If you like what I do, consider buying me a coffee~ https://ko-fi.com/budace
Trained on the same current dataset I used to train the regular SD Inflation LoRA found here: https://civarchive.com/models/210476/inflation-lora
I also started a Telegram channel where I share random generations that I make: https://t.me/InflaitableWonderland
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.
Description
FAQ
Comments (33)
Really cool LoRa! Did you use image classifiers / some other technique for tagging? Can't imagine manually tagging all the images
I manually tagged every image in the dataset. It's really tedious XD
He really did do all this by hand. He's completely bonkers
@budace damn dude thats some dedication. I'm not this dedicated, looking for ways to automate this kind of tagging (Look up VLMs if youre interested)
@septagon Problem I have had is auto taggers don't recognize the size controls that I add to my training data. And especially don't auto tag stuff like -monorump- and -spherical inflation-. So I just take care of it myself XD
@budace yeah its too bad it doesnt do the more niche tags, im trying to fix it by training it on data sets with custom tags tho
@septagon Also I am lucky enough to have access to the current version of Photoshop so I have been going "generative fill" by hand on each image to get rid of text and speech bubbles. Your image cleaner looks interesting. I guess I'll have to get back into using Comfy and try it out X3.
nice
seems to have some tearing issues when tying to get it to focus on monobutts
Yeah Monorump is still a challenge. It works best if you apply some leg tags too. Like inflated legs, or puffed legs. It ain't easy training in so many concepts into just one model, especially stuff that no checkpoint understands like monorump XD
Will you be exploring more on the head inflation? Y'know, making it more customizable like round head, scaled up head without puffing it up, uber head, etc?
I hope to, it depends on finding good training images for it. With this version I included it just to see if it would even work in the first place.
@budace Well, good luck with that. I'll be waiting to be able to do some big head mode.
have you thought about training a flux lora? Or sharing the dataset for someone to try to do it? You've crafted a truly great lora here.
I have thought about it a lot. But I don't know if I have hardware good enough to train flux myself and my current dataset is too big for the 1000 file limit to train on Civit.
@budace Now that is curious to me, that you used such a giant dataset. I imagine that's why you've managed to get such good results. I do wonder though if you could train smaller, individual things, such as breast expansion or spherical expansion or w/e individually. As someone who's done training himself and attempted to train similar concepts to you, I must admit you've trained perhaps the best lora for this subject that I've ever come across.
@LostIdeas One reason I didn't want to do individual loras is because I want to be able to mix things together. And things start get wonky when you layer too many loras together. My dataset is big because I have over a dozen different concepts and various mixes of each of those with many different viewing angles. I'm taking a break from training this model for now to focus on some characters but when I get back around to it I am actually going to try and reduce the number of images. I'll have to scrap the current dataset and start over.
@budace That makes a lot of sense. However, I think that in some ways, at least for this, you've succeeded. Rather than starting over, you could start with trying to prune the dataset instead; of course, that would just be my suggestion. I would still love to see your dataset, if only because I would love to try and train one of the concepts on a FLUX model; you've managed to do something great here though, so I trust your judgement.
@LostIdeas The plan I wanna try is limiting myself to no more than like 5 images per concept. Way easier to do that process on a clean slate than pruning thru over 2000 current images XD. I plan on keeping the current dataset in case that doesn't work out so I won't lose all my work. Also yes I tagged each image used in training this lora by hand
@budace you absolute mad lad. That's a lot of work, and I respect it. I do think that it might be a bit too few if you only have 5 images per concept, but also, you are the expert, so it could work.
@LostIdeas I apologize for ever giving anyone the impression that I actually know what I'm doing XD. It's basically been really tedious trial and error so far. And that's just another thing I want to try XD
@budace I mean, as someone who has also trained things, and talked to a lot of people who have trained things, I can say that everyone is doing everything from trial and error. I know some people who have trained other similar loras to an extent, such as the person behind Symmix, who has a discord server. Still, You're doing good work, so at the very least, you've got that going for you.
I have to admit that this is one of the best loras out there. The only thing that I've noticed when experimenting with different models is that it doesn't play well with realism models; which isn't surprising. But it's curious to me since of how well it does with many different art models.
Hopefully we'll see an Illustrious version soon, which could work with this sharper and better.
I plan on branching out to other XL models. I already had a Noob model planned too. It will be a while as training XL takes a long time XD
@budace WAI-NSFW, NTRMix (what also Noob) and Hassaku XL is making good results and it somenting working last time Pony, but yeah. Fuji right, we needed Lora IL possibly, because it concept not always good more strange because Gelbooru.
WAI-NSFW possibly here good, if inflation more NSFW content.
https://civitai.com/models/827184?modelVersionId=1183765
https://civitai.com/models/926443/ntr-mix-or-illustrious-xl-or-noob-xl
https://civitai.com/models/140272?modelVersionId=1240288
@budace Pony version lora may be able to work with the latest WAI-NSFW version 9.0, as Pony support was reportedly improved.
Also includes training material and much more.
I recommend WAI-NSFW for training or testing previews, if you yet decide soon to making an Inflation IL version.
VERY NICE
I read this in Borat's voice
i have promblem bc generate me only back view arts
Any chance for pear shaped inflation? aka the inflation focusing on the lower body while keeping upper body normal?
That's actually in the v6.0 model under -monorump-, It works best on that model when paired with -puffed legs- or -inflated legs-, but it's inconsistent. I'm working on the update right now actually and I'm putting a lot more effort into making that more consistent.
@budace Thank you. <3 Do take your time.