Hello to my 108 followers, I decided to retrain my Giselle lora since I used 7th anime and not very good images for it and all that shebang so I went and redid all the tagging and used better images. Anyways, the lora was trained on MeinaHentai @768p so the performance of the lora is much better than v1. I recommend using DPM++ 2M Karras and 1.5x upscaling if you got the hardware for it. My recommended prompt is simply 'giselle' and maybe twintails if the lora does a ponytail. Thats about it though, I have been uber busy and I will be even more busy starting from January so I think I will just retrain my older models as a sort of last hurrah until I go into hibernation for a while. If you liked the update give the model a review or favourite it, or just a positive comment if you want to :). <hr />Version 1: So basically I'm not very good at making models, so probably this lora is less than ideal performance wise, so i guess if anyone can give me some tips on improving it I'm all ears. Anyways, this is Giselle from Project QT, I used like 35 images all manually tagged with lovely close ups of the things that matter and then I used <a target="_blank" rel="ugc" href="https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb">Lora Trainer.</a> to train it. I used <a target="_blank" rel="ugc" href="https://huggingface.co/syaimu/7th_Layer/blob/main/7th_anime_v1/7th_anime_v1.1.safetensors">7th anime v1.1 </a>to train it with. Giselle is customisable by some miracle so if you don't really like her fox ears just don't put it into the prompt, this works quite well with her face and head so you can make her have long hair, twintails, ponytail and all that, but I'm not sure about the rest of her body parts. Recommendations to prompt her in my limited and unprofessional testing: Portraits: Giselle, red streaked hair, bangs, twintail, headband, fox ears Full body: ^ and, thick thighs, navel, legs maybe This lora actually makes pretty good portraits but its very picky in what sampling method it likes so id recommend using DPM2 Karras or DPM++ SDE. Please post some nice picture cause i really want to see if its only my very low powered pc thats producing these mid pictures or its an issue with the lora itself. Thank you 😘

Giselle - Project QT

I'm not sure about the training epochs I used Dadaptive instead of Adam8Wbit

Files with SHA256: 313f67589d84ee6374184c8479ffafeda8c4b48350b0be63bf107f79585d6e11