matthias-wright / art-fid Goto Github PK

Thanks for your excellent paper and well-organized codes. I have few questions about the experiment settings in the paper though.

In section 4.1 of the paper, you said that "Style images are sampled from the WikiArt dataset [70] and the BAM dataset [78]." and "we compute the ArtFID with samples of 50k images." and "For each style transfer method, the ArtFID is computed with 5 different samples containing 50k images each.".

I have few questions about the experiment settings.

1. Are the style images sampled from the same style in WikiArt or do you use many styles seperately and calculate the mean ArtFID?
2. What is the meaning of "50k images"? 50k content images or 50k style images? If it refers to the number of content images, then how many style images do you use?
3. What is the meaning of "5 samples containing 50k images each"?

Thanks for your excellent paper again and I am hoping for your response!

Hi experts,
The e.q.2 is sum of LPIPS and FID_infinite.

1. LPIPS between content and stylized images more higher is better, but opposite on e.q.2.
   ref:
   https://arxiv.org/pdf/2201.12543.pdf
   https://arxiv.org/pdf/2208.00921.pdf
   https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9244092
   https://arxiv.org/pdf/2006.01431.pdf

2. FID between style and stylized images is not good for distribution of holistic statistic.
   something like context and structure difference from style and stylized images.
   for example: stylized animal dataset and original style image from Monet
   SIFID from SinGAN in which use single images pair distribution
   https://arxiv.org/pdf/1905.01164.pdf
   IS_infinite: The Inception score is the expectation of KL divergence distance between two sets of generated images
   https://arxiv.org/pdf/2006.01431.pdf

Do you have been some experiments about these two metric functions?

Would it be possible to share the code used for training the inception network? It would be very helpful!

My e-mail address is [email protected]

Thanks very much!

Hi,
I tried "pip install art-fid" following your instruction, but it does not work. (screenshot uploaded)
Could you please check your version?
Thank you!

Would it be possible to share the code used for training the Inception network? It would be really helpful.
Thank you in advance, and thank you for your contribution.