So I've been messing around with Stable diffusion a lot and all the different models out there including NAI, F222, 1.5, the hpreg one on here, and some other mixes. I tend to lean towards realism style images and am currently testing the waters training yet another model on some images from the site. I'm not expecting it to work all that great as I've found that even though variety of photos helps to train new AI models it needs consistency in the focus. IE. Using photos of a woman with different makeup causes artifacting around the eyes etc... Anyways what I'm hoping to do is find a model either willing to help me out by getting a simple variety of photos while maintaining some consistency (Belly shape, belly button, body, etc) Either that or to find a good trove of a model to dig out pictures from. I'm ideally thinking nude but would like to train it on clothed photos as well after some experimentation. Unfortunately I am not in the position to pay for these photos but am completely willing to upload the model here after I work on it and get it to an adequate state.

Here is a list of what I am trying to accomplish:
-Better shape and size, Ideally a full term average pregnancy or multiples. Aiming for realism and not absurdity
-Consistency with belly buttons and anatomy of pregnant bodys as even the best anatomy models seem to struggle when it comes to pregnancy
-Want to start work with naked full term pregnancy and move from there to clothed and even first and second trimester concepts
-Any other recommendations will be taken into consideration.

Also of note- Face is not required at all for any photos. Actually it would be ideal not to have a face in most of them.

I am fairly new to this but have experienced success in training concepts before. Thank you for any help and feel free to message me/post here if there is anything I have missed. Depending on how this current model turns out I may wind up posting it for others to use.
I will also post generations depending on how the current model turns out.

This first take on the model has honestly surprised me. Definitely requires further work but with very simple prompting with facial restoration enabled at a low CFG I got some pretty astonishing results. I have already noticed the belly button dilemma is completely gone though there are strange deformities that can be further improved upon. Here is a first look at a batch of AI generated pregnant woman with very little effort.

Notes
-unprompted selfie-style images are common.
-Hands tend to wind up misproportioned and in strange places.
-Very obvious influence in most bellies from the photos I used

Amazing outputs, and this is only the beginning. While vanilla SD can do passable photographic realism, all too often the details such as the pregnancy bellybutton or other subtle things just don't appear, and if they do they don't seem to upscale well using img2img. Anything generated beyond 512px height starts to get real weird real quickly. These are already so much better than that.


I've also noticed that any model trained on actual photos seems to improve the probability of better hands.

(November 17, 2022, 2:50 am)GravidaSomnia Amazing outputs, and this is only the beginning. While vanilla SD can do passable photographic realism, all too often the details such as the pregnancy bellybutton or other subtle things just don't appear, and if they do they don't seem to upscale well using img2img. Anything generated beyond 512px height starts to get real weird real quickly. These are already so much better than that.


I've also noticed that any model trained on actual photos seems to improve the probability of better hands.

Exactly, This is already a serious improvement over what most SD models understand as 'pregnancy' which tends to just be adding a round shape to someone's torso or occasionally thigh. There are definitely still points at which this loses the belly button which I'm going to work on. As well as being able to determine the body surrounding it but that's going to require a lot more effort and source photos. Right now I tend to get large tits with a large belly which is great but I'd love to be able to fine-tune it and get smaller breasts and different size bellies. Been messing around with merging it with f222 which is trained on anatomy(I think more or less just a porn model) but haven't gotten too lucky. Anyways here are some more images

very interesting. how many images were in your training set? been thinking about doing my own model, but unsure of how much finessing it requires. was this done using dreambooth?

(November 17, 2022, 8:24 am)subthresh15 very interesting. how many images were in your training set? been thinking about doing my own model, but unsure of how much finessing it requires. was this done using dreambooth?

This was done in the google collab with 53 images. I've moved to A1111 dreambooth UI now. 6000 steps for training. The main thing that takes finessing is nowhere which photos to choose as that can cause a lot of issues. I'm hoping to get a slightly more refined model today and then to train in on other concepts such as clothes. I also have a 3090 in my own setup so training this at home is an option versus the google collab which can be frustrating. Will probably be uploading more images soon after the current dataset trains

Here is a couple images from the latest iteration. Experimenting with merging a base model with one to help specify more on the looks of the person as well.

Hey, I trained the hyper sized models if you have any questions about training feel free to ask. I've lost count of how many I've trained at this point.
Glad to see a more realistic model too. The default belies in SD were pretty disappointing, but I found it didn't take too much training to fix up the belly buttons. The fingers attached to the bellies are a much harder problem XD

My dataset is all CGI and Digital art, so it was difficult to keep prevent the model from drifting away from realism, but found that you can still get away with lots of synthetic data before it got too un-realistic hence the v1 and v1.1 models