>>43716679>Correct. What you prune is what the lora focuses on learning, that's why it's important to have things you don't want tagged in the image so it doesn't learn them on accident. Even so, if the learning rate's too high, you have too many repeats, etc, it may end up learning them anyway. It's very finicky. An example of mine would be when once I had a "phone" tagged, but the lora latched onto it and would try to gen the character holding a phone, or something square like it unless I pu it in the negative prompts. If you solely focus on the outfit, it has a much greater potential to work not only better on other characters, but should also reduce the number of images you need for it to learn it. The more complicated the subject of a lora you train the more images you need to train it.Okay, I'm gonna prune things again using this advice and take out all the outfit related tags. This is a pretty small dataset (34 images total), so should I take out other tags such as background and expression related tags? Also I do notice that 6 images have a swing so maybe I should take that out too to avoid your phone example.
>swear it used to have more. It used to call for powershell to train, looking at the dates on the bottom it's been updated recently. Continuing my reading and troubleshooting on it now. Mine doesn't run popup.bat correctly, so I'm looking into that. If you can get that to work for you and have more questions though let me know.I'll try the popup.bat after I prune things again and I'll let you know how that goes!
