The First Sample Image Is Amazing
#1
by
jrrjrr
- opened
The contrast in the first sample image between the illustrated character and the photo-realistic background is remarkable. It feels like the model understands foreground vs background. I think I remember some model trainer write about using a different weight emphasis between the center and the periphery of an image. Maybe that is how this model works? Favor weights from illustrations for the center and weights from photographs for the edges? It is a great effect that I never noticed in a model before.
Perhaps the reason is that images similar to the URL below were used for fine tuning.
https://www.pinterest.jp/hakoniwa852wa/
But I think it's also possible that she's using another new idea.