All the example generations were done on the same seed, indicating good training.

P.S. I didn't describe the man and woman's appearance, because back then, the pose description was tied to a specific hair color or body type. You'll have to provide your own clues.

In theory, it would be possible to increase the dataset by 3-5 times with different hair colors/body types, but then the number of steps would have to be increased by 3-5 times, while maintaining the proportion of images per pose, which would most likely lead to unsuccessful training.