03/12/2025
added a legacy LoRa (Areola Flux Mary/aka Mar_eola Flux)
03/07/25
Added v3, saggy, v2 and v1 were burnt out - v3 hit the spot I was looking for.
Added MetArt_tits - enjoy
03/03/25
Added v3, large, v2 was good, but kept it local.
Model Training Summary
Dataset: Trained on over 40 close-up images of large areolas, sourced from various locations (first round).
Captioning: Captions were generated using Koyha_SS with GIT captioning.
Training: Conducted on Civit's onsite trainer, primarily using default settings, targeting approximately 2,000 steps.
Training Configuration:
{ "engine": "kohya", "unetLR": 0.0005, "clipSkip": 1, "loraType": "lora", "keepTokens": 0, "networkDim": 2, "numRepeats": 20, "resolution": 768, "lrScheduler": "cosine_with_restarts", "minSnrGamma": 5, "noiseOffset": 0.1, "targetSteps": 2025, "enableBucket": true, "networkAlpha": 16, "optimizerType": "AdamW8Bit", "textEncoderLR": 0, "maxTrainEpochs": 5, "shuffleCaption": false, "trainBatchSize": 4, "flipAugmentation": false, "lrSchedulerNumCycles": 3 }
Notes:
A lower weight was used to help the model generalize beyond close-up images. Inpainting wasn’t specifically tested, but previous LoRas have handled it well, so feel free to try it out and share your feedback or thoughts.
Support:
Buzz is nice, or a like.. or a post using the model...
Next Steps:
Version 2 is currently in progress as of August 23, 2024.
09/18/24
figured out how to add versions!!!
Description
V1
- works "ok" with other character LoRas - still testing.
- depending on model, 20-40 steps (f8)
- CFG between 5-7
Large Areola
FAQ
Comments (11)
great creations - works very well. can you give me a hint how to train body parts lora? what conditions must the templates have and how many do you use? did you create the LORA on civitai?
for this one, I used onsite civitai trainer (concept), with a variety of images cropped to either 1024x1024 or 2048x2048, kept as much hair, belly, arms, teeth, etc.. out of frame as possible, captioned with koyha_ss GIT captioning (basically: a close up of a womans breasts, breasts, <a key word for the source pic>, then let git do the rest.. ) I'll post more details in the description. Lots of trial and error and LOST buzz though for sure.
@rickets_xxx thanx - i guess i will try my first LORA soon. buzzed you some credits to build more ;-)
Amazing Lora. It works like a charme.
And thank you very mutch for publishing your training configuration and some background information how to create a concept Lora.
It seems you have used the standard configuration for Training Parameters?
Please could you explaine how you get targetSteps=2025.
In my case, with the same configuration:
maxTrainEpochs=5 and
numRepeats=20 and
40 images civitai always calculated:
targetSteps=250.
Thats quite different from over 2000....
Thanks for the feedback! - I try to target 6-8 MAX epochs, and try to have at least 20 pics, I set the repeats up and up until I get about 2200-2500 steps MAX (for concept)., sometimes that means 40+ repeats (its worked so far) it jacks the cost of buzz way up though... I haven't tried anything less then 2000 steps for a FLUX LoRa yet.. and Lately have discovered turning down the batch size to 2 is better.
@rickets_xxx Thank you for your reply. Your information is really usefull!
With Civitai I just tried the standard settings - and I'm not really satisfied with the results.
To increase epochs and repeats to get higher targetSteps makes sense.
One more question about this please: Normally, when I increase the targetSteps (lets say from 1000 to 2000 steps) I would decrease the unet-learning-rate accordingly: (unetLR: from 0.0005 to 0.00025 for example).
As I see, you didn`t do this. Have you ever tried to change the leranig rate or get you mostly good results with the standard setting of unetLR = 0.0005?
And thank you for the hint with the batch size of 2. I will try that next time too.
@arkinson I have been leaving the learning rate at the default, have had good luck so far.. dependent on the subject of course, just tried a revised version of this model and blew away 4000 buzz because it didnt work out how I hoped, might be a learning rate lesson here.
@rickets_xxx "might be a learning rate lesson here" :-) Funny answer!
Yes I know, the try and error method is sometimes really frustrating. That's why I'm very grateful getting some basic informations from other creators. Especially because the civitai documentation for the settings is not really helpfull.
Thanks again - and of course good luck for new fantastic Lora`s!
I don't know what the problem is, but when inpainting a photo I get birthday cake sprinkles on the areolas. I'm pulling my hair out, wth did I do wrong? How did others get it to work??
maybe drop the level down to .6, .7, seems to be the sweet spot, and keep it at the end of the prompt (my opinion) - also, are you using forge?
Just want to comment that I've had no trouble with inpainting at various weights.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.