Datasets:
How many for DB?
Thank you for this generous dataset! I’ve had more success with using fewer class images. Do you ever use thousands? What kind of results are you getting in terms of speed?
@okaris Thank you! So, normally the lowest number of images I use would be 2000. I've also had great success with using 4000-5000 images as well. It can be a bit slower when using regularization images, but nothing too crazy in my experience. I've heard of people using far fewer regularization images with success as well, but I imagine that it may be different depending on your training image content, what the base model already knows, and what sort of outputs you are looking to make.
@ProGamerGov Thank you! Could you share the settings you use for Dreambooth. How many steps, batch-sample size, gradient accumulation etc? Would be nice to start with better settings than try and explore (with such a lengthy training) :)
@okaris You can see what settings I use here: https://huggingface.co/ProGamerGov/Object-Taped-To-Wall-Diffusion-V1 and here: https://huggingface.co/ProGamerGov/Min-Illust-Background-Diffusion
It looks like these were generated using stable diffusion. In some sets, the images are distorted / contorted or the usual bad anatomy stuff. What affect might these have on the base model being used for DB training?
bad anatomy stuff. What affect might these have on the base model being used for DB training?
I just realized that this is what was spoiling my results!