Trans Girls Flux

Hi! I made a nice writeup now, and accidentally clicked away and it all disappeared, so I'll make this short :') lol

Firstly - SHOUT OUT TO @niestyboi !! He is awesome, he gave me my first ever tip! A whopping 10'000 buzz! I am very thankful, it allowed me to create this LORA. Thank you very much! :)

Hi there!

So I made a concept lora for you guys. it makes naked trans girls. yessssss!

Why did I make this?

Because when it comes to NSFW.... FLUX SUCKS! I thought it was cool you made Andro flux. you made lora's that make pussy. lora's that make cock. god I hope I can use those words so freely here lol. well. NO GIRL DICK LORA's??? Yea I saw a few "futa" lora's but this is not that. this is a lora to make trans girls.

Trigger words examples: trans, trans girl, trans woman, transgender, penis, testicles, flaccid, erect, small breasts, medium-sized breasts, large breasts, small nipples, big pink nipples, puffy nipples, anus, white girl (I specified the ethnicity in my training data, so you can change it :-) should work! xx)

WHICH SCHEDULER TO USE:

DEIS + BETA (44 steps) 

(or do 80 steps if you want very good images - when you find a good seed! can go higher too. experiment)

UNI_PC_BH2 + BETA (42 steps)
^^ this one is GREAT! looks the BEST but most pics are bad so u have to really cherry pick. V2 will be better I swear!

it is NOT DDIM like it shows in the example images, maybe someone can correct me but I wasn't able to select "DEIS" as an option sorry about that.

I used bigger network DIM and ALPHA than before so I trained it on more concepts at once. it seems to get the boobs really nice sometimes. I added "big puffy nipples" as a trigger cause I'm trying to use this to make porn of myself so I needed to train it on how mine look and that's like most accurate description lol. it can do the size pretty well too. so when describing breasts don't say like "she has small boobs", rather be like "her top is lifted up, revealing her small-sized breasts"

You have to really speak like you are joy caption. lol. IYKYK

How to use it? with the model-only lora setup for flux. I will upload my workflow later.

How to prompt it? pretend you're a LLM :'D cries lol. I will fix prompting style in version 2! But really, make sure you read the example images in the post and see their prompts. copy those for now I know its long and tedious but at least u can be specific with details, I tried to label as much as possible.

How I made dataset:

1. Downloaded a ton of HD "content" style pics like from onlyfans girls, pictures they have publicly available online. I tried to do all different people just randomly, ones that I thought I liked the look of their pictures. very realistic looking like good iPhone photographs not professional porn styles which you can't use as content. I do onlyfans too and this was my goal when making this lora - to make onlyfans pics of myself like optimal ones, that look like I took them with my phone. So i downloaded like 1000 pics or so. way too many. took me a whole day. just ones I thought were GREAT. and that were over 1024x1024 pixels.

2. then I picked I think it was 150 images for this dataset. First, I used a manual tool to place a square bounding box on each image, over which area I thought contained the most relevant information for a 1024x1024 picture. I scaled and cropped EVERY image MANUALLY to 1024x1024 so its in the PERFECT position. auto crop seemed dumb.

3. I used Joycaption to label all of them. then used GPT-4o in a custom comfy node, to transcribe each prompt into more natural style, and to make sure to describe the subjects as "trans girl" not "girl", and a few other things like it reorders the captions to be better for training. had to MANY openai calls lol, as its NSFW most times it says no but some prompt engineering u can make gpt 4o api do anything.

4. then i went over them all manually, making sure to correct all the captions manually. took long. but still its very average dataset now when I look back. I can't wait to see what I can do with a better dataset with 300 images. or 500 even. god I'll HAVE to use vision models to automate that. maybe multimodal LLM models? anyone can give me advice on dataset prep specifically captioning LARGE amounts and cropping effectively in smart but maybe efficient way would be a godsend. I happy to like spend 100-200 hours manually curating a dataset, checking EVERY image and reading and modifying EVERY caption - but it would feel way better if I knew I was going in the right direction at least cause I wasted a lot of time making this database, picked some bad images for example which can spoil the whole lora.

5. made sure to include a good mix of full body, bottom half of body, weird angles, but there are some bad ones. while making this I learnt a lot, and next time I've got some cash to do this again or if someone would like to contribute to V2, I will use my next Buzz to train on a completely new dataset. going to recaption everything on new images.

In V2 I will also change the trigger words, to be more 'casual'. and prompt style will be SHORTER like 10-15 word prompts in V2 if that's what you guys want?

Anyways I have rambled again. enjoy making trans girls lol and sorry about the insane animal creations you might see now and then if using other schedulers or combining with other lora's lol. good luck using this please be kind it is a rough LORA lol but you can make nice images with the right workflows and tricks, most of the time! :) find a good seed - stick with it and change prompts. another way to get good pics.

Known Issues

Penis-hand, penis-in-hand, detatched girl penis. Phantom penis. Missing penis. four balls. penis-fingers. oh the horror.