Didn't check back on this for a while because I kind of cracked it, got bored, and moved on. But should probably answer a couple of questions that people asked that I never got back too. (I hate it when people leave shit hanging).
Here's some examples of where I'm up to again(I did a bunch of generations here, all of these used no image to image). I included 1 that isn't inflation as well just to demostrate on the High quality ones that I've made some progress as well (basically pretty close to proving the original annon who said that 'these at best would match the quality of the shit the community draws - and other then a few errors I'm within the ball park of 'good' artists in general.
I can get realistic to artistic (close enough) in everything but completely spherical quality - the other anon figured that out already so I won't bother unless he wants to share some info.
>>69267
Nearest neighbors. train half your model using the match to get it to understand a loose concept - then move it towards the specific. The idea is if you have a big belly, you seed the model with pregnant in the tags then, switch it out after a couple of training steps to a lighter weight (stop the model, redo the tags, resume training). What the training model is doing is it's actually (removing) the things that you list in the model, but also looking for similarities at the same time. Basically it's a balancing act towards that.
For example, in the standard model of SD you'll list "big breasts" that always leads to 'voloptuous hips" the trick is basically you do a step wise training to move the embedding slowly to an 'isolated concept'. This is caused by 'concept bleed'. You'd (if your using the NovelAi / waifu diffusion model, a funny thing happens whe if you use the oppai tag, and the boobs tag together you'll end up with 4 breasts because of this, as they tag them as seperate items due to oppai loli (thin hips, big breasts)).
Basically - be VERY specific in what your training it for... the more you describe the model, the better it will be as a training image. It's good to glance at the dictionary of terms - then get very specific.
if your curious - read up on how the tagging section of the model works for recognition of tokens in the language phase. People are focusing on the images - not the 'words' section. basically what you list in your model will be related to what you get out.
>>69412
this was done via the novel AI leak. Use DIIM for best results. 32 sampling steps and a custom embeddings that I can't remember for the life of me. I used several and alot of options.
Right now - as a major aside - that's important is people haven't checked the 4chan work that's been done on this - the hyperpreg model actually contains flags for (belly inflation) (body inflation) and if you merge it and train (slightly) retrain the model with other models you can actually get pretty 'realistic' results without actually having to train a whole new model from scratch - you can actually get models that are NEARLY in range of good models that can be trained further with a leep start. (you can litterally save do hundreds of thousands of dollars of stuff on a few bucks).
Otherwise for general embeddings - it's better to have 4-5 images that are almost identical, rather then 500 that are disimilar. You use about 2/4 embeddings for a concept (things), 16 for an artist 'style'.
link for their model database: https://rentry.org/sdmodels - worthwhile checking out.
>>71405
I didn't, this was done via soft seeds of embeddings. Because if you check how stable diffusion works below the hood it's a noise algorithum. What your looking for is basically stuff that creates a large 'circle in the noise' of what will be generated at an early step and transition that into a real model. Basically in the above cases I was using a micro embedding trained of a few artists.
When I do training, I have an early 'soft bail' - if the embedding doesn't look like I want in 3-4 steps at a low go at only something like 0.05/10/10k I bail on training the model. The reason is std will hard pull if you've gotten the concept correct 'isolated' (see above) into what your training it for.
also another trick that I've learned recently as a side note - is you can actually train a hypernetwork in something 'better to 'learn' it - then a different network to 'apply it'. this is worth it's weight in goal - especially if you merge (say hyper preg, balloon diffusion, 1.4, and other models).
also never use STD::1.5 - it's trash - they removed the titties for it.
>>71929
I honeslty have no idea, usually what I do is run the model and see if it does a 'hard swing' on the outputs (a new epoch every 100-200 steps) and a test image. The algorithumn should hard filter towards the target if you've set up the data / embeddings correctly in the setup. A GOOD data set normalizes fast, very fast, then will move away again. In the first 1000 steps you can see generally if you're getting 'what you aimed for'.
Generally you want to use images that are similar for simply stuff and low 'noise' (similar). So a single pose of 'inflation' only requires 1 image. (also make sure you name you keyword nonsenes not to get overlap as well.