to Slurm array jobs. I am currently parallelizing the generation of 75,000 images across the University of Leeds’ Aire HPC cluster, utilizing 15 shards running concurrently.

It’s a logistical nightmare of checkpoints, prompt scheduling, and node timeouts, but the resulting dataset is going to be bulletproof. Instead of “visually similar people,” the network will be forced to unlearn true identity representations.