Sure, I'll share some general tips. Not going to write a full guide since things change a lot and there are existing guides e.g. https://rentry.org/lora_train.

1. Collect a high quality dataset with lots of images (minimum ~15-20). I recommend using https://github.com/Bionus/imgbrd-grabber and setting it up so that it'll download the tags as well. There is some setup you'll need to do, but it's somewhat covered in the lora_train guide. For the "text file content" part, I use %all:,excludenamespace=general,unsafe,separator=^, %
so that it's seperated by comma and not by space. Also, you'll need to escape parenthesis. I have my own script for this.

If your images can't be found on booru websites. I use google images/artstation/twitter and I manually tag them. You can use wdv1.4 tagger, but I would manually check them since tagging is very important and tagging wrongly can make your model very worse.

If you're going for a specific look, make sure all the images are consistent. Sometimes you can't avoid this because artists just draw it differently.

2. If you're training on a single folder. I recommend using a repeat of 1. This allows you to have a model everytime it's trained on all the images allowing you to pick the best model.

3. I find my models usually good at 30-40 epochs, but this depends on a lot of variables. You should do a X/Y plot using prompt S/R to test each model.

4. I also do X/Y plots for mixing with other LoRAs to see if my characters are not overfit.

5. I prefer to make my models usable at weight 1. Others recommend overfitting and adjusting the weight to get the best images, but I avoid this to make things simple and not have to figure out the best value. Also, adjusting the weight is a linear change whereas going for weight 1 is non-linear, so I suspect it's better to do it this way, but I haven't tested it.

@Makalandia Not sure what dreambooth you're using, but I'm using kohya's script. See https://rentry.org/lora_train#how-to-use-this