Thank you so much for sharing what you've worked so hard on! I wish I could wrap my head around the training side of AI. My own hardware isn't up to it and I'd rather spend the little spare time I have for this hobby making pics, than playing with a cloud service without knowing what I'm doing. 

---

I can't train but I have been playing around with the available models. For those who haven't played around with model merging: The "add difference" option in automatic1111 lets us add the effects of a training to other models without the dilution of the original model. Here's a demo using a 50/50 merge of Analog Diffusion and Elysium v1, merged at 0.5 weighted sum with your model, then merged with the added difference between your model and SD 1.5. Cherry picked because the output wasn't too bad to begin with.

"young woman posing by the lake, short tight denim hotpants and fashion print bikini, pregnant belly" Steps: 28, Sampler: Euler a, Seed: 523726071, Size: 512x640, no negative prompt, CFG at top row:

   

Even without LadyBabyloin's tokens (keywords) there's more of a belly and more detail on that belly, for the add difference merge compared to the weighted sum merge. Here's where the magic happens:

"young woman posing by the lake, short tight denim hotpants and fashion print bikini, pregify belly"

   

Yes there's imperfections but that's a part of AI fake photos, esp. with the ancestral samplers. The difference in detail is clear esp. with the texture of the skin. It's not hard to tidy things up with a bit of editing and use img2img for upscaling, to get something that looks relatively natural. 

   

Here's the same prompt/settings using just LadyBabyloin's Preg Teen model:

   

---

The add difference merge effect also seems to work positively with the Anime-style models, such as Anything 3.0:
   
"young woman posing by the lake, sexy school uniform, pregify belly, youthify, teenify"
Steps: 28, Sampler: Euler a, CFG scale: 9, Seed: 523726071, Size: 512x640

This works for other models that have been custom trained, if you know which model was getting trained (and usually this information is available). In the webui's Model Merge tab, Have the model you want to modify as A, the trained model as B, and the model it was trained from as C.

---

Thanks again LadyBabyloin, your pregnant teen model can do a great job of enhancing pregnancy anatomy and details for almost any outputs. Love that profile pic too, you have amazing talent.

Thank you for all your kind words, and your example!

Here's some test shots from my next step, day job has me busy so IDK when it'll be solid

They look amazing

(January 8, 2023, 3:58 am)LadyBabyloin Thank you for all your kind words, and your example!

Here's some test shots from my next step, day job has me busy so IDK when it'll be solid

What kind of prompts are you giving for the first and last pictures? I really like this last set you have with the new stuff here, absolutely my most favorite. I have quite a bit (kind of) like this already made/generated I could send you that I've fine-tuned a bit, if those would help with your model training? Though they are not actual photos, and AI generated ones. They are pretty close to these pictures, but I had to go through a lot of img2img and txt2txt and photoshop to get them to look good. So they likely would be suitable for AI training also.

Try doing a merge difference with protogen and the preg checkpoints and adding Elf into the mix. Get some pretty interesting results that way as well. Not sure if I should post pics as it seems that belly buttons and hands tend to get mixed up a bit sometimes. Otherwise you get some pretty young looking Elves with preg bellies.

Prompt looks like this:

Positive prompt : A digital picture of a (Pregnant Elf girl with long pointy ears with a (young body that is teenify), (her beautiful young face is youthify), (her large round belly is pregify))

Negative prompt: disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, mangled, old, surreal, unsymmetrical, asymmetrical, text, error, cropped, jpeg artifacts, worst quality, low quality, signature, watermark, censored, Pixelated

Why do all my pics keep coming in blurry using the model? What's the fix?

(January 15, 2023, 10:21 pm)Aggression Why do all my pics keep coming in blurry using the model? What's the fix?

Depends on a few things, what graphics card are you using, is your stable diffusion up to date? what settings or prompt are you using, are you using a merged checkpoint or stock checkpoint that was posted in this thread? If my picture shows up, settings "should" be something like thats in the picture for best results that i have noticed.

Heya, was this the last iteration of the model? Was super curious if it had been finished.

Kinda curious about this too if the last post was the final version. I know there are LORAs that do the same where you can take a realistic model of people and convert them into pregnant pics. However, that might be too much downloading for some. CivitAI is where i got my models from if anyone was curious.

LoRA, LoHA, and LyCORIS have (IMO) totally outstripped finetuning an entire model in both efficiency and quality. Therefore, if I do release another model, it will likely be a merger between this one and a few custom LyCORIS models, but even that is pretty unlikely honestly.

That said, I had a bit of obsession with LLM chatbots on my off time for a while, and adjusting to using AI gen at work took some of the joy out of it. But I've been working with Stable Diffusion and ControlNet again at home lately, and if I produce a good related LyCORIS I'll certainly post it. I also cannot recommend ControlNet inpaint enough for making morphs or touching up generations.