The reproduced results differ significantly from the results reported in the paper.

[Achestnut00]

Dear Author,

I have successfully run makeup_transfer.py following the instructions in the README file. However, after training on CelebA-HQ, the ASR in FAR@0.01 differs by more than 20 points from the results reported in the paper. Could you please advise on how to address this issue to successfully reproduce the results from your paper?

Thank you!

[HansSunY]

Thank you for your kind words and interest in our work. In our publicly available code, we reduced the strength of attacks. If you’re aiming for better attack performance, you can try running the following command for enhanced results: python main.py --makeup_transfer --config celeba.yml --exp ./runs/test --do_train 1 --do_test 1 --n_train_img 200 --n_test_img 100 --n_iter 6 --t_0 60 --n_inv_step 20 --n_train_step 6 --n_test_step 6 --lr_clip_finetune 4e-6 --MT_iter_without_adv 2 --MT_adv_loss_w 1.2 --model_path pretrained/celeba_hq.ckpt --target_img 3 --target_model 3 --ref_img 'XMY-060' This configuration result in stronger adversarial attacks, although it may impact the visual quality of some of the generated images.

…

-----原始邮件----- 发件人:Achestnut00 ***@***.***> 发送时间:2025-01-05 23:07:22 (星期日) 收件人: HansSunY/DiffAM ***@***.***> 抄送: Subscribed ***@***.***> 主题: [HansSunY/DiffAM] The reproduced results differ significantly from the results reported in the paper. (Issue #8) Dear Author, I have successfully run makeup_transfer.py following the instructions in the README file. However, after training on CelebA-HQ, the ASR in ***@***.*** differs by more than 20 points from the results reported in the paper. Could you please advise on how to address this issue to successfully reproduce the results from your paper? Thank you! — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Achestnut00]

Your response has successfully resolved the issue we previously encountered, and we have successfully reproduced the attack results described in the paper.
However, when attempting to reproduce the FID, SSIM, and PSNR metrics, it seems that Table 2 in the paper does not specify how these values were obtained or which model was attacked to derive these metrics. Additionally, the provided code does not appear to include the calculations for these evaluation metrics.
We would greatly appreciate it if you could clarify our doubts regarding this matter.

Thank you!

[Achestnut00]

[signifance]

Hello, I am running this code
python main.py --makeup_transfer --config celeba.yml --exp ./runs/test --do_train 1 --do_test 1 --n_train_img 200 --n_test_img 100 --n_iter 6 --t_0 60 --n_inv_step 20 --n_train_step 6 --n_test_step 6 --lr_clip_finetune 4e-6 --MT_iter_without_adv 2 --MT_adv_loss_w 1.2 --model_path pretrained/celeba_hq.ckpt --target_img 3 --target_model 3 --ref_img 'XMY-060'
An error occurred -- MT_adv_loss_w 1.2 is invalid