weight: 0.8-1
First try of flux training style lora, 123 data sets trained 2100 steps, using natural language labeling, no specific trigger words, you can test the return image and give feedback
Description
FAQ
Comments (2)
Hi. Thanks for your work, i just made some tests with your Lora, and I really like the style (I will not use it for Mangas, but I like the fact that it does not do gradients. And that's perfect for my use cases).
It seems to me that starting the prompt with "Illustration of [something]" give more coherent results than the same prompt without this keyword.
A little drawback is on text interpretation, we loose this but I think that's because there is not or little texts in your dataset (I have read that it's needed for Flux Lora - seen that on reddit if I remember well). Edit : that's a little better with 40 steps (I usually only use 20s with flux).
Anyway, the whole lora is pretty good, I will follow you looking for future improvements if any. Beau travail, Charine, et un grand merci ;)
Thank you for your feedback, there are still some problems with flux training, for example, the limit on the number of tokens, in addition to this, I use a high learning rate that may lead to overfitting, I have a plan to train v2, can look forward to it