If you like what I do, consider buying me a coffee~ https://ko-fi.com/budace
I also started a Telegram channel where I share random generations that I make: https://t.me/InflaitableWonderland
Model 9.0 Update
I did some major changes to a lot of the tags because I think for this whole time I have been competing with existing checkpoint knowledge of the more common concepts like breast expansion. I really didn't want to but I changed a lot of the tags with the whole l33t speak thing. The examples grids should show all the new tags but for people who actually read the description I'll list everything here too.
Breasts:
b00bexpansion, hyperb00bs, blimptits, boobwall
Butt:
assexpansion, hyperass, blimpass
Belly:
0verwe1ght, b3llyexpansion, hyperb3lly, blimpb3lly
Thighs:
th1ghexpansion, hyperth1ghs, blimpth1ghs
Hips:
h1ipexpansion, hyperh1ps, blimph1ps
Full Body:
b0dyinflation, sph3ricalinflation
Those are pretty much the only tags that changed. Thinks like monorump and head inflation are still the same because those are already concepts outside checkpoint knowledge. I try to give examples with the rest of the modifiers in the image grids. but if I get enough comments I'll list those down here too.
Model 8.0 Update
I did some pretty heavy editing of the dataset so I consider this a new overall version. I think these are getting much better at preserving checkpoint styles and letting character details come through. It gets a little muddy with the more niche concepts but you can kinda get around that with more detailed prompts.
Core inflation concepts
I tried training in each concept on its own by tagging them as they appear in the training data. I added a little bit of size control with the size tags. You can mix and match any of the different body part concepts but I would not recommend mixing the size tags that come with them.
Inflation concepts:
Breasts
Hips
Thighs
Butt
Belly
Size tags
Expansion
Hyper
Blimp
Each concept above can be mixed and matched any way you like. The only thing to keep in mind is that the blimp sizes don't have a space in the word. This was to prevent the AI from just making images of actual blimps XD. Also the ones for breast and butt are Blimptits and Blimpass because those roll off the tongue better. Also since breasts are the biggest focus for most expansion art, those get a fourth size with Boobwall
Secondary concepts:
Body inflation
Spherical inflation
Head inflation
Monorump expansion
Hyper Monorump
Living Inflatable
Body inflation and spherical inflation are for that full body inflated look. Body inflation is for that puffy look that happens just before the subject fully rounds out. Supplimental tags for these are inflated arms, inflated legs, puffed arms, puffed legs. Inflated and Puffed can be mixed with pretty good effects but they work best together. I used "puffed" for that "donut" limb look but if used the word "donut" the AI would just throw actual donuts all over the image XD. I don't recommend using hip or butt tags with either of these but who knows it might still work. I tried to make this model as customizable as possible.
I split monorump into expansion and hyper because I have images that have very wide range of sizes. The leg tags work really well with those too.
Living Inflatable has quite a few supporting tags as well. Permagrin is one for that painted on smile look. For nozzles and handles I tried to tag separate areas but sometimes the AI still throws extras in. For handles on the hips use hip handle. For nozzles I tagged in a few different areas, navel nozzle, hip nozzle, nipple nozzle, tail nozzle
Finally there's a few other ancillary concepts I tried to tag but they are wildly inconsistent. I tried to tag in different looks for weight and floaty. Heavy, weightless, and buoyant. Each concept responds to those a little differently, but it still needs work.
There's Lip Expansion for some nice thick juicy lips lol
I tagged in a few floating concepts as well. Like floating on its own works well, but there's also floating on back, and floating on front. But those are a little inconsistent still. This model is a never ending work in progress lol.
__________________________________________________________________________________________________
Model 7.0 update
I had to completely rebuild the dataset again. I tried adding tags for clothed and nude to prevent things like nipples showing up on shirts and stuff like that. As always the core concepts are shown in the grid images.
For more info on the core tags and other concepts trained into this model keep reading below~
__________________________________________________________________________________________________
Model 6.0 update
Please expand and read the full description.
This update is kindof a reboot. Overall I think it performs way better but there's still some oddities I need to fix but I've been working on this for a month and I just want to post what I have XD. Like I don't know why it wants to keep putting inflatables in unfinished houses. But anyway, tags remain mostly the same as below.
Monorump and head inflation work much better on Pony here than my SD1.5 models. But Pony still has trouble separating the hyper and blimp tags. Still needs work but I'm pretty happy wit this one.
__________________________________________________________________________________________________
Model 5.0 update
For reference. The way my version naming should read is the number like 05 is the 5th major revision of the dataset itself. The name like Polyfur is the checkpoint I trained this version on. The e010 part is how many epochs I ran training.
Anyway on this model I made a major change in adding monobutt as a tag for a new fun shape. Also other fun tags I tried training in like "torn topwear" and "torn bottomwear" altho it's still a bit inconsistent. Also more expression tags trained in like "looking down at self"
Also got tired of e621 models when I try to specify "three-quarter view" 90% of the time it defaults to a rear view. So I tagged "front three-quarter view" and "rear three-quarter view" so experiment around with those.
I'm still not 100% happy with how this one turned out but I've been playing around with settings and training methods for a month now and I just wanted something to show for it. It still does the basics REALLY well.
Hide
Description
Trained on Illustrious
FAQ
Comments (18)
The new tags does bypass the filter, but sometimes it.. kind of flipflops, sometimes it'll detect it and switch to yellow at random. And now its switching to yellow regardless of where I put the new tag. I hate this overly sensitive filter.
seeing as the male lora seems to be a bit outdated, i wonder how compatible this is with male characters?
"a bit outdated" my man I updated it barely 3 months ago xD. The female inflation Lora works on male characters as long as you want to give a male character feminine curves. But none of the trained male expansion tags exist in this model. But stay tuned~
@budace alr, thanks dude, and thanks for the great models
@hogcranker9000 I do need to update the male model with new tags tho eventually. But I got a big thing I'm working on getting done before christmas~
Legend, thank you again. That main closeup, damn 👀.
So one thing I'm noticing that I'm experimenting with the tags is that there's a lot of bleeding between tags at the hyper level. For example, the boobwall tag often causes it to create bellies instead of breasts. There's some similar issues with other bits, but most of the tags leading up to the maximum tags seem to work well.
These are all things i have been trying to iron out for the 2 years I've been working on and updating this model. I guess it's the price I pay for trying to make an all encompassing inflation model. The tweaking never ends~
@budace Figured it was worth noting, if only because overall, I think there's a lot less bleeding than in earlier versions. It's also just this one tag I've noticed so far. Only other thing that seems to have an issue is posing; but that's a model issue more than a lora one.
@LostIdeas you can kinda see which checkpoints I gravitate to in my examples. BananaSplitz for hand drawn style and TigerMilk for realistic
@budace Interesting. I've never tried either; I have mostly been using WAI for testing since it seems like the most popular. Guess I'll have to try them!
@budace Hate to bother you again, but in the course of my training, I've encountered something I wonder if you've found a way around, specifically with spherical inflation tags and sizes. The single hardest thing to train the model to do seems to be getting that spherical shape, particularly in terms of proportions. What methods or tagging do you do to avoid it ignoring things like say, sunken hands tags?
P.S. thanks again for your earlier tagging advice to only tag things that you want it to learn, that's improved the quality a lot already!
@LostIdeas Doing the l33t tag like I did this time around helped a lot too. But it's always the one the model seems to learn last. To get it to properly show up I always train this model for at least 30-40 epochs
@budace Oh wow, that's a lot. How many steps is that? Like, how many repetitions/batch size do you use? I imagine that 30-40 epochs would be nearly 100k steps, right?
@LostIdeas I use the non-Prodigy method so the steps for each epoch is however many images you have in the folder. For this huge dataset I have been training at batch size 2 so that does cut step count in half. But yeah it still comes out to around 40-60k steps overall
@budace I mean, even using prodigy, you set how many steps/epochs you want, it just has a learning rate of 1. That said, in the past I found that training beyond like, 6k steps was often overkill. But I'm going to have to up the steps then, see if that works better to get better results. Probably avoids me training 40 versions of the same Lora. I'm guessing you're not using the CivitAI trainer, because that would cost like, $10 a training session, lol.
@LostIdeas Yeah for characters I hover closer to around 1000 steps. But when you're training a lora that combines over 20 different concepts it needs a lot of training. And yeah I have a PC setup as my training workhorse locally.
@budace A good plan. For me, I could train locally, I just prefer to use Civit's trainer since I already pay for it lol. But yeah, I tend to find characters and styles to be far easier than concepts; even not doing as many as you, I find concepts to be the one thing that's escaped me. Hopefully not as much going forward! Thanks again for your help!