expectedwrong hindsight

You Are the Dataset Now

On the particular mistake of training a LoRA on your own face too many times.

1 min read 182 words #machine-learning #lora #flux #meta-ray-bans #mistakes
hindsight — still happening

Overfitting a LoRA on your own face is still exactly this easy and exactly this embarrassing. The training pitfalls haven't changed. The self-awareness hasn't helped.

There is a specific moment when you realize you've overfit a model on your own face.

It's not a gradual dawning. It's a confirmation — you look at the outputs, you already knew, you just needed to see it rendered at full resolution to make it official.

Trained a Flux LoRA on photos I took with the Meta Ray-Bans. Used too many of myself. The model has now learned, with great conviction, that the concept I was trying to capture is my face. This is not what I was going for. The glasses were a secondary concern. The face was supposed to be incidental. The model disagreed.

The Ray-Bans at least got the glasses right. Portrait mode, good light, decent angle — the frames came through clean. So the LoRA knows what the glasses look like, and it knows what I look like wearing them, and it has fused these two facts into something that is technically a success and practically a curse.

The correct number of selfies to include in a training set is, apparently, fewer than however many I used.