I’m tired of the low riding, bowed things that “panties” always makes, so I’m seeing if I can make Illustrious generate a particular type of panties with a moderate set of training. V2 uses retagged data from Anima conversion and was possible because Kohya was randomly stable one night.
You can help support making models like this by posting stuff you've made with it to the gallery using the "add post" button, that helps me earn back the training cost at no cost to you.
Supporting: crotch seam (it's actually an existing tag), navel (helps control length), [color] panties, [pattern] panties, (naturally)
Definite Negative: lowleg panties, bow panties (Illustrious and Pony always need this for panties, even if not using this LoRA, unless you explicitly want them), seam, seamed legwear (use with crotch seam)
Negative for fit: groin (makes it follow particular anatomy a bit more. I suppose you can also use it as a positive if you want it. "Highleg" can normally help with the same issue in other circumstances but it ruins the effect this LoRA is aiming for. )
Negatives for troubleshooting: lower body, head out of frame, photorealistic, photo (medium) (these are framing and style tags near omnipresent in training data that normally shouldn't pose an issue, but if they do, use them as negative), cameltoe (some styles like to include this in panties)
V1 Trigger: H_L_Pan
31 pics: 4 random anime art+2 lineart from style charts+1 western comic page I had handy+bunch of pics from major brands/retailer websites. If this works, I'll probably try some other types of panties, so tell me if there's any types you'd like to see (non-binding, but helps to decide an order).
V1: This was training when an outage on the LoRA trainer occured. This led to at least epoch 2 and epoch 3 having the same preview pics and the epoch number starting over. Not sure this received the right amount of training. I DLed a few other epochs to be sure.
V2: Uses revised tagset from Anima version, trained with Kohya on Network Dim/Alpha at 16/8 for lower file size.
Description
Uses dataset from Anima and Network set to 16/8 for lower file size/fewer artifacts. Actually managed to train on Kohya without issue.





