When preparing dataset to train a char lora, should you resize the image as per the training resolution? Or just drop high quality images in the dataset?