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
Added 10% more images, and improved a small amount of tags. Otherwise same dataset as v3 but trained on NoobAi_vpred
500 handpicked images extracted from my bubble_butt classifier dataset
Training config:
Optimizer ADOPT
--optimizer_args "decouple=True" "weight_decay=0.001" "betas=0.9,0.999" (not sure if this is the best config, still tinkering)
Edit: after some testing, don't use weight_decay with DoRA, it hurts the results.
--zero_terminal_snr
--v_parameterization
See bubble_butt_v3 for the params I didn't change
FAQ
Comments (23)
The return of the god of Thick. Welcome back <3
Thanks!, new model new experiments.
The return of the Ass GOD! Thanks!
It's beautiful
"slimthick" is a key tag?
Yea, but there were only 60 images with that tag in the datsaet. Not sure if it works.
@throwawayjm Thanks! I see that the tag adds an interesting narrow waist effect. Are there any more labels of this type on your Lora? I'm still testing.
"assheavy" is the only other custom tag that might do something. Its meant to make the breasts, thighs, smaller, and focus weight on the ass. Its similar to bottomheavy, but with slim thighs.
@throwawayjm and... Is "from front" other key tag too?
@nokiasupreme973 Yea `from front` `from behind` both have over 150 images tagged
Can you share the dataset so that other users can train your Lora at different checkpoints? The speed at which different checkpoints with different AI engines appear is overwhelming, and I only see two versions of your lora: Pony and Noobai XL V-Pred 1.0 version. I think this is too specific, and your Lora has a LOT of potential. You could share it with specific licenses like non-monetization, obligation to tag you, etc.
If you want really want to build your own similar dataset, I released all the custom tags for my datasets here: https://huggingface.co/datasets/thojm/hyperfusion_classified_tags_export
You can pick out the image_ids tagged "bubble butt" and train on those. But you would need to download the images yourself. This was an easier solution than sharing my datasets every time someone asked for them.
@throwawayjm yes, sorry, I meant to say “training data” and I said “metadata”, I confused terms. Thanks, best regards.
Sorry @throwawayjm I did not explain myself well. I understand that training a lora for each checkpoint costs x buzz (600 buzz?), so, I propose you to share the training data (images, tags, etc.) so that other users can train your Lora for each checkpoint. Obviously this would be accompanied with a license and usage requirements (e.g. obligation to tag yourself as creator).
However, I understand that you don't want to do this. I think it would be easier to pay you with Buzz and ask you to adapt your Lora to x checkpoint that a user wants. Your Lora is only adapted to Noobai V-Pred 1.0, but this checkpoint is very specific and there are more Illustrious based checkpoints (ex: WAI NSFW, NTR Mix V777, Hassaku, etc.).
@nokiasupreme973 Yea, I understand. The tags export above is the best option for collecting all the images I use in training. If people are willing to download those images filtered by the tag they want, then they can build their own datasets that would match mine (I'm training on the same images). I'm just too lazy to do anything further, and have never been motivated by money anyway (it would stop being a hobby at that point).
Please retrain for illustrious when you can.
I don't plan to train on illustrious since it is only trained on anime content and way too limited for my tastes.
@throwawayjm Fair enough,
where tf do I put DORA???
@throwawayjm Did you see the new versions of illustrious? They released version 1.0 and the second version. They plan to release a third version.
Yea, unfortunately those didn't interest me since its mostly geared towards anime, and its not a v_prediction model as far as I could tell. I picked noobai_vpred because its diverse dataset makes training my concepts easier, and training an SDXL finetune without v_prediction is more time consuming.
Can you make a Lora Illustrious of this pleace?
Yea, finally got around to trying out WAI_IL. Seems decent enough.