Update v2: retrain with maybe-possibly better settings.
Look at images for usage examples.
Training notes:
Cogvlm 150 tokens several sentences + minimal danbooru
lr 0.0002
1800 steps
dim/alpha 16/16
