Krea-2 version

80% same dataset.

requires more prompting in human language, simple tags work poorly.

ZIT version

Trained on 247 images.

For women in panties, use lower values ​​(0.5 - 0.7) since most of the dataset is without panties.

For a lower viewing angle, it is better to use a landscape aspect ratio, and for a high angle, a portrait one. (tag words: low angle, mid angle, hi angle. But this doesn't work very well.)

Trained on comma separated tags.

Works on base checkpoint.

If a girl is lying upside down, try adding "in the foreground is the huge ass of a lying girl."