About reproducing

[liujikun]

Hi,

Thank you for your great work! Have you ever cleaned the lhq dataset? I used the lhq_1024_jpg dataset to reproduce the effect, and the FID can only reach 9. I tried to fine-tune on the open source model, but the initial FID was as high as 24.

Best,
JiKun

[universome]

Hi! I am sorry for answering that late. Yes, we indeed preprocessed the dataset (as specified in the Experiments section) with the procedure described in Section 3.3/Appendix C/Algorithm 1.

Here is the script that we used to preprocess the datasets (we use a threshold of 0.95 for LHQ).

Also note, that in our Table 1, we provide results for the 256x256 resolution, not for 1024x1024 (for 1024x1024, our model achieved FID/∞-FID of 10.11/10.53).

That's strange that the initial FID was so high. But maybe those "unconnectable" images (that are removed by our preprocessing procedure) really change the distribution so much.

[universome]

Hmm, the currently released dataset has images sorted by their InceptionV3 likelihood (from least to most probable). I think in the above script, one should also shuffle the images. I will update it

[Muccul]

@universome Hi, some confusion about your jupyter of preprocessing.
LHQ256 dataset contain 90k data, and rest contain 70k after split(15k train, 5k val), but in your jupyer file, why filtering number is 167k with 0.99 threshold?

[Muccul]

@universome Hi, some confusion about your jupyter of preprocessing. LHQ256 dataset contain 90k data, and rest contain 70k after split(15k train, 5k val), but in your jupyer file, why filtering number is 167k with 0.99 threshold?

btw, if we get a subset of data, How to divide it into training set and test set ? and calculate fid just on test set in subset for your paper result?
Looking forward to your reply, thank you!