>>40776823Yeah, I actually just wrote up a post about it coming to a similar conclusion.
>>40774536>>40774771I found and tested the other asanagi embed. I can kind of understand the complaint about it fucking with the likeness; it has a much stronger effect than mine does. For example, with my embed, I didn't have to specify that La+ is flat chested, but if I failed to mention that with the other asanagi embed, it would give me big, chesty La+ like in pic related.
There are four key differences in play here: the number of images in the training, the number of steps, the initial keyword, and the curation. Mine was trained on ~750 images for 50k steps using only 'girl' as the keyword, while the other was trained on 135 images for 150k steps and 'voluptuous' as the keyword. Finally, I suspect the other embed's training images focused on asanagi's more voluptuos girls, while mine included flatties and lolis.
The result? The old embed sticks harder to a certain style than mine does; it will generate outputs that are closer to asanagi's typical style, while mine is more flexible.
