Training Images
Thanks again for your great contribution to AI art community. Question, how many training images did you use for this model? Did you train the character and background styles separately or together in one shot. I was experimenting with training on Ghibli style but not getting good results. I was mixing characters, backgrounds from different movies for a total of 40 images (20 characters, 20 backgrounds). I trained for 4000 steps / LR:0.000001 and used class prompt (and 1500 reg images) as "artstyle" Any insights are appreciated.
Also interested, wait for more details of training.
Nitro has this type of questions quite a lot, he may answer in more detail for Ghibli model specificaly, not sure, but for now, here is the guide he made for making a Style model
https://github.com/nitrosocke/dreambooth-training-guide/blob/main/README.md
Thanks, I read everything there is from Nitro before posting this question :)
Hi there!
I used 118 instance images total with 24 images of random scenes and objects, 13 animals and the rest are 81 person images. Settings where 15k steps at LR 1e-06, with 1500 class images of "illustration style".
Only thing I see with your dataset might be that it's too many background images, that might be too much. Even with mine the scenes sometimes overpower the rest of the prompt.
You rock nitro!
Hi there!
I used 118 instance images total with 24 images of random scenes and objects, 13 animals and the rest are 81 person images. Settings where 15k steps at LR 1e-06, with 1500 class images of "illustration style".
Only thing I see with your dataset might be that it's too many background images, that might be too much. Even with mine the scenes sometimes overpower the rest of the prompt.
@nitrosocke
Thank you so much for this!
What is the source for class images "illustration style"? Where can I get these or generate them?