Reproduce the results of the LNP method in the paper.

[wyczzy]

Thank you for your meticulous work, but I have a question. Using the pre-processing of LNP and the pre-trained weights of the classifier, it is not possible to reproduce the results of the LNP method in the paper.

[mmmn143]

@wyczzy I got the same result for LNP. For stable diffusion datasets, acc is very low (about 50-58). Have you found a solution yet?

[iamwangyabin]

same with you, i think its about dataset

[ksxdtla]

Thank you for your meticulous work, but I have a question. Using the pre-processing of LNP and the pre-trained weights of the classifier, it is not possible to reproduce the results of the LNP method in the paper.

"My LNP test is very slow. How did you solve this problem? Can you reproduce the weights test for other models mentioned in the paper?
If you have any good methods, please let me know.“

[iamwangyabin]

i find the reason, you need no crop & no resize & batchsize=1 for LNP.