Done a lot of experiments, which is the reason version 0.4-0.9 and their sub versions isn't there. More pictures has been added to the training set, including images generated by the previous models. Captions has been reformulated in attempt to make the model aware whats worn under the anorak.
It's not strictly needed to use a trigger word, just asking for a jacket, and enabling the Lora has a effect.
This version was trained with Adafactor, lr 0.001, batch size 8 and cosine with restarts.
Done a lot of experiments, which is the reason version 0.4-0.9 and their sub versions isn't there. More pictures has been added to the training set, including images generated by the previous models. Captions has been reformulated in attempt to make the model aware whats worn under the anorak.
It's not strictly needed to use a trigger word, just asking for a jacket, and enabling the Lora has a effect.
This version was trained with Adafactor, lr 0.001, batch size 8 and cosine with restarts.