I trained a LoRA on my own face. Here's every mistake.

What I was trying to do

I wanted a Flux LoRA that could put my face in a scene without turning every portrait into the same lighting setup.

What actually mattered

The first dataset was too big and too same-y. Cropped phone photos, repeated backgrounds, captions that just said my name. The model memorised the room.

Eighteen photos with distinct captions — clothes, camera, background — beat the sixty-photo pile. Rank and alpha needed a short sweep, not a weekend of random numbers.

What I would not repeat

Training on images I also used as the site hero, then being surprised when the hero looked like the training set. Hold a few frames out.

The portraits on this site are generated from that LoRA. The writeup is the part I can stand behind: curation beat volume.