About Model Training
I have experienced couple of your models, especially the paper-cut and it feel super great. Can you share your experience in training models? For example, how many photos in training set and class/reg set as usual ,training steps, training Epochs, the learning rate , etc. Thank you very much!
Hey mate! Thank you for the kind words.
I generally train with somewhere around 20-40 images, depending on the quality of base images. I try to keep my images as varied as possible with regards to subjects when training a style (eg: animals, landscapes, portraits, objects, vehicles, etc) ; This helps with making sure the model only learns the general style aesthetic and not the actual subjects, although, this is not an exact science.
Step count depends on the amount of images used. When using the LastBen Notebook, I was using the number of images x 100. With text encoder training set to 30%.
When I use local dreambooth from within the WEB-UI, I generally use Number of images x 30 steps with EMA training and Text encoder Training Enabled with 1e-6 LR polynomial, continue for images x 40 steps without text encoder at 4e-7 LR polynomial, then finish with images x 10 steps with text encoder enabled at 1e-7 LR Constant.
I usually crop my images to 640x640 and train them with a batch size of 2.
Thank you very much for your thorough reply ,it was so helpful.one more questions, Would you tell me How to set the Warmup and the Adam optimizer in WebUI dreambooth?
I dont know much about the adam optimiser
Warmup (which i dont use) is generally best for 10% of total steps. I.e:- 1000 steps total training will have 100 steps warmup
Get it. THX Bro!
can we make some SD2 768 model ? <3 <3
What's the best size for this model? 512x512 or 768x768?