Underboob Style - Clothing that Reveals Underboobs
This LoRA is designed to help depict underboobs where the lower part of the breasts is exposed by clothing that end just before the breasts do.
The trigger phrases "that shows/exposes/reveals her underboobs" will activate the effect, and you can enhance it with "underboob style".
In most cases, this works well, but since the Flux base model isn’t familiar with clothing that stops beneath the breasts, the LoRA sometimes has to work against the base model. Because of this, the underboob effect may not always appear consistently in every perspective or clothing item.
The LoRA also allows for creative shortening of various types of clothing - winter jackets ending right under the breasts create a particularly unique look! :D
I’m always open to feedback, especially on how it performs across different styles and perspectives.
Description
FAQ
Comments (5)
If you don't mind sharing, what service and settings did you use to train this? (onsite, kohya, etc.)
I recently tried training something similar (underboob cutout in my case, example) but the fully trained LoRA barely picked up on anything even after 3000 steps.
Hi :)
I have used ai-toolkit. For SDXL I had made extensive use of kohya, but I get along better with ai-toolkit, especially because pausing and resuming training is a lot easier.
I also have to say that underboobs was something that wass really not easy and I almost had the feeling that the Flux base model would “fight” against this concept.
The model here is the third iteration of the underboobs lora. The two versions before that I used to create better training material with a lot of selection and culling(I hope it is the right word here). I have the feeling that it is then better picked up by the training.
I used 85 images in 3 different resolutions (ai toolkit does this automatically). I have trained 5700 steps. I included new training material from 2000 steps onwards and reduced the learning rate. From 4000 onwards, I reduced the learning rate again so that fewer new concepts were taken from the images.
I put about 12-14 hours of actual work into the Lora. The training time for the 3 models together was about 8 hours, which is on top of that.
I can't tell you whether this is the right approach, but I think it's going in the right direction. At least for me. :)
Do you have further questions?
@Catalorian I've been using ai-toolkit as well. If yours took that many steps for what I would consider an easier effect compared to what I'm trying to get, I probably need to train my dataset for way way longer.
How do you reduce the learning rate partway through in ai-toolkit? I've searched repeatedly for defining LR schedulers on ai-toolkit with no luck.
@TopazStudio
I do it by hand :D I usually let the training run for 3000 steps. If 3000 steps were not enough, I adjust the config file accordingly and start the training again. AI Toolkit picks up the last training status from there.
If it looks promising, I stop the training process and use the version that looks best to me in the sample images (I delete the other saves) to continue training at a reduced learning rate from then on.
Maybe this is all bullshit and I could have it easier... but I think it's incredibly hard to get information about good training settings :D
Spectacular LoRA, combined with other LoRAs works perfect. Very well balanced.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.