thanks for these great work!

Heres some more I think my model is getting pretty good.

Please please post the model when you think its done, if you can't because its a stupid license issue ((those shouldn't be a thing with diffusion models imo)), then can you tell us the mix, and maybe post the photos you trained with? I'd love to use it with the amazing new hyperpreg lora that was just released.

Only thing it needs imo is stretch marks, the linea nigra, and bigger outies.

awesome! it would be cool see pregnant girls having contractions in different places, i don´t know maybe a pregnant teacher suddenly having contractions in class or a girl in a party hahaha it just an idea XD i don´t know this is so cool, man! good job

(March 9, 2023, 8:24 pm)Justaperv Please please post the model when you think its done, if you can't because its a stupid license issue ((those shouldn't be a thing with diffusion models imo)), then can you tell us the mix, and maybe post the photos you trained with? I'd love to use it with the amazing new hyperpreg lora that was  just released.

Only thing it needs imo is stretch marks, the linea nigra, and bigger outies.

It can do all of that, yeah, I don't think I will be posting my model for kind of the reasons you put, with or without using custom models just don't want to be involved in all of that. Because I did training, no matter how many models I mixed, pregnant navels were just way off, and they did not have a good breakdown of what we in our community like to see. I will say this, my best results came recently when NOT using Dreambooth, but instead using EveryDreamTrainer2: https://github.com/victorchall/EveryDream2trainer. It uses a different method than Dreambooth which I believe is way more effective, instead of taking what you present and "merging" it with what the model already knows, it just trains the new model.. I think there is a breakdown, but the Github is VERY detailed, and its surprisingly easy to operate. I also did not just dump photos into a folder and do "token" and "class", I literally captioned every single picture, I used a problem to use BLIP 1.4 captions to assist with getting 65% of the way there, but still I verified each file til I was complete, adding my own phrases, removing other phrases, and trying to catch captions that were completely inaccurate or wrong. I would include different types of pictures, so if a woman for example had a great body and ugly face, I would tag "unattractive face", "sexy pregnant body",I had about 800 images so it took several weeks doing a little bit at a time, but I think the results speak for themselves. I still think I need to do more work on in certain settings, I think its kind of skewing my results there, but everything else is looking pretty damn solid. I can put in twins, triplets, quadruplets, 20-42 weeks pregnant, small-huge belly, tiny-huge breasts, heavy breasts, lactating, stretchmarks, no stretch marks, "suddenly pregnant", underbelly, prepregnancy clothes, etc. I also captioned "female hands", "accurate skin complexion", "detailed skin", etc. Everything I could to get rid of the problems I see with other models. So there is a long list of keywords that you can do if you use captions. The other problem is I am so damn picky, I keep improving it, and as long as I can keep improving it.. the model isn't done. Here are some examples of pregnant women with stretchmarks and linea nigras, just kind of picked examples, so if you see goofy hands etc, I just picked them fast.:

That looks awesome, I'll definitely look into that plugin instead of dreambooth. Could tell us the models you merged? For the hand issue theres a negative embedding called bad-hands-5.pt.

Found it. https://huggingface.co/yesyeahvh/bad-hands-5/tree/main Of course as is normal with stable diffusion, it doesn't solve the issue all the time, but it definitely helps.

Do you think it would at least be possible if you made a lora or textual inversion for the stretch marks, lineas, and belly buttons? The hyperfusion model that just came out is excellent at making large bellies, even if they aren't as realistically shaped as the ones here. But it lacks those skin details since it was made for anime.

That's an amazing model @bo27. It blew me away when you started it and it just keeps getting better. I can understand the hesitation for releasing it, the AI art community has a lot of good things going but also gets weird in some ways and once a model is out, it's out. 

Thanks for the dreambooth alternative recommendation.  I should probably start doing some training of my own, I actually kind of like the idea of going through and tagging a few hundred selected photos for a training dataset. I think I've pushed photorealism as far as I need to with model blending (I think I'll post it on civitl as a new model rather than update Gravy Mix, once I've produced some demo outputs) but realistic belly details are still sorely lacking esp. proper flattened/pushed-out bellybuttons and linea nigra. I did dabble with textual inversion months ago but the results were poor, probably due to insufficient tagging and not really knowing what I was doing but at least I could run it locally. The results you're getting are inspiring me to try again, hopefully the free colab will get me started. All the instructions for training I've seen are for either an individual or for a style, not for a body type or specific features but I assume that a well-captioned dataset will be a good start, and the documentation for EveryDreamTrainer2 looks way easier to understand than the other guides I've seen. Perhaps the hard part will be taking a break from doing AI and going back to just scouring the web for awesome preggo pics, just like the good ol' days  

Quote:The other problem is I am so damn picky, I keep improving it, and as long as I can keep improving it.. the model isn't done. 

I know that feeling, just blending has that effect, but for once I'm satisfied with how far things have come and the outputs I'm getting. From the sound of things AI-assisted model merges are eventually going to blow my efforts out of the water anyway, and the hardware to try and run that sort of arrangement is outside of my budget, so I think I'll be skipping on trying to improve the blend for a while and just focus on making pics, and getting started with AI training.

Loving the ones with like 3 pregnant hotties in the bar. Those women are looking very... blessed and fruitful.

(March 10, 2023, 7:46 am)GravidaSomnia That's an amazing model @bo27. It blew me away when you started it and it just keeps getting better. I can understand the hesitation for releasing it, the AI art community has a lot of good things going but also gets weird in some ways and once a model is out, it's out. 

Thanks for the dreambooth alternative recommendation.  I should probably start doing some training of my own, I actually kind of like the idea of going through and tagging a few hundred selected photos for a training dataset. I think I've pushed photorealism as far as I need to with model blending (I think I'll post it on civitl as a new model rather than update Gravy Mix, once I've produced some demo outputs) but realistic belly details are still sorely lacking esp. proper flattened/pushed-out bellybuttons and linea nigra. I did dabble with textual inversion months ago but the results were poor, probably due to insufficient tagging and not really knowing what I was doing but at least I could run it locally. The results you're getting are inspiring me to try again, hopefully the free colab will get me started. All the instructions for training I've seen are for either an individual or for a style, not for a body type or specific features but I assume that a well-captioned dataset will be a good start, and the documentation for EveryDreamTrainer2 looks way easier to understand than the other guides I've seen. Perhaps the hard part will be taking a break from doing AI and going back to just scouring the web for awesome preggo pics, just like the good ol' days  

Quote:The other problem is I am so damn picky, I keep improving it, and as long as I can keep improving it.. the model isn't done. 

I know that feeling, just blending has that effect, but for once I'm satisfied with how far things have come and the outputs I'm getting. From the sound of things AI-assisted model merges are eventually going to blow my efforts out of the water anyway, and the hardware to try and run that sort of arrangement is outside of my budget, so I think I'll be skipping on trying to improve the blend for a while and just focus on making pics, and getting started with AI training.

Sad to hear you're taking a break, though the time off I'm sure is sorely needed, and probably way more important than generating porn. But I wanted to pop in and say your work has been hugely inspirational, as someone who is just now starting to get deeper into stable diffusion and image generation, and the models and LoRA that currently exist have been immensely helpful though not quite to the level I'd like them to be, and not for the subjects I want to do with them (as it's all stuff related to though not explicitly preg and hyperpreg), so seeing your work here has been a godsend for reference and inspiration. I'd kill to get a copy of the model you're using! But regardless, hope you come back refreshed, rested, and inspired.

(March 9, 2023, 9:42 pm)bo27

(March 9, 2023, 8:24 pm)Justaperv Please please post the model when you think its done, if you can't because its a stupid license issue ((those shouldn't be a thing with diffusion models imo)), then can you tell us the mix, and maybe post the photos you trained with? I'd love to use it with the amazing new hyperpreg lora that was  just released.

Only thing it needs imo is stretch marks, the linea nigra, and bigger outies.

It can do all of that, yeah, I don't think I will be posting my model for kind of the reasons you put, with or without using custom models just don't want to be involved in all of that. Because I did training, no matter how many models I mixed, pregnant navels were just way off, and they did not have a good breakdown of what we in our community like to see. I will say this, my best results came recently when NOT using Dreambooth, but instead using EveryDreamTrainer2: https://github.com/victorchall/EveryDream2trainer. It uses a different method than Dreambooth which I believe is way more effective, instead of taking what you present and "merging" it with what the model already knows, it just trains the new model.. I think there is a breakdown, but the Github is VERY detailed, and its surprisingly easy to operate. I also did not just dump photos into a folder and do "token" and "class", I literally captioned every single picture, I used a problem to use BLIP 1.4 captions to assist with getting 65% of the way there, but still I verified each file til I was complete, adding my own phrases, removing other phrases, and trying to catch captions that were completely inaccurate or wrong. I would include different types of pictures, so if a woman for example had a great body and ugly face, I would tag "unattractive face", "sexy pregnant body",I had about 800 images so it took several weeks doing a little bit at a time, but I think the results speak for themselves. I still think I need to do more work on in certain settings, I think its kind of skewing my results there, but everything else is looking pretty damn solid. I can put in twins, triplets, quadruplets, 20-42 weeks pregnant, small-huge belly, tiny-huge breasts, heavy breasts, lactating, stretchmarks, no stretch marks, "suddenly pregnant", underbelly, prepregnancy clothes, etc. I also captioned "female hands", "accurate skin complexion", "detailed skin", etc. Everything I could to get rid of the problems I see with other models. So there is a long list of keywords that you can do if you use captions. The other problem is I am so damn picky, I keep improving it, and as long as I can keep improving it.. the model isn't done. Here are some examples of pregnant women with stretchmarks and linea nigras, just kind of picked examples, so if you see goofy hands etc, I just picked them fast.

Incredible stuff, your work through this whole thread keeps blowing me away. I absolutely can't wait for this model to be shared whenever you feel like it's ready. It's such a fight to get any model as of now to do the sort of things I want, but this will go a long way to making it better. Ever considered making it into a LoRA rather than a checkpoint or a merge?