Trained with Leco score (https://civitai.com/articles/5416)

The dataset is 5k images that have the most reactions on civitai (positive label) versus 10k images that are randomly chosen (negative label).

Weight 0.2~1.5

Remarks:

• 1.5 is not the best weight, pick a good seed, and then try weights from 0.3 to 1.5 (they all give different results)

• trained on AnyLora with LecoScoreV2 for 3000 steps along the PartiPrompts

• training a score on this dataset is unstable (because the difference between very good images and ok images is very small, the network can pickup different characteristics of trending images each time)

• the criterion (trending images) is bad. The main reason for this is that it can indicate a large number of followers or a very 'mainstream' theme for the image. I would prefer the number of positive reactions per view (the eye with a number when you click on an image), but this would be very noisy as well.

• dataset is included (if you want to train something on it)

• example images are sometimes rendered within the Lykon universe (this was trained on AnyLora, then use Dreamshaper, Fast Negative, Easy Negative), I had to remove the Lykon negatives for other models

• -> alpha version