stability diffusion vore

[RoTheHuman]

Thanks for the link!

You don't need to train it yourself. Just replace your current model with another one.
Training your model is really hard both for dataset(100K+ tagged images) and hardware (20GB+ of VRAM)

The method I linked to apparently doesn't require making a whole new model yourself. It basically takes an existing model and adds more data on to it, which is (apparently) very accessible, even with a system that's not totally decked out. If I give it a try, I'll definitely report back here.

[RoTheHuman]

I found even more models to try out, including a big one trained on danbooru images and tags, another based off of that danbooru one but focusing on kink content, one based on rule34, and more. There's even one that seems to be specifically about anal vore! I'm not into that so I won't be tinkering with it, but if someone else does I'll be curious how well it works. https://rentry.org/sdmodels

If I get anything could out of them I'll share.

[RoTheHuman]

Well, this is the closet I've managed to come so far. Using the model based on images from danbooru or the "Zack3D Kinky" model definitely works better, and with some effort you might be able to get some all right mawshots out of it. But actual vore with someone being swallowed or digested is hard to do. I don't think people included much vore when training these models, so it doesn't come out easily. And a lot of the stuff I get is, not unexpectedly, a bunch of weird surreal anatomy all morphed and merged together in an unsettling way. But hey, maybe that's your kink!

Anyway, I think the trick will be to actual train a model on some vore images, complete with tags and all that! And not necessarily even a new model, but just adding to one of the existing hentai-focused ones would probably work well. Maybe I'll try out that anal vore model, just for the hell of it...

[jimpickins]

Currently exploring using dreambooth to train vore images. Still experimenting but I have managed less cursed images than normal SD produces. Long way to go still, but anyone saying "this cant do vore" is wrong

edit, here is a best result from my current model. still work to be done, but its 100% possible

[jimpickins]

Sharing my newest mawshot model for Stable Diffusion. Trained using dreambooth. You'll need to prompt it using "maw face". Face is the class and needs to be there to get good results. Its still not great but I think with clever prompting it is usable, excited to see what people can make with it.

https://mega.nz/file/5qogXApK
V5gK6tIxNtjq54sfGy2n0FEeRGh66hL5qUFhKgBRcqc

[taylorsm]

You can wrangle some nice internal scenes out of it with a bit of effort. Never managed to get the actual act of eating/swallowing with a mouth to be coherent though. At best, I've managed to fit a blobby monster mouth over the prey like a hat. See attached.

EDIT: I suspect it's possible if you iterate it the image between photoshop and the AI. For example, in the second pic, you can shop her upper half into the monster and leave the legs dangling, then tell the AI to repaint over that.

Spoiler'd because NSFW:

Spoiler: show

[jimpickins]

You can wrangle some nice internal scenes out of it with a bit of effort. Never managed to get the actual act of eating/swallowing with a mouth to be coherent though. At best, I've managed to fit a blobby monster mouth over the prey like a hat. See attached.

EDIT: I suspect it's possible if you iterate it the image between photoshop and the AI. For example, in the second pic, you can shop her upper half into the monster and leave the legs dangling, then tell the AI to repaint over that.

Spoiler'd because NSFW:

Spoiler: show

00248-3504280838-(((masterpiece))),best quality, illustration, straight body, (detailed face), {messy hair}, squeezed 1girl sliding down wet hole.png

00056-2916819604-(((masterpiece))),best quality, illustration, {{pose}}, full_body, (detailed_face_1.5), {messy hair}, 1girl, {touhou_project},.png

Do you have a prompt for the first one? Or was it done with lots of img2img?

[RoTheHuman]

Seconding a prompt for the first one. I tried tons of times to generate an internal image with no luck at all! Also, what model were you using?

[taylorsm]

I used NovelAI's anime model. No specific prompt. Recursively used img2img with different prompts until I liked what I saw. That specific image you two are asking about used this web image as the base (nsfw)(https://gelbooru.com/index.php?page=pos ... id=1968314). I manually flipped it 180 as a last step because the AI simply cannot do upside-down at all.

However, I find you can also prompt a non-vore base image yourself and do the same process. For example, I prompted this:

(((masterpiece))),best quality, illustration, full_body, (detailed face), {messy hair}, 1girl, {touhou_project}, (beautiful detailed girl), {beautiful detailed glowing eyes}, nude girl lying on pink fuzz with arms behind back, full body, on back, huge black aura

nsfw:

Spoiler: show

...and eventually img2img'd it into this:
nsfw:

Spoiler: show

Last iteration prompt was:

(((masterpiece))),best quality, illustration, straight body, (detailed face), {messy hair}, squeezed 1girl, restrained, (fetal position), (looking up), covered in sweat, looking up, detailed tentacles tunnel, nude, nipples, medium breasts, scared, open_mouth, arms_behind_back, arm_up, inside_tight_hole, dripping liquid, detailed_tentacles, (beautiful bright eyes), wet, partially_submerged, water slime, liquid dripping from above, crossed_legs,

So the AI is really flexible. The AI doesn't really understand "vore" or "internal". Instead, I find that variations of "tentacles" and "slime" are good for getting an organic appearance.

[taylorsm]

I set the denoising really high in img2img and somehow the girl got turned around, so have this great internal head-first view.
nsfw:

Spoiler: show

less sexy and more close up generation:

[jimpickins]

Found this which may help for posing and creating specific art

https://www.reddit.com/r/sdforall/comme ... to_direct/

[taylorsm]

Made a mawshot with the previous generation using mspaint and a lot of img2img and inpainting.

[taylorsm]

Better, more feral looking maw.

[taylorsm]

You really need use a ton of inpainting to make the AI drawings good. I touched up the previous pic and make the left-side arm more realistic. Had to regenerate it about 100 times lol. Also added collarbones, made her eyes more shocked-looking, and made them golden cause I like golden eyes. I think I'm done with this pic for now.

[jimpickins]

Is there any interest in starting a Stable Diffusion Vore discord server for sharing and researching this?

[whitewendigo]

Could be cool. Would be nice to not have to pay stupid amounts of money for simple vore art.

[taylorsm]

Sounds like it could be productive for sharing tips. I'm super new too this as well. Literally started playing around with it last week.

[ryanshowseason3]

Wanted to make some internals for my game. Took a while to get the hang of it and definitely used endoscopy stuff for image to image but eventually I got something new that looks reasonably good.

Only catch being a lot of my attempts were censored by dream studio for violating terms and services. Not sure how internal backdrops do that but ok.

What services are others using?

Filename has the tags used to generate the image.

[AlluringPredation]

NovelAI(Paid) or Local(4GB+ GPU), more info; https://boards.4chan.org/g/catalog#s=sdg

[taylorsm]

Tried a monster-vore cross-section this time. It was pretty annoying to get right. Pred's some kind of giant sea serpent thing.