The failures it took to get the handful that worked
This one comes out of my own curiosity: how far can open, local AI models for video and sound be pushed on hardware you actually own?
Local model formats change constantly and are barely documented. Keeping a pipeline stable across different hardware is a dark art.
CUDA, PyTorch, LoRAs and ComfyUI. A lot of rebuilding. Blood, sweat and tears.
What surprised meNothing surprises me anymore.
Short generated clips, some with sound, and a much bigger pile that didn't make it.
Every tool that works in production started as something like this. Knowing where it breaks is how I know what's ready to use on client work.