:D
Description
FAQ
Comments (6)
Amazing results. How did you do it? I have tried to make Arsh-style lora, but the result is not satisfactory, would you be interested in trying?
The lora file contains the training parameters, other than that just set it to save every epoch or every ~200 steps so you don't need to worry about under/over training, for images grab as much as possible, you can also use manga cutouts (don't use the whole page unless its a single scene, just crop the region of interest) or augment the data with cropped regions (face closeups, penis closeups), for tagging I use this https://github.com/toshiaki1729/dataset-tag-editor-standalone , first swinv2-tagger-v2 at a threshold of ~60%, then convnext-tagger-v2 at 70%, after that you remove the duplicate tags and do a cleanup on things you deem important, could be something removing penis tags on images without penises, fix flaccid/erection tags or adding a tag for something the tagger didn't recognize
@Struramill Sent you a PM 😊
This is probably going to see a lot of use in my installation of SD. Thanks for making and sharing! Do you have (or have plans to make) Pony XL versions for the previous ones you made for 1.5? (especially Homri, Beater, and Nayoshi)
Some other guy made a beater one https://pixeldrain.com/u/JNko96Ld
Homri https://pixeldrain.com/u/4GgU9wHB
Danzi https://pixeldrain.com/u/cbn2geWX
I'll keep the nayoshi in mind
@eeeeeqeq Awesome thanks. I'm still fairly new to SDXL (but not SD1.5), so I haven't really looked for a lot or LoRAs yet, so thanks for the links. I really like your SD1.5 nayoshi LoRA.