Purpose
Create more variety an accuracy in breast sizes, and allow for prompting of more descriptive breasts, while also improving the quality of the images.
In addition, improve the looks of the breasts in general and hopefully reduce the slight ai look to skin and faces.
I am working on creating several, of which this is the first.
Prompting
Trigger words: a-cup, b-cup, c-cup, d-cup, dd-cup, f-cup, g-cup
These trigger words, along with flat-chested or flat chested are what is used to create different breast sizes. Further keywords listed below help to define other characteristic.
Nude, topless and bottomless: Helps guide the output to the appropriate lack of clothing.
Enhanced, round, natural: Adding one or more of these words with breast will push the output towards the appropriate breast type.
Small, medium, large, very large: Adding one of these words will suggest the appropriate breast size. This is very finicky, but often describing body type helps.
Skinny, thin, slender, fit: Describes the body type, and tends to aid in making the breast sizes more accurate. Often you will have to add the word physique to this, but it's not consistent. However it also tends to drop breast size by one.
flat chested, flat-chested:: To produce a flat chested woman. Z-Image seems to ignore this quite often, but adding "skinny" or "slender" does increase the probability.
You can combine them such as:
"A skinny fit woman with medium-sized perfectly shaped b-cup natural breasts"
"She has large round dd-cup enhanced breasts"
"medium-sized natural c-cup breasts"
Limitations
Since it was not trained on the full body, but waist up, and focused on the chest one side effect is it can occasionally produce worse anatomy full body.
Description
This is the initial version I made. I had finally gotten a close version with Z-Image Turbo, but decided to redo it as Z-Image Base.
It was trained on 174 images of women of various sizes for 64 epochs.
This is one of serveral Loras I have been working on for a while, originally for Chroma but decided to switch to Z-Image.
Eventually I plan on having the following lora types:
Sizes: Which is this lora and is mostly trained on and upper body and can butcher
Women: Give more natural and accurate women full body with better anatomy.
acts; Create accurate acts in various positions
Once these are completed the final outcome will be to take all datasets and a few others to come up with a more general purpose lycoris for scenes.
It was trained under ai-toolkit with the following changes from default:
Training Resolution: 768
Quantization: None
Training method
This model was trained on descriptive prompts generated by JoyCaption Beta 1. I then ran the prompts in a ComfyUI workflow, through the base model and generating 4 outputs for each prompt.
If 3/4 of the outputs matched the scene the prompt was kept. If half were good it was edited and tried two additional times.
The images that didn't get to 3/4 (which was about 30%) I ran through Qwen VL. If I couldn't get them to work then I threw the photo away.
Key words
Main keywords: a-cup, b-cup, c-cup, d-cup, dd-cup, f-cup, g-cup
In addition they are in the following categories.
descriptive: enhanced, natural, round, perfectly shaped, ,
general size: tiny, small, medium-size, large, very large, huge
Dataset composition
Below is the dataset breakdown, and keywords under each category
a-cup - total 18
natural:2, tiny:14, small:4, :3, firm:4, perfectly shaped: 1
b-cup - total 19
natural:4, small:16, medium-sized:1, firm:4, perfectly shaped: 1
c-cup - total 21
natural:13, enhanced:7, small: 6, medium-sized:13, large:2, round:9, firm:5, perfectly shaped:7
d-cup - total 23
natural: 13, enhanced:7, medium-sized: 15, large: , very large: 1, round:7, firm:5, perfectly shaped:4
dd-cup - total 29
natural: 5, medium-sized: 7, large: 19, very large: 2, round:18, firm:1, perfectly shaped:3
f-cup - total 31
key words - large: 20, very large: 8, huge:4, round:8
g-cup - total 30
large: 6, very large: 17, huge:7, round:13
None - total 3
All but three images included cup size. This was done to try improve the success rate of when prompted. These prompts all had and nothing no other key words included for these 3. In testing I found that adding skinny, or slender to the woman's description give it a higher percentage, but still is a bit hard to get.
I originally started with 25 images each of cup size, but had to increase them somewhat linearly to get larger sizes. I suspect this is due to the images Z-Image base was trained off tended more towards small than large
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.