This is actually difficult to solve logically, because the model relies solely on black and white images to identify, for example, a person's clothing and hair color, which is already difficult to achieve. Taking the human eye as an example, it cannot judge what color the clothing and hair should be after converting black and white images to color. I tested a relatively plain style Chinese manga myself, and the consistency of characters was still very good. It may be difficult to maintain the consistency of characters in manga with a girl style and Korean style. Currently, I can think of two methods: one is to write accurate prompt words to maintain it, and the other is to refine a Lora model using the color images and animated works leaked from the manga as a dataset to achieve it