After reading on this forum, I see that the amount of quality training data is very important. I've seen suggestions on making sure you have 1k-10k pictures at varying angles of both the source and target. However, my situation is that I can get 5k pictures of the target, but I can only get barely 200 pictures of the Source. I believe this should be a somewhat common scenario. Does anyone have any advice to help raise the final swap quality in this scenario? Or is there anything that anyone has tried that seemed to work for them?
Some ideas I had.
some sort of data augmentation techniques to generate more pictures of B
select a different model other than the default model
train for longer (right now I'm at 24 hours of training. Results so far are ok, but still not satisfied)
change hyperparameters such as learning rate, batch size, etc
