an embedding based on generated images
this embedding is trained on a small batch of generated images. Works best with 30-35 steps, Heun or DPM++SDE
Natnicha.pt