Thank you for your kind words😁. I trained these models using free GPU resources. Unfortunately, those resources have run out now, and I'm not sure when I'll be able to train the LoRA model next.

@spoilersmii Sure, here's the general process I follow for data processing: Training set: Bulk scrape One Piece screenshots from fancaps → Use the dghs-imgutils package to filter out Nami's images → Categorize the images into three tiers based on their attractiveness → Perform super-resolution upscaling on medium-tier and above images using the IllustrationJaNai_V1_DAT2 model → Tag the images → Set the number of repetitions, with more repetitions for more attractive images. By the way, due to the filtering by dghs-imgutils, most of the mid-to-high tier images are upper body shots, so I sourced some full-body images from the web to include in the training set. Regularization set: Scrape One Piece images from fancaps but exclude Nami's, for the regularization set (this part could also be chosen as part of the training set). Additionally, to avoid overfitting while increasing the intensity of training, I fetch 500 appealing images from Danbooru to use in the regularization set.

the dataset include some pretty fanarts.

maybe

Firstly, your image cannot be opened properly. Secondly, it is quite common for the Pony model not to follow the prompt instructions accurately. Of course, you can also complain that this LoRA has overfitted; however, due to the poor capabilities of the base model used for training, the LoRA model adapted has to take on too much responsibility. Lastly, I have already released the Illustrious-XL version of the LoRA.

pony v6