This is a female chastity LORA. The type of chastity belt is the one that has a silicone lining so you can see the color on the outside. This time, I've trained it in two stages. First, using the full set I got. But then, I used that LORA to generate AI images, focusing on getting as much of a flat/anime style as possible. With those synthetic images and a subset of the original image training set, I trained a new LORA using the first LORA as base. Comparing the output of the two LORAs (version A and version B), I saw that indeed the second one didn't tend towards a 3D style as much as version A, so probably this validates the approach. From now on, I will train the LORAs in two or more stages. This is also useful by the way when the original training set has too many close-ups, since I can force (e..g using controlnet) the AI to generate cowboy shots or full body shots, and then re-train the LORA with those. All trigger words involve the made-up word Chastix. The reason is that I tried to avoid "chastity belt" or varieties of it, because "belt" will most likely get the AI confused. I think this model does pretty well all kinds of angles, different lining colors (black lining, pink lining, and so son), as well as the metal bra and thigh cuffs (also wrist or ankle cuffs). The reflections and details feel pretty good. You can use "anus hole" to have the AI generate an opening in the belt for... purposes. What this model doesn't do well, despite my intentions, is showing a girl holding a chastity belt. Sadly, it doesn't work that well. You may get 1 out of 10 generations looking decent. I really like the idea of having LORAs that not only show the characters wearing x fetish item, but also holding it or interacting with it.Also training it with synthetic AI images led to a degradation of details such as the padlocks or chains when there's a metal bra or thigh cuffs. This has taught me that maybe when I train version 2 of this LORA, I should first "burn" a first model, so it really learns all those details, and then generate the most 2D images possible that preserve these details. Otherwise, the model "forgets" about those details if the synthetic images don't preserve them. I've tested it with many illustrious derived models, and it doesn't seem to tend towards a 3d style as much even with the worst "offenders" such as the base.

Files with SHA256: de948fc1391faab572c361d3d48599a91d41600550fa61c47ae6a6f22f16777c