NOTE FOR A Cup Only v0.5

This is a test to try and get the LoRAs working better. Switched over to AI Toolkit, increased the rank, cleaned up the data set a lot. This is trained on just 60 images so it doesn't have a good understanding of anything other than A-Cup Breasts. See version notes for all tags.

I'm training a new version now that uses a greatly expanded dataset, that will be a few days out.

NOTE FOR v0.15 v v0.1

v0.15 may or may not be better for your needs than v0.1. The only difference between the versions it that v0.1 is 17 epochs and v0.15 is 29. While v0.15 seems to follow the keywords better, it also seems to generate more blurry images and sometimes generates tiny heads. These are experimental and I take no responsibility for any weirdness. That said please let me know any weirdness as I am going to adjust the dataset and start another training from scratch later this week.

Read This About!

This is the start of an attempt to create a data set that will produce exceptional amateur candid nude photographs using Flux. At this point it is still very early in the process. I have a collection of over 50,000 images sorted into categories, but they need to be tagged and captioned.

This is created for Flux1 Dev. I have used it on Flux1-dev-fp8, I have not tested it on any other versions.

For my initial test with Small Breasts I have created a LORA which has the following inputs:

• 200 - Images of women with A Cup Breasts - Captioned by Joy Cap and then manually edited to include the appropriate tags.

• 200 - Images of women with flat chests - Captioned by Joy Cap and then manually edited to include the appropriate tags.

• 100 - Images of women with B Cup Breasts - Captioned by Joy Cap and then manually edited to include the appropriate tags.

• 150 - Images of women without makeup, captioned with Joy Cap and then manually edited to include the appropriate tags.

• 150 - Images of women over 40, captioned with Joy Cap - These helped to resolve an issue with early tests that refused to do anything other than barely legal.

• 100 - Images of vaginas captioned with Joy Cap - This was an attempt to get some ability to generate vaginas. NOTE: Joy Cap uses the term Vulva, so that's what you should use until I tag them.

I limited this LORA to 1000 images, and for testing and time sake only the 500 that were important to small breasts were further tagged and had their captions edited.

This LORA is trained at 4250 steps, 17 epochs.

All images are generated with a very minimal ComfyUI workflow using Euler and Beta, CFG 1, Guidance 3.

Very minimal prompts were used such as:

An Age late 40's woman with Small A Cup Breasts with Large Nipples, Bumpy Areolas, Dark Areolas, topless. She is Sitting on a stool in an intimate setting. Wearing no makeup.

An 18 Years Old woman with Small A Cup Breasts with Medium Nipples, Oval Areolas, Light Areolas, topless. She is In a desert landscape in an intimate setting. Wearing no makeup.

Tags:

Here is a list of the tags this should understand. Note that while breast size is usually adhered to, I am still having a hard time getting it to adhere to Areola details. Sometimes you get lucky, and sometimes you don't.

"A Cup Breasts",

"B Cup Breasts",

"Bumpy Areolas",

"Dark Areolas",

"Dark Pink Areolas",

"Different Sized Areolas",

"Eraser Nipples",

"Flat Chested",

"Flat Nipples",

"Inverted Nipples",

"Invisible Areolas, No Areolas",

"Large Areolas",

"Large Nipples",

"Light Areolas",

"Medium Areolas",

"Medium Nipples",

"Oval Areolas",

"Pierced Nipples",

"Pink Areolas",

"Puffy Nipples",

"Sagging Breasts",

"Small Areolas",

"Small Nipples",

"Tan Lines"

As I get more of the data sets tagged and captioned I will continue to work on various LORA's to tackle specific elements. From there my next step will be to combine it all into a final checkpoint.