Training image samples: samples_scaled_gs vs samples_gs

[SeverianVoid]

Can you explain what the difference is between the samples_scaled_gs images and the samples_gs images produced during training? In all my attempts to train a model the samples_scaled_gs images seem have the correct structure of the my training images and the samples_gs just seem to be a lot more random maybe just the style of the training set I am not sure. But all my attempts to use the trained embedding also seem to more closely match the samples_gs random images even though the samples_scaled_gs images generated during training look pretty good.

[1blackbar]

rinongal/textual_inversion#19