Really enjoy timeless thank you!
Could you provide a few more details on how you trained timeless? There are plenty of dreambooth tutorials, but I find this one so artful and most tutorials detail with characters not styles. A few things I'm curious about, learning rate, image sizes, number of example images etc. Thanks again!
Thank you so much!
Timeless was trained on Stable Diffusion v1.5 with updated VAE. The dataset ended up being about 130 images, which I cropped to 1:1 and resized to 512x512. I trained with with a batch size of 1, LR of 1e-06 with a polynomial lr_scheduler. I also used prior preservation with a class dataset of 1500 images (simply made with the class prompt "photograph".) It was trained on a 3080ti 12GB w/ 8bitadam optimizer. And the checkpoint I ended up going with was 14000 steps.
Style training is really similar to character training. It all comes down to the dataset. People like to mess with all the training parameters, but imo nothing will improve your model more than improving the dataset.
Hope that helps!
This is gold thank you! I don't suppose you remember if you used bilinear/bicubic/other resampling during the resize? Did you also caption each image separately?
No wonder there's so much coherence in the imagery. You probably just saved me weeks if I ever manage to get my act together and produce a model. I'm getting wonderful space age images out of timeless, thank you! I hope to use it to produce images some music I'm making if I ever finish the music!
I don't remember what resampling was used in the resize. It was to make the images smaller, not bigger, so I wasn't too worried about that as I would be if upscaling. And it was trained with dreambooth class/instance training, so no captioning needed. I just trained the instance as "timeless style"