It Didn't Learn My Face. It Learned One Photograph.

This is the failure that actually matters if you want to USE the model. I wrote a prompt describing the room I was sitting in. By 8000 steps the model ignored every word of it and rendered the background from my old videos instead - the one background that was in every single training clip. It didn't learn me. It learned that photograph. The more I trained it, the less control I had. At 250 steps the prompt is in charge. At 8000 the prompt is decoration. Full breakdown - 32 checkpoints, one face, four failure modes measured: https://youtu.be/w_7EpghF8fw #LoRA #AITraining #StableDiffusion #MachineLearning #AIVideo