Train
Teach a model something of your own, on this computer: a picture model your subject or style, from a few pictures, or a chat model your writing voice or how to answer, from your writing or example chats. What you train it on never leaves this computer.
How it works
- It trains a LoRA: a small add-on (around 12 MB for Stable Diffusion 1.5) that teaches the picture model your subject or style. The model itself doesn't change.
- Every picture is labelled with your trigger word, so the model links the word to what your pictures have in common. Put the word in a prompt to call it up.
- Stable Diffusion 1.5 takes a few minutes on a recent graphics card; SDXL takes around half an hour and needs a 12 GB card. Training uses one graphics card, and stops models on it first if it needs the room.
- On the Pictures page, choose it under Style. If pictures copy it too closely (every dog becomes your dog), lower its strength.