some question about the normalize operation

[yuese1234]

hello! Congratulations on your work ！I noticed that in your white-box attack code, the combined image was normalized again, as follows：
normalized_images = normalize(combined_images).to(torch.float16)
Isn't this equivalent to normalizing the original image twice?

data_dict = item_process_func(question, target_words, image_file, tokenizer, data_args)
This code has already normalized the image once

[pyogher]

Hello @yuese1234 , in our HADES white box images, we concatenate the black box images with the generated white box images. The item_process_func is used to process the black box images generated in the second stage, and normalized_images = normalize(combined_images).to(torch.float16) is used for normalizing the white box images. Therefore, the images will not be normalized twice.