So, here's something I haven't seen any posts about, and I apologize if this actually goes into draw, but... is it possible to train a model to understand inflation proportions? I've seen lots of models about anime and the like, but I'm curious what it would take to really train something given the further outside the 'norm' something is the harder it would be to train.

I imagine you couldn't use a textual inversion embedding; hypernetworks could be used to try and nudge something towards a mean I suppose. Dreambooth might be better but idk how good it would be?

Like, I could see how one would train it to do slightly above average sized breasts. But what about truly massive ones? Or blueberries or the like? Could this be possible? I feel like I'm seeing enough of the art on this board that someone else has to have considered this as something worth doing.

I suppose as well, if nothing else, this could be a thread to demonstrate how far you've pushed the models. We can all find things that fit into a normal size range anywhere, but I feel like we inflationists should be aiming higher than that.

It's not incredibly difficult to train a model. It's just really time consuming if you're doing the entire training by hand. You need to start with a significant amount of images and label them with appropriate descriptors. Luckily once you start to get good results, you can shift over to training with generated data and auto populated labels but that can produce shoddy results if not trained enough manually. All in all, I've been looking into the process while using DA's image bank and tags for labeling but haven't got the patience to sit down and label possibly thousands of images.

>>69232 (OP)
https://github.com/balloonboi/balloon-diffusion

>>69249

that's a good link! I am curious however what KINDS of images he's training it on; and I eagerly look forward to 1.2.

>>69245

of this, I'm certainly well aware. I've trained models before, but I've never managed to train them on anything that requires a great deal outside of the norm.

For example, I wouldn't know where to start to train, say, a spherical body, because I wouldn't know what terms to use for it to understand what that sort of thing means.

One thing that may be of use to you if you're using anime images is deepdanbooru, which is an autotagger that has an extension in automatic's repo; that certainly makes autotagging faster though as you said, it can vary in terms of accuracy. I can say that if you're using NAI's model at least, or waifudiffusion, the danbooru tags will work. Not sure with other models.

>>69249

apologies on a double post, but I did some digging into the model link posted, and it's worth looking into.

However, you can already get better results with everythingv3 and the NAI leak, though I suspect this is because the maker of balloon diffusion doesn't list what terms he used to tag the various images. If it's based on waifudiffusion like he claims, then it's using danbooru tags, which are... questionable for what he's trying to do it seems?

But, worth paying attention to, perhaps.

>>61940 (Cross-thread)
>>62189 (Cross-thread)
>>67390 (Cross-thread)

when i place the command "git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git" it say that it isn't recognized as an internal or external command

Can it be trained for full body inflation, blowkiss or other expansion types like fully rounded blueberry also can it do fictional characters or celebs? I.e. Ariana grande and Beth Sanchez for example?

>>69263
Novel AI does anime and figuratively nothing else. I don't know why people keeps saying it's "better" when it's extremely samey and pigeonholed in comparison to most other model checkpoints: including waifudiffusion.

>>66320 (Cross-thread)
>>70420

You're not entirely wrong about this. I personally find EverythingV3 to be better.

But novelai seems to include lots of subjects in its dataset that make it more able to generate expansion related content, particularly breast and belly expansion.

Hype?

It’s onion time
It’s bunion time

>>72654
THIS. CANNOT. BE. HAPEEENIIIIIIIINNNGGGGGG. AROOOOOOOGWUGAWUGAWUGAWUGAWUGAEEEEEEEEEEEEEEEEDDEEEEEEEEEEIhbffvwruoifnwrgoi fowing
Onionsbunionsonionsbunionsonbunonbunonbunonbunrtagwhwjsisiskmdnejsjsk

>>72685
Someone got loose from Nurse Rached’s ward goodness gracious! 🤣
Merry Christmas and seek help y’all!
🦔 🍿

Did somebody say bunions?? 🧅 🤢 🤮

bump

>>69232 (OP)
Ironically preg models are really good and pregchan tries to make them look less like balloons.

>>75468
This.

bump

>>69249
>https://github.com/balloonboi/balloon-diffusion

sorry I am very new to this what do I do with the ckpt file from the drive?

Tutorials?

>>71922
>>72685
>>72701
>>74240
>>77187
Dead thread.