Trigger word: ral-fluff
☕ Buy me a coffee: https://ko-fi.com/ralfingerai
🍺 Join my discord: https://discord.com/invite/pAz4Bt3rqb
Description
FAQ
Comments (12)
How does one even train a lora on a non-existent before concept? Is it like selection and breeding in biology where you start with the few successful generations based on nothing and pure luck and then train several generations on it?
Hey there seedmanc! Thanks for asking this question. All my training data is based on DALL-E generated images. This is how i generate my images for the LoRA training :)
Why not generate them on SDXL?
@gfhdgfhfdhgfh because a year ago I trained the LorA for SDXL
@RalFinger link?
@gfhdgfhfdhgfh what do you mean by link?
@RalFinger from beginning... "All my training data is based on DALL-E generated images." I asked you why not on SDXL...
@gfhdgfhfdhgfh 🤨
SDXL can generate better and more realistic images. Why did you choose DALL-E?
)))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))
it's a smile
@gfhdgfhfdhgfh One reason might be that it could work better. If you train a lora on images created by the very model you make the lora for, I would expect problems to quickly pop up as biases and mistakes the model makes get multiplied and reinforced.