This Lora uses an experimental guided training to ease the formation of the neural network.
Instructions are given through the images themselves. For this training run, a color association method is used.
The images in the training set are divided into two identical side-by-side images. On one side, the components of the female genitalia are colored distinctly.
In the image captions, the image is described as having two identical sides. Then, a description of the anatomical component to learn is given as well as the color of its colored region.
This isn't a general use Lora. The data set lacks a reasonable quantity of far and medium distances crotch. While there are some male pictures with color associated male genitalia, it's only a handful.
This training method is very efficient for complex concepts, such as the female genitalia as seen here, or hands, etc.
Link to article: https://civarchive.com/articles/3870/guided-training-with-color-associations-for-complex-concepts
V2 expands the dataset. It shouldn't be very proficient at male genitals, but it's better. It should be better at farther distances, too.
Description
Expanded the training dataset. Changed the training settings a bit.
FAQ
Comments (8)
Very interesting approach.
I think it's worth making an article about it, with more information and examples.
this is innovation. i'll try it out and see if it works.
I am actually working on several merge, I will try it out to see what it does in my merge ;) thanks for the works
An anime one would also be appreciated. I've tried to make one myself but I just don't understand the process at all. I know some loras exist and such to try and correct it, but it just isn't the same.
so is this meant for inpainting, to fix bad pussies? as i gather if i were to use this for general txt 2 img i'd create mostly vaginal focused shots right?
for sd 1.5 please?
What version is the recommended one with better results? The v2 or v3? The v2 is coming first on your recommended list, that's why it's confusing.
I strictly do full body generations. The lora by itself wasn't strong enough for full body renders. Whenever I tried full body shotss the vagina would get deformed instantly. Putting the lora strength too high also drastically altered the picture in a negative way. I fought this by smashing a few vagina loras together at a low strength, usually around 0.3 so they wouldn't affect the overall picture. Turned out quite good, now im getting decent vaginas in 8 out of 10 pictures. The failure rate with a single vagina lora was like 9 times out of 10. This picture was a single generation, no edits other than putting it into img2img to upscale it, but for that the vagina looked good too.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.