That Wasn't Me. And Neither Was the Voice.

For every video on this channel, the background behind me was fake. In this one, so was I — and so was the voice for the first minute. I trained a character LoRA of myself on 28 clips of my own face, saved a checkpoint every 250 steps to 8000, and rendered the same prompt through all 32 to find where the sweet spot actually is. 164 hours of training. It was finished after 39. Full breakdown, all four failure modes, and the measurements: https://youtu.be/w_7EpghF8fw #ai #LoRA #AIVideo #MachineLearning #ContentCreation