LSUN dataset

[tedious-turtle]

Hi,

I'm currently trying to setup the training data according to your description. However, I struggle with recreating the subset which you used for the real_LSUN category.

The link that you provide leads to the LSUN github page with a download script for the "scene" categories (bedroom, bridge, ...). However, according to the names from your real_lsun.txt, I assume that you did not use the "scene" categories, but the "objects" categories (airplane, bus, ...) instead. Is this correct?

I then proceeded downloading the object categories from http://dl.yf.io/lsun/objects/ zip files, extracting the contained mdb database files. However, I'm now facing two issues:

1.) The image data in the mdb file is in .webp format, whereas your .txt image list refers to .png images. Did you just convert the .webp images back to .png and is it thus safe to assume that using .webp as the input images will yield the same results, or did you use some different image source with original .png files before lossy .webp compression?

2.) The .txt image list uses the category name followed by 5 digits (decimal) for the image identification -- unfortunately, it is not obvious to me how to map this 5-digit number to the 40-digit-hex key that is used for storing the image in the LSUN database. Can you provide more information on the mapping you used?

[RCorvi]

Hi, for the real lsun data you can take the images from the dataset of the 2020 CVPR paper from Wang et al.: "CNN-generated images are surprisingly easy to spot...for now".

[tedious-turtle]

Thanks a lot!