search

tgxs002 / align_sd

Better Aligning Text-to-Image Models with Human Preference. ICCV 2023
https://tgxs002.github.io/align_sd_web/
Apache License 2.0
265 stars 9 forks source link

positive and negative sample data for LoRA training #9

Closed huxian0402 closed 1 year ago

huxian0402 commented 1 year ago

image Hello, I am very interested in your work and I am trying to reproduce your results. Would it be possible for you to share positive and negative sample data you used for LoRA training? @tgxs002

tgxs002 commented 1 year ago

Sure, but it will also be uploaded to onedrive.

tgxs002 commented 1 year ago

@huxian0402 Try it out: link. You may need to subsample half of the negative images to reproduce. Looking forward to your feedback!

huxian0402 commented 1 year ago

@huxian0402 Try it out: link. You may need to subsample half of the negative images to reproduce. Looking forward to your feedback!

Thank you for providing the training data. It was successfully downloaded. However, could you please clarify why this step was necessary? "need to subsample half of the negative images". I see that you used 37k positive samples and 27k negative samples for training. Are you suggesting that sampling should be done according to this ratio for better results?

tgxs002 commented 1 year ago

No, I used 37k positive samples and 21k negative samples. My experience is that using all negative samples will lead to worse results.