What did you used for training target / tutor signal?

[nat-chan]

Dear authors,

Recently we planned to used your model to evaluate the illust generation performance in our paper.
But, I could not found what did you used for target in the training process.

Danbooru2020 tag["score"] or tag["favorite"]?

I am looking forward to your responce.

BEST regards,

[grexzen]

This model was trained using images from 3 sources that were scored by people, so the model uses image path and rating for retraining.

For just evaluating your gens, you can just embed all images, then run them through the inference py and get the score.

But not sure how that would evaluate your paper, unless it is about making gens look better pre or post process. Otherwise, it can only tell you how much the people that graded the original model this runs on would like them or not from 1-10.

Keep in mind this was trained on gens and no gens, so it does a poor job on gens only, since it over scores them down.