search

Ekko-zn / AIGCDetectBenchmark

189 stars 17 forks source link

Reproduce the results of the LNP method in the paper. #15

Open wyczzy opened 5 months ago

wyczzy commented 5 months ago

1710037867178 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 commented 4 months ago

@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 commented 4 months ago

same with you, i think its about dataset

ksxdtla commented 3 months ago

1710037867178 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 commented 2 months ago

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