So I've been spending the last few days generating a bunch of images, with mixed success.

Some of it is overtly fetishy, some not. I was also trying to generate some images for a character in a fantasy 'verse I play around with, and some of those came out interesting too - most of the ones I was trying to make were supposed to be SFW, but the model I'm using insisted on a lot of topless shots. She's a priestess of a goddess of motherhood, so there are some pertaining to that facet. (This is the gal with box braids and beige robes who turns up from time to time - she's not really meant as fetish fuel, but hey, if the foo shits, wear it.)

I had some trouble getting it to generate a consistent number of people, and getting it to generate more than one person, where they're not essentially twins or at least look like sisters, is a real trick. Ages are weird - I usually didn't specify age (and when I did, it was either "21-year-old" to get younger-looking gals or "35-year-old" to get mature gals), but I found if I said "girl" rather than "woman" or "mother", it would randomly give me a kid, and I didn't want that. So after a few failed attempts, I consistently asked for groups of "women" or "mothers". (Except the gal with the box braids - I consistently called her a "witch" after some experimenting.)

I was able to do some labor/birth poses, but couldn't get it to show crowning or anything like that.

Also, this model likes making southeast Asian looking folks. My prompts never specify ethnicity (though I do sometimes specify "dark skin" - the black women are thus specified), so I'm not really sure why it's doing that, but "black hair" or "dark hair" seems to result in vaguely Asianish looking folks, while "golden hair" results in folks who look northern European.

EDIT: Removed the dupe.

More to follow later - I'm trying to convince it to put together certain scenes. Maybe that'll eventually pay dividends.

Some oddments!

Two of the most interesting, though not exactly the hottest:         

Gaining several cup sizes the day before delivery, apparently:     

Hot coochie!     

Not sure where this came from, less kink, but she's pretty:    

Talk about engorged!     

The action is clearly out of frame, but:     

Honk if you love amazons:     

Some one-piece love:     

Amazon lesbian mothers, because why not?     

She'll feel better once she pumps some of that milk out:     

Full-term sisters:     

And for those who like dresses:     

great images~

Which model/software are you using?

If you are using WebUI A111 or ComfyUI, it appears that most of the issues you are mentioning can be reduced or mitigated by prompting and some other extensions

Yeah, I'm using ComfyUI, and still figuring out how to make most of it chooch. I'm more surprised that it's worked as well as it has without having to do a lot more fiddling. I'm using the PhotoSomnia checkpoint and the Pregmix Lora for now, too. (I've tried blending in a few other Loras, with various results and haven't given up on that by a longshot.)

The weirdness to me is in figuring out what prompts will do what, under what conditions. Area constraints and prompt concatenation will probably get me at least a good chunk of the way to being able to consistently generate two *different* people in one image. Just from those I posted, there are plenty with multiple subjects, but they usually look pretty similar. There's a few where there're both blondes and brunettes - I still can't get it to do that consistently, but a few do come out that way. If I specify "dark-skinned" for one, everyone in the image will be, too, and usually the same skin tone and similar facial structures.

Also, using a name in the prompt has frankly funky results - it'll be consistently the same person, right up until you make a small tweak, and then it'll be someone entirely different! Learning what does what, under what circumstances, is half the fun! (and I've probably only produced about 1400 images total).

Currently monkeying with area conditioning and multiple prompts, with some interesting results. Alas, generation is slow as hell on my M1 MBP. Gonna look into some ways to speed it up, apparently there are some.

(August 7, 2023, 10:10 am)bigkatie Yeah, I'm using ComfyUI, and still figuring out how to make most of it chooch. I'm more surprised that it's worked as well as it has without having to do a lot more fiddling. I'm using the PhotoSomnia checkpoint and the Pregmix Lora for now, too. (I've tried blending in a few other Loras, with various results and haven't given up on that by a longshot.)

The weirdness to me is in figuring out what prompts will do what, under what conditions. Area constraints and prompt concatenation will probably get me at least a good chunk of the way to being able to consistently generate two *different* people in one image. Just from those I posted, there are plenty with multiple subjects, but they usually look pretty similar. There's a few where there're both blondes and brunettes - I still can't get it to do that consistently, but a few do come out that way. If I specify "dark-skinned" for one, everyone in the image will be, too, and usually the same skin tone and similar facial structures.

Also, using a name in the prompt has frankly funky results - it'll be consistently the same person, right up until you make a small tweak, and then it'll be someone entirely different! Learning what does what, under what circumstances, is half the fun! (and I've probably only produced about 1400 images total).

SD 1.5 based models are trained on lower resolution images, so their accuracy drops when the prompts become complex. SDXL works pretty good on generating multiple people with different faces based on simple prompt like ("a group photo of pregnant women"). It works on generating two people with different ethnicity as well (but does not work well when the prompt includes a third person).

For people name (like celebrity), in addition to LoRA, you may also try text embedding (it works well some time). Also, you may try to fixed the seed when you want to make a small tweak, without changing the face of the subject (I am using this trick to work on progression photos on SD 1.5 based models).

Hope that the above will be helpful!

So I'm trying to create a semi-specific scene, and some interesting stuff - less overtly kinky than just "hey, these gals are hot", but some interesting accidents nevertheless:

More stabbing in the dark:     
The gal on the left is prompted as being in a beige tunic, but the only trace is that wisp on her arm. The gal in back is less in focus but I can fix that.

This one has a bit of a "damn, girl!" vibe to it:     

This one came close to getting what I was looking for on the non-kink side of the house, but... no second gal.     
The second gal is actually supposed to be a midwife, which is why I'm poking at getting her to be not nude and not pregnant. (The top one was from later, as you can see, I got that *partially* working).

This one has some mild kink factor, at least, if you use your imagination:     

I like this one, but it's drifted so far as to be nearly off-topic - maybe if you squint you can tell the gal in the back is knocked up:     

This one is pretty close to the mark, actually:     

Is that Summer Glau?     
(Not specifically meant to be)

Twins!     

Again, close:     

Reading in, but I get a sort of "Damn!" vibe from this one too, and it's pretty close to the mark:     
The acolyte/midwife gal is wearing pretty much exactly what I had in mind.

This one is interesting:     

Some shooting in the dark came up with this:     
She's showing, if you look closely. I forget exactly what the prompt was for this one, but the result was wildly off the mark. Still hot, so I kept it.

Another barely-showing, but... maybe nursing?     

Another shot in the dark:     

The rest are a random smattering that fell out while trying to create one particular scene, that aren't it, but I liked anyway.

(August 8, 2023, 10:13 am)monyek

(August 7, 2023, 10:10 am)bigkatie Yeah, I'm using ComfyUI, and still figuring out how to make most of it chooch. I'm more surprised that it's worked as well as it has without having to do a lot more fiddling. I'm using the PhotoSomnia checkpoint and the Pregmix Lora for now, too. (I've tried blending in a few other Loras, with various results and haven't given up on that by a longshot.)

The weirdness to me is in figuring out what prompts will do what, under what conditions. Area constraints and prompt concatenation will probably get me at least a good chunk of the way to being able to consistently generate two *different* people in one image. Just from those I posted, there are plenty with multiple subjects, but they usually look pretty similar. There's a few where there're both blondes and brunettes - I still can't get it to do that consistently, but a few do come out that way. If I specify "dark-skinned" for one, everyone in the image will be, too, and usually the same skin tone and similar facial structures.

Also, using a name in the prompt has frankly funky results - it'll be consistently the same person, right up until you make a small tweak, and then it'll be someone entirely different! Learning what does what, under what circumstances, is half the fun! (and I've probably only produced about 1400 images total).

SD 1.5 based models are trained on lower resolution images, so their accuracy drops when the prompts become complex. SDXL works pretty good on generating multiple people with different faces based on simple prompt like ("a group photo of pregnant women"). It works on generating two people with different ethnicity as well (but does not work well when the prompt includes a third person).

For people name (like celebrity), in addition to LoRA, you may also try text embedding (it works well some time). Also, you may try to fixed the seed when you want to make a small tweak, without changing the face of the subject (I am using this trick to work on progression photos on SD 1.5 based models).

Hope that the above will be helpful!

I'm going to start messing with SDXL. Also, the idea of holding the seed consistent when I get a good one, or just one I want to tweak, makes a lot of sense.

A lot of the ones I just posted look like they're trying to be the same character, because they mostly are - I almost have the description dialed in that gets me someone pretty close, ish, most of the time... ish.

So I've been meefing around a bunch more, and came up with a few interesting ones.

Still working out how to get scenes with multiple people to come out right, and how to reliably suppress nudity when I want to. (Reliably creating it? That's easy as instant ramen.) SDXL is being much trickier, so I've fallen back to 1.5-based checkpoints for now until I have some time to sit and focus on that.

Here's a few SDXL ones that don't suck:             

One of the few SDXL spit out that felt actually close to what I asked for, though, look at those legs - this gal's gotta be at least 6'2". Which, hey, I like that, but this character is only supposed to be about 5'7", so...     

This is close-ish:     

Back to 1.5 - This one's not bad at all. In fact, for what I'm envisioning, it's close:     

Tried to do img2img, succeeded in knocking up the girl who's not necessarily supposed to be. Still, mo' belly, mo' better, belike.     

You could parse this as the gal on the right leaning forward to breathe through labor - I'm certainly going to interpret it that way!     

More labor-related, though the roles seem to be reversed - shouldn't the midwife be behind? (Also the midwife is pregnant too, but... hey, that can happen!)    

Better pussy detail, but she's barely showing. Cryptic pregnancy, I guess? But hey, she sure is top-heavy:     

A few more attempts at the "motherhood camp" idea I poked at earlier:         
These two are ready to drop!


This last bunch is one particular character - she came out surprisingly, but not perfectly, consistent. Kate her has a story - in this one she's supposed to be about 20 years old and about to deliver twins.
                                                                               

Finally, one of a few of another set - this is Kate too, albeit later, when she's in her 30s (I'd say maybe closer to 40 in this image). Not pregnant in this one, or at least not obviously so, and this particular attempt produced mostly not-pregnant not-porn, so hey, but I liked this one. I'd like to see if I can get this one to show up - or something like it - along with her now-20 daughters.