Second iteration (the first, frankly, was bad).

Removed some lower quality image from the datas, added some other, splashed with personal generations, re-tagged everything, smaller tensor size, increased shadowing.

From tests, this seems a good generation improver, oriented toward a less-animesque style: the cover image is a simple prompt with high CFG to show the baseline.

As I forgot '2D' and '3D' as tags, it doesn't particularly react to that disctintion anymore. No big deal, honestly: you can prompt a 2D generation anyway with the right prompt.