Hi, I use the same parameters for most loras, currently using dylora with "network_dim": 32, "network_alpha": 16, "conv_dim": 32, "conv_alpha": 16. The secret is a good dataset. For this I had ~200 images that I found on yandex, google, pinterest, and some other places. If your dataset is too small, try training a lora anyway. Then generate more similar pictures with that, then train a new lora with your new synthetic dataset.

@ThexSam set the amount of steps in such a way so 1 epoch will be ~5000 steps. So for a 100 image dataset name the folder 50_thing. Note that with a >1 batch size it will actually be less steps but whatever. I save the model every 2 epochs. I set a very big epoch count (like 100) and just look at generated samples to determine when I should stop. I start training when I go to sleep and when when I wake up it's usually good enough and I stop it (Note I'm training on an RTX 3090 with batch size 4, on 640x640 setting) I noticed that models don't really overfit in this period so don't worry about just taking the latest saved model. Also note that you can change the sample prompt during training by changing the txt file, it's very useful. I generate a sample every 250 steps.