>>5685

I don't train Loras myself, I'm just making/helping to make a few datasets for the training of a few Loras. If you have enough good images I'd recommend you request a Lora training in this thread: >>4264

Also there in the opening post you'll find a list of criteria for the quality etc.. It says there that at least 20 images are needed, but from recent experience that seems to be too few, hence my recommendation of 40-50 pics at least.

You can mostly prepare the dataset pretty easily yourself, thus saving Entropy's time; if you are using a1111/forge you can install this tagger extension: https://github.com/picobyte/stable-diffusion-webui-wd14-tagger.git

Once you've got all pics in order in one folder run the batch process (there should be a new tab in your a1111/forge after you install the extension and restart the ui), select Interrogator model "WD14 moat tagger v2" and hit "Interrogate". That will create a .txt file with tags for each pic. To further easily edit these .txt files I also recommend you install this little taggerUI: https://github.com/jhc13/taggui/releases/download/v1.30.1/taggui-v1.30.1-windows.7z ...you'll need to decide on a triggerword for your concept and add that to all .txt's. In your case simply going with "broken fingers" may work, but there's no guarantee before trying. Also for pony add "score_9, source_anime" to the start of all tags. So each .txt should start with "score_9, source_anime, broken fingers" in your case.