I decided to make this LoRA both as a test of a scraping script and an examination of a different way of training.

Usage:

• As with most character LoRAs, a strength of 1 works best.

• Be specific with the clothing you are looking for, otherwise, you will just get something revealing.

• The keyword smiling tends to give the best faces.

• It works best on realism models, but can even work decent on others like Dreamshaper.

Some quick notes on the training:

• I used full BF16 training, my usual precision.

• I used WD14 captioning which performed a lot better than BLIP in my opinion.

• Trained at 1080x1080, but I probably could have pushed it further.

  • It seems to have a preference for more vertical images because of the dataset.

• I believe it is because of the captioning method, but it took a network rank of 128 to get the face right. Otherwise, the LoRA was just getting from the neck down decent.

  • Perhaps this is a clothing LoRA as well given generic tags? Curious to see what could be done with this tidbit.

Once again, this was created as a test. It's not the greatest character LoRA I have ever made, and didn't really take much effort to make. But that's the point of this little experiment. It does, however, produce some good results with a little experimenting.