Model is intended to create a linked nipple piercing to a collar.

Trained specifically to generate a somewhat thick and detailed chain.

Trigger word is: ncc

while training words are: ncc, cha1n_l1nks, nipb4r, nippul, 3xtremepull, collar,breast_apart, squeezed_breasts

Good additional words are: huge nipples and sagging breasts,

Also, I felt like adding words such as: nipple pull, huge_chain_links do help with the generation of more high quality ones at the cost of occasionally generating more mistakes such as additional chains or hands pulling nipples.

Since it's my first ever lora, and I would like to learn/understand if what I did is right or wrong I will explain what I did in part below.

Hopefully, if my process is correct, it could be helpful to someone else so why not explain.

Dataset:

Used 108 pics which I made thanks to https://civarchive.com/user/retoto lora and also I would like to thank him for assisting me into creating this lora.

All the dataset if made of pictures of the same style which is what I usually use, and edited more than half of them with photopea to fix all the random chain bits and remade some of the piercings/links.

Tagging:

here I put my only trigger word which is "ncc" at the front of each picture and then added different tags which were relevant for the picture so that when you recall those tags you are more likely to get those feature vs others.

In this case since the chain links were detailed enough I labled them cha1n_l1nks. Written this way so that It wouldn't prompt random chains like "chain links" would.

nipb4r in this case is meant to recall the "nipple bar" piercing that you see at the side of the nipple.

Obviously collar will be present in each picture.

Every time the nipple has a higher curvature towards collar than nipples i've tagged with nippul.

And when the pull was the strongest I've added 3xtremepull with the hopes of recreating the effect as consistently as possible.

Here is an example of squeezed_breasts , which I've been applying to every pictures that had breasts together.

And lastly when breasts are apart I've tagged with breast_apart

Training setting:

I've used all the default ones of illustrious except for the "Train Batch Size" which I put to 1.

In the end I've done:

10 epochs

2 repeats

1 batch size

2160 steps (since I had 108 pictures).

I've tested epochs 7-8-9-10 and 10 is the most consistent one.

Here are my questions:

As you can see from the tags I described everything of the person wearing the object so that the trigger word wouldn't recognize those features as part of the concept.

Also, I've added visula cues such as from above, from behind, side view, from below so that when calling those words a appropriate ncc would appear.

However, this time I felt like that describing the background was going to be overkill and that It would do more harm than good so I left it alone.

Is everything correct or did I screw up some steps?

Is the style going to take over other styles?

Could have I used different training settings that would have produced a better result?

Feel free to comment or pm me what is good/wrong about this one, or even if you have better parameter suggestions I could train another version.