search

PeterWang512 / CNNDetection

Code for the paper: CNN-generated images are surprisingly easy to spot... for now https://peterwang512.github.io/CNNDetection/
Other
826 stars 171 forks source link

Creating the same training set as used in the paper. #1

Open saksham-s opened 4 years ago

saksham-s commented 4 years ago

Is it possible to get a method to curate the exact same training set as used in the paper?

PeterWang512 commented 4 years ago

I haven't check whether my training set generation code behave deterministically (e.g., generating same images via different machines, or extracting the same lsun images from lmdb). Meanwhile, we're looking at potential releases of the training code and training set.

However, if you would like to train your own progan model, you can generate progan images directly from their repo. Moreover, you can also make lsun real images by downloading the lmdb files and process it using the create_lsun function from dataset_tool.py in the progan repo. This is exactly how we collected our training set and test set. However, if you make your own training set, you will need to create your own progan/stylegan/stylegan2 testset to avoid train/test leakage in the real images. Thanks!