This LoRA generates bubble butts for a number of SDXL/ZIT models. Its nearly the same dataset as my SD1x bubble butt but with a handful of larger sizes added, and better tagging.
Use the tag "bubble butt", with (optionally) one of the following size tags "medium ass", "big ass", "huge ass", "gigantic ass", "hyper butt" along with similar Hyperfusion tags.
If you want front facing images try "from front", or "ass visible from front" and negative "from behind, looking back, ass focus" at varying weights.
v4 ZIT Changelog:
Couldn't resist training a lora on Z-Image-turbo. Went pretty well, so ill probably be training more when the base model drops.
This one is just a quicky, probably not fully trained, but good enough as you can see.
Dataset is fully anime/furry, so you need to include realism related tags and words if you want that.
v4 WAI_IL Changelog:
No changes, just re-trained on WAI_IL-v14 because I was in a tinkering mood.
v4 noobai_vpred Changelog:
Not much changed. Added a few more images/captions/tags, and trained on noobAI_vpred_1.0 with somewhat similar settings to v3
Same tags as v3, you already know what to expect
"hyper butt" has a greater effect on this noob_vpred version as well if you want it really large.
Description
Training Notes:
Kohya's trainer
LR 2e-4
frozen text encoder
dim 16
alpha 8
conv_dim 8
conv_alpha 4
batch 4
GA 8
700 images
network dropout 0.3
tag dropout 0.1
scale_weight_norms 1
ip_noise_gamma 0.02
min_snr_gamma 1 (its been suggested that latent diffusion models be set to 1 or 2)
soft_min_snr
keep_tokens 2 (for bubble butt, and the size tags)
FAQ
Comments (36)
Heck yeah, I was hoping that you would grace us with your presence in the Pony space <3
Yea, finally found a training config I'm happy with, so its smooth sailing now. Might do a few more LoRA updates in the future.
@throwawayjm I think that Pony is incredibly capable, so it should be more able to produce your particular kind of greatness.
@waifuliberator Yea, its picking up on concepts really quick with ponyxl which is nice. Training on the base sdxl model was just too slow to bother.
@throwawayjm My particular favorite thing about Pony is that it inherently understands many, many characters from pop culture. Additionally, it can produce stuff like cellulite, and breast veins.
love this any chance of hyperbreasts for pony?
I could, is pony/autism not already doing hyper breast sizes well enough?
@throwawayjm Consistency is probably my personal problem, and adding specific angles for tags may cut down on its data set. Going (hyper breasts:1.5) sometimes risks blowing up the image, too. Even then, sometimes it can be a struggle to get past roughly torso size. So, anything to help nudge things along would be cool, imo.
@WeisseningBlitz I already convinced myself to give a new hyperbreasts dataset a try on sdxl. If it turns out ok ill post it.
@throwawayjm it is, maybe my command prompts are not good :( I think I will try and find a guide then. And SDXL would be great! :)
This is AMAZING! I've been playing around with it and I get THAT BUBBLE every time.
Yea, and its nearly the same dataset as the 1x version. Just goes to show how much better fine tuning on ponyXL is than 1x models for these concepts.
@throwawayjm That's what I've been hearing about PonyXL is that people are taking their 1.5 data sets and upgrading.
I am not necessarily complaining, but this model definitely wants to make breasts larger also.
Yea I imagine it will considering the dataset is mostly hyper sized content. Just put breast size tags in the negative.
@throwawayjm lol. I most certainly will not. :-D
Is training data available for this?
Do you mean, training details, tags, or the images themselves? Training details are in the sidebar. Tags I could post if there's interest.
Flux?
Probably going to wait on a decent base model fine-tune first. Just experimenting with Flux in the meantime.
@throwawayjm dev ggufs are very good, they just cant get a big butt like flux in general its so phobic to butts
clip_skip=1. I Usually use any of the DPM++ 2 samplers.
@throwawayjm thanks!
Can you train for the Illustrious Checkpoint please?
I was thinking about training a v_pred XL model, so potentially the noobai v_pred variant. Waiting to see how nvidias new sana model looks if they release it soon.
@throwawayjm v_prev XL model? Why?
@throwawayjm Hi, I have been testing Epsilon Pred 0.5 version, and I have to say that it is a bit superior to Pony Diffusion. I accessed your Discord and was recommended “checkpoint-e0_s10000.safetensors” I'm looking forward to your Lora adapted to this checkpoint!
@megachina31687 Yea, its pretty much the only choice for v_pred SDXL models right now. Training my own from scratch would take over a year to do, so its better to start with theirs when its done.
@throwawayjm Great! They recently released the 75% version of Epsilon Pred. Looking forward to your new version!
@throwawayjm I've been testing the 75% and 100% version (v1.0), but I'll stick with this one. It's a very trained version they have.
@throwawayjm I just wanted to cheer you up. I hope everything goes well. If you need help testing or adding ideas, I'm here. I use an RTX 3080 to generate the images. I don't use any servers.
@megachina31687
I'm just waiting on the 1.0 vpred version of noobai to finalize. Got the dataset ready to go when it does.
@throwawayjm understand!
@throwawayjm The "V Pred 1.0" version is out of the oven! I can't wait to try it!
@throwawayjm I share an article about Noobai-XL: https://civitai.com/articles/10622