NOW WE HAVE BANANA|TORPEDO and SAGGY TITS!!! See the stand-alone lora file version at the top
I was tired of other LoRAs without the areolas textures and details. Most of the nipples I got until now using Flux was some soulless plain areolas without the depth it requires.
Then I curated a mid-size dataset with the best breasts and areolas I could find today in the internet.
Stand-alone lora files for more specific breast types:
long nipples - Nipples that you can bite easily
puffy nipples - Small breasts with the tip of them covered with a dense nipple
veiny areolas - Veins appearing under the skin
dark nipples - Dark nipples with more contrast to the natural skin
banana tits - Similar to puffy, but focused on breast shape instead of nipples only
saggy breasts - breasts hanging a bit due to the gravity
ghost areolas - Areolas with almost the same skin tone as the breast
empty tits - Similar to saggy, but empty
flat tits - Flat chested woman
oval areolas - When it's cold, the areola shape may change to oval. Good texture
Types of breasts trained on general lora: (Not trigger words, but the types of images used to train)
big areolas - Rounded medium to big areola size
dark nipples - Dark nipples with more contrast to the natural skin
pale nipples - Pale nipples that appears slightly
big nipples - Nipples that you can bite easily
puffy nipples - Small breasts with the tip of them covered with a dense nipple
saggy breasts - breasts hanging a bit due to the gravity
small areolas - Areolas with a small radius
veiny areolas - Veins appearing under the skin
banana tits - Similar to puffy, but focused on breast shape instead of nipples only
silicone tits - The spheres from the 90s. It focuses on shape
empty tits - Similar to saggy, but empty
flat tits - Flat chested woman
oval areolas - When it's cold, the areola shape may change to oval. Good texture
bumpy areolas - When it's cold, the areola texture becomes bumpy. Good texture
wrinkled areolas - When it's cold big areolas becomes wrinkled! Good texture
Bonus that I added in case the full body appears, to have something right down there:
hairy pussy - hairy pussy of a woman standing
shaved pussy - shaved pussy of a woman standing
You can use multiple loras, but the image may degrade... It's just how it works sadly.
Description
Version 0.1 beta.
FAQ
Comments (28)
"one miss fp8 training"
Did you mean AI Toolkit? Because unfortunately, I got a CUDA error trying to train the 9b on my 4060 with 32gb RAM.
I'd like to try the one that can train fp8 but lacks advanced settings, would that be SimpleTuner?
I usually train using Kohya. They are still developing. The script is missing fp8. AI toolkit is working, I'm using it, but the text encoder offload is broken, but it's easy to fix the code (there is a thread in the github issues). I trained with a 4060 16GB VRAM e 32GB RAM
@Skullkid I have about same setup but I get out of memory errors. Could you share config or details on teh offloading part? Thanks
@Skullkid Hey can you please upload your ai toolkit training config file ? I'm really struggling with the memory and all along with the parameters even though im using almost the same setup as you. Would appreciate your help !
@HyperXSen I gave up from AI Toolkit, musubi-tuner use way less memory. Soon I'll update the article about training, because on musubi you have to run scripts manually, but it should be integrated in Kohya SS soon.
On Toolkit I run like that: type "lora", adafactor optimizer instead of adamw8bit, resolution 768 (may need to use 512 only), batch 1, quantize qfloat8 for both model and TE
unload_text_encoder and cache_text_embeddings true. I run on Linux, then the OS use only 100MB vram. On Windows this may be a problem.
@AiLEarner the offloading on AIToolkit is broken. Check here where I suggested a fix in the code:
https://github.com/ostris/ai-toolkit/issues/653
Also, I gave up fom AITollkit to use musubi-tuner, that use the scipt sets from kohya, then it fits more the training I'm already used to.
Have you tried SimpleTuner? It also has Flux.2 Klein support since quite a while. And it supports regularization, eval, ... all the stuff you need for real training
@StableLlama I've tried... to make it work. Installed WSL, Ubuntu, Docker Desktop, then build the image, but then it wouldn't run. Given it took like 50 or 80 Gb of disk space or something like that, I decided to give up. Now I'll try Musubi Tuner as the OP suggested.
@tsunamix instructions are here: https://github.com/kohya-ss/musubi-tuner/blob/main/docs/flux_2.md
@tsunamix ok, I don't use Windows, so I don't know what ugly things it does.
But I can run it quite well locally as well on cloud GPU providers (vast.ai, RunPod and modal.com were all working for me with SimpleTuner)
Its hard to make small tits using it in edit mode, any advice? But nice work :)
Yes, I trained with small tits images but it seems that the lora rank size is still not taking effect in aitoolkit training.... should be fixed in later updates, when the training tools finish implementing all training functions.
@Skullkid I’ve done some testing, and if you insist a bit with the prompt, it actually does what you ask. Thanks a lot!
@Vyxen808 That's good. I already found a problem in my training configs, I'll update the model once I train it again!
@Vyxen808 Could you share a prompt for really small breasts, please?
on V2 it's a bit easier now
So annoying and stupid they censor models like that, we are biological not mechanical, for f's sake, grow up already, gosh. Good job on the LoRA btw.
Agreed but the backlash they'd receive would be astronomical. I think it's stupid.
@Keroro_Gunso I know exactly the generation and type of people who f*ed it up for everybody. knowing nod
@eyeonyou other than those people. It's a new technology. People are scared. They don't know what it's capable of or where it's heading. For most people, AI is just something that flutters in & out of their world as moments pass. No thought into what it took to create it. What model. Nothing. To them, its all monolithic and homogeneous. When big bad "AI" does something wrong, the entire industry takes the blame.
They can't help it. It's just sheep shit.
Pretty soon everyone will have to wear a hooded cloak in real life. Apparently bikinis and bathing suits are a problem now. Once everyone's in the cloaks then they'll have to start wearing gloves because bare fingers will be a scandal.
Holy shit its really youuu!!!.... youre back! your flux 1 dev nips loras were the absolute pinnacle of its kind.... Really glad youre back in the game!
so to say "top of the nipple" ^^
@Mech4nimaL took the words right outta my .....pants xD
This is great! What are you using for lora training?
AI Toolkit
I'm going to change to musubi-tuner now that they fixed flux2 training. Aitoolkie seems having something wrong as the training is not converging.
@Skullkid ah, good to know! i hadn't heard of musubi-tuner, i'll check it out. thanks!