And to add, if I just do pregnant, not aiming for pregnant with hyper preg (which I am not really aiming for anyway), prompt for photo and add negative prompts for anime, animation, etc. I still get "normalish" looking pregnant cartoons as the result.

(October 29, 2022, 10:39 am)bo27 And to add, if I just do pregnant, not aiming for pregnant with hyper preg (which I am not really aiming for anyway), prompt for photo and add negative prompts for anime, animation, etc. I still get "normalish" looking pregnant cartoons as the result.

make sure you're using the V1 model

(October 29, 2022, 6:48 am)subthresh15 yeah i had a look at them in the exif viewer, they're useful, ty. i tried a bit more with V2 to get photorealism, and it's sort of like, it doesn't look like a drawing per se, but it's definitely still obviously digital art/stylised rather than a "photo" of a hyperpreg woman in the real world. obviously that's tricky, because there aren't any "photos" of real hyperpreg women anywhere on the internet, let alone in a danbooru dataset. morphs are the closest we have, and a lot of them are pretty dodgy looking. approximating photographic textures from a danbooru dataset is always gonna make things look pretty stylized, even with strong negative weightings of art and digital art, etc. V1 is interesting because it can do properly "photographic" looking images, and also regularly get a decent-sized belly. but it's far less coherent/consistent with belly size/not deforming stuff than V2. i'd be super interested in a revamped V1. what was it trained on originally?

i think for now i might be stuck with inpainting if i want properly large bellies in V1 with a more photographic style. there seems to be an inherent tradeoff between belly size and photorealism, which makes sense considering there are no women of hyperpreg size irl, and hence nothing totally relevant in the training set.

My dataset has changed a lot since V1 as well.  V2 is a smaller cleaner dataset focusing on larger sized bellies, so the results are usually better overall.  If I get the time, ill retrain V1 on it to see if it turns out better at realism.  But your right that there are not many realistic hyper pregnant images to train on.

(October 29, 2022, 3:04 am)bo27 I can't seem to not get it to create an animated drawing, how do you get this to create photo realistic images?

Use the V1 model and use prompts like "photorealism", "realistic", "hyperrealism", and Negative prompts like "((digital illustration)), ((art))"
It's all about forcing the model to look more relaistic since most of the training data is digital illustraion.

After a lot of tweaking, I figured out you can combine stable diffusion models, giving it a certain weight to one another, so I took the hyperpreg v2 and combined it with a model I found. I sometimes will go back and forth between the models. Anyways, results are good. My video card is on fire.

(October 30, 2022, 4:11 pm)bo27 After a lot of tweaking, I figured out you can combine stable diffusion models, giving it a certain weight to one another, so I took the hyperpreg v2 and combined it with a model I found. I sometimes will go back and forth between the models. Anyways, results are good. My video card is on fire.

Nice!,  I spend to much time testing my new models I need to try out more realistic renders instead.

(October 30, 2022, 4:11 pm)bo27 After a lot of tweaking, I figured out you can combine stable diffusion models, giving it a certain weight to one another, so I took the hyperpreg v2 and combined it with a model I found. I sometimes will go back and forth between the models. Anyways, results are good. My video card is on fire.

which other model did you merge it with? was it zeipher f111 or something? looks good

https://huggingface.co/runwayml/stable-diffusion-v1-5

I got the 1.5 version, and I just mixed various different levels. So I created like 8 different model types, I explored til I got what I wanted.

If anyone wants a quick prompt to get started this one works pretty well with V1 (and now that im trying v1 again I see how far I've come with v2, lol)

Prompt: a (((huge hyperpreg pregnant))) girl in a tight sexy clothes. by elvgren, mucha, tarakanovich, rutkowski, trending on artstation, ((realistic)), (beautiful face and eyes), (outdoors), 30mm photo
Negative Prompt: (digital illustration), (art), ((((anthro)))), (((mammal))), (((text))), ((hooves)), cropped face, bad composition, two heads, long neck, missing limbs, deformed arms, deformed hands, extra fingers, (((extra legs))), (((extra arms))), ((poorly_drawn_face)), ((poorly drawn hands)), ((poorly drawn feet)), ((out of frame)), (((long neck))), (((tiling)))), ((bad hands)), ((blurry))

(October 31, 2022, 12:49 am)bo27 https://huggingface.co/runwayml/stable-diffusion-v1-5

I got the 1.5 version, and I just mixed various different levels. So I created like 8 different model types, I explored til I got what I wanted.

what values did you use? was it like 85% 1.5 and 15% V2? did you use weighted sum? looks great