Actually, I trained a haps style model to generate this character, whom I really like. However, due to the insufficient number of its images in the haps Lora training dataset, the model failed to converge properly. As a result, I retrained a new Lora and took the opportunity to test high-resolution parameters during the training process.

Negative prompt :

unfinished, monochrome, greyscale, signature, url, artist name, username, text, watermark, English text, bad quality, worst quality, worst detail, sketch, censor,