search

jonasricker / aeroblade

[CVPR2024] AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error
https://openaccess.thecvf.com/content/CVPR2024/html/Ricker_AEROBLADE_Training-Free_Detection_of_Latent_Diffusion_Images_Using_Autoencoder_Reconstruction_CVPR_2024_paper.html
40 stars 6 forks source link

dectection for other datasets #3

Closed SleeplessBegonia closed 4 months ago

SleeplessBegonia commented 5 months ago

i get the distances of my datasets as distances.csv,i want to know how to use the distances.csc for detection?

jonasricker commented 5 months ago

In general, low reconstruction errors (high values in distances.csv, because we take the negative) indicate that an image is generated. For binary detection you need to determine a threshold based on your setting. For instance, you could compute the distances from a reference set of real images and select the threshold that correctly labels 95% of those images as real. This is what we did to compute TPR@5%FPR in Table 3.

SleeplessBegonia commented 5 months ago

thanks