A lora for generating photos of human character mascot costumes.
There aren't many loras focusing on general human character mascot costume photos out there, and most models (including recent ones like Flux, Qwen Image and sometimes ChatGPT) requires quite a bit of prompting to get the style right, hence this is created to fill in the gap. Recommend to turn the strength up to more than 1 for more prominent effect. Tested to work with strength 1.2~1.5.
The dataset used to train this model comes from Hunyaun Image 3.0. A set of around 20 4:3 images containing one girl character human mascot costume (primarily Japanese anime style characters) performing various poses, wearing different outfit, with different hair style, and with different backgrounds. The images are first center cropped (to remove the watermark on the bottom right of course), upscaled by 4x_Ultrasharp, then downscaled to 1280x1280 for training.
The lora was trained based on Illustrious V1, and is observed to not perform well when paired with other base models such as haruka (however models finetuned from illustrious such as wai illustrious seems to work), and some details like plushy texture or photorealism may not come through when using incompatible models. Also, the model is not explicitly trained on images with more than one character, with characters other than female human characters, or weird poses or style, thus may underperform in those untrained circumstances.
As the training images came from Hunyaun Image 3.0, all licenses of the original model applies here. Please read the full article at their huggingface page.
