Hello,
Thank you for sharing the code and your great work on "Edit Away and My Face Will not Stay: Personal Biometric Defense against Malicious Generative Editing." I tried to reproduce the results for learning and research training purposes using the public code in this repository.
Specifically, I ran experiments on 2,000 validation images from the CelebA-HQ dataset
, setting the hyperparameters according to Appendix Table A1 and Table A2 in the paper.
However, my reproduced results fell short of the performance reported in the paper. For example, while Table 1 reports a CLIP-I score of 0.648, my reproduction yielded approximately 0.82. Similar gaps appeared in other evaluation metrics as well.
Could you clarify if there are any additional details (e.g., preprocessing steps, random seed settings, normalization strategies) that might be necessary to match the reported results?
Thanks a lot for your time and for making the code available!
Hello,
Thank you for sharing the code and your great work on "Edit Away and My Face Will not Stay: Personal Biometric Defense against Malicious Generative Editing." I tried to reproduce the results for learning and research training purposes using the public code in this repository.
Specifically, I ran experiments on 2,000 validation images from the CelebA-HQ dataset
, setting the hyperparameters according to Appendix Table A1 and Table A2 in the paper.
However, my reproduced results fell short of the performance reported in the paper. For example, while Table 1 reports a CLIP-I score of 0.648, my reproduction yielded approximately 0.82. Similar gaps appeared in other evaluation metrics as well.
Could you clarify if there are any additional details (e.g., preprocessing steps, random seed settings, normalization strategies) that might be necessary to match the reported results?
Thanks a lot for your time and for making the code available!