v3:I retrained LoRA with a smaller parameter size and now it only takes up 28MB of storage. Although the fitting is not detailed enough, perhaps some people need it.v2:I changed the clothing trigger word from <code>bodysuit</code> to <code>white_comet_bodysuit</code>. I retrained a smaller LoRA using the original train-dataset and reduced overfitting. However, due to the fewer epochs of training, the fitting of the character features is not as detailed as the first version. It may need to be combined with HD restoration, or use more iterations.Nevertheless, I do not have enough resources to train. This version was trained for 5 hours using my RTX3060.v1:This is the LoRA model I trained to generate images of Kiana Kaslana, the protagonist of the mobile game Honkai Impact 3rd. She wears her original battle suit and is known as the "White Comet". (But I forgot to use this trigger word, my mistake.) The model's trigger word is <code>Kiana Kaslana, bodysuit</code>.I recommend using the following settings for best results: weight ratio 0.6, step size 50, SDXL checkpoints, and Euler A sampler. You can also combine this model with the third-party LoRA library LCM by setting the LCM LoRA weight to 1 and using the LCM sampler and 7 iteration steps. The model is based on SDXL. I also tested this LoRA using SDXL-Turbo checkpoint. However, due to SDXL-Turbo's shortcomings in generating high-resolution images, I recommend using a resolution of 768*768 or lower. Hope you like my model and share your feedback with me. Thanks.

kiana_kaslana_white_comet

I retrained LoRA with a smaller parameter size and now it only takes up 28MB of storage. Although the fitting is not detailed enough, perhaps some people need it.

Files with SHA256: edcb9b8a7eb9a8604dfdf6e079710a9d091eef2aa44b6e819a4dd7028bb5e0cb