awsaf49 / artifact Goto Github PK

Thank you for make the dataset public, it is a really big contribution to the field!

When you generate these fake images, the ones from COCO-dataset, do you also save at some place the caption of the fake image generated? I can see it is not provided in the metadata but recent studies have shown captions plays a very important role in identifying such images and my research also revolves around image+caption embedding.

If possible can you also release the captions for these real/fake images?

I was reading the paper: https://arxiv.org/pdf/2302.11970.pdf

Still I am unsure, which folder belongs to which category. My guess:

Real:
"afhq",
"celebahq",
"coco",
"ffhq",
"imagenet",
"landscape",
"lsun",
"metfaces",
"cycle_gan"

Fake - GAN:
"big_gan",
"cips",
"face_synthetics",
"gansformer",
"gau_gan",
"generative_inpainting",
"lama",
"mat",
"pro_gan",
"projected_gan",
"sfhq",
"star_gan",
"stylegan1",
"stylegan2",
"stylegan3",
"taming_transformer"

Fake - Diffusion:
"denoising_diffusion_gan",
"glide",
"latent_diffusion",
"palette",
"sfhq",
"stable_diffusion",
"vq_diffusion"

Is that correct?

Hello, thanks for such excellent contributions for public ! I have some questions about the information of metadata.csv for every synthetic data folder.

target : The label for the image, which is either 0 for real or 1 for fake.

However, why 6 appears in several metadata.csv of synthetic datasets instead of 1 ? What does it mean? Besides, 3 is also exists in glide/metadata.csv.

Looking forward to your reply ! Thanks a lot !

Since the data is sorted by sources (and in some parks there are only real ones and in others only artificial ones), the question arises how to properly organize them for feeding into the model input. Is it necessary to shuffle the data somehow?

Hello, I want to ask will you open source code?
thanks a lot.