I said to myself... Fuck it.. Qwen needs some love... Put it in taggui, who cares if the captions wrong. Don't care anymore, prodigy can kiss my ass and so can the lora rank and boom. This is how this lora was made, about 12 hours of generation time. 5090 aint cutting it, I NEED A FARM OF THEM. I NEED 16K RESOLUTION SHARK TITTIES. I NEEEEEEEED IIIIIIIITTTI'm still on my burnt out journey of none stop AI, it's going to ruin my life.

This isn't the best lora, but it does provide high resolution images which you can see the goosebumps on the vagina (Close-ups). I think the pussy spread is ruining it. I don't like humans so don't expect much.

Default QwenImage settings (OneTrainer) 24GB lora config version

images 515

repeats 3

resolution 1024

rank 64

V0.2

Hot damn, saved by using my gut instincts, tried sampling 1 step further and it resulted in catastrophic failure. Decided I used the sample option instead of stopping training and do it raw. This would of originally take 24 days to train but since I was like, you crazy, I cut the resolution down. Apparently adding 200 more images sucked like life out of my computer.

I barely tested this, and you know what, I don't really need to. With these results I don't think sifting through 3 hours just for anything else and maybe even worse isn't worth it. Also fuck the 75 token limit, more words = who the f knows territory.

I realised that I haven't gone over 10 epochs in Qwen, how? I'm pretty sure it's my bad training habits.

This lora was made from bottled water sweat and tears made from the clouds of a volcano.

Also went to see Zootopia 2 while this sucker was going, it was basically empty inside, just like my -- um...

This lora likes close-up for some reason, putting in negative does nothing. Well this is technically a good thing as it does pretty well at it. Not using Euler results in much better results, res_multi seems to be the hot one.

Trained resolution 512

Rank 16 (5090 is poo poo)

images 731 (This killed my RAM and my computer, at this point you need a H100)

NEW TRAINING WORDS.... Yes, I fixed up the captioning and added more prompts for breasts. I'm actually surprised the first version worked because the captioning was like 25% bogus

SAMPLE FROM TRAINING: naked female, 1girl, breasts, nipples, solo, realistic, looking at viewer, nude, lying, photorealistic, large breasts, long hair, armpits, upper body, dark brown hair, brown eyes, on side, brown hair, watermark, logo, large areola

shit I just read it and the caption is fucked.... Lets just pretend it didn't happen. If this was IRL I would've gotten fired.

Also I forgot to mention that I added watermark and logo in the training prompt and well, lets just say 600+ images have watermarks in it. And for that many, I haven't seen one yet. My 5090 deverses a rest and a soda pop (Fuck coffee, coffee is for the weak)

Trigger word: naked female

[small:medium:large] (This was a bitch to do, my depth perception went to -15, size may be wrong for medium and large since I have no real measurement)

breasts

nipples

areolas

SPECIAL WORDS

goosebumps

sagging breasts

veiny breasts

grabbing own breast

puffy nipples

MANY MORE I GIVE UP

Negatives (Not necessary but know knows)

watermark

logo

english text