There are some incremental improvements in Version 3.0.

More face variety, less of the "beautiful woman Flux Face" lady
Better response to different poses and camera angles with less weird anatomy.
Slightly bigger breasts without using extra prompts, allowing you to scale back if you want to.
Profile shots no longer completely unrealistic.
Unrealistic nipple bump artifacts are much more rare.
Red-haired and auburn-haired prompts generate more realistic results.
Used "topless" keyword on every training image with visible nipples. You can use that in your prompts if you want those results.
Used keywords "cleavage", "sideboob", and "skindentation" where appropriate. You can incorporate those words into your prompts and it will help guide the results. As always with AI images, that isn't guaranteed.

I added many more images to the training set bringing the new total to 376 images with more faces and poses included. To create the example images for this version, I ran a random prompt on auto-repeat for hours and made 247 images. Out of those, only 9 had bad anatomy ranging from messed up fingers to extra limbs.

I didn't work as hard at the NSFW quality this time, so honestly v2 might be a bit better for that.

Prompt tip: I found that hair color prompts will often override other prompts that cue about nationality, race, and skin. My random prompt below doesn't mention hair color for this reason. If you want unusual combinations, try asking, for example, "an asian woman with dyed blonde hair". I tried to use "dyed" in places where appropriate in training photos, which should make that a useful prompt term.

Overall, I think this version is more consistently high quality with more variety and flexibility.

Random prompt I used to generate the samples:

Photo from a __cameraangle-flux__ camera angle. A beautiful __skin__ woman is __action__ in a __cyberdelia/locations/cyber_rooms__. She is wearing __trendy_outfits__. Her hair is __hairlength__ and __hairstyle__. Cinematic high quality photograph.

BustyWomen - Flux