Experimental: popular demand request for a niche. It's definitely rough around the edges.
Start at 1.0. You might need to bump up if the draw is incomplete, and hiresfix at 1.5x helps with distance shots. For Anime models, increasing focus on face helps.
Anime Starter Prompt:
kaopan 1girl,<lora:kaopan_v2:1.4>
Photo Starter Prompt:
<subject>, kaopan
or
<subject>, (Kaopan:1.2), frill, lace , closeup <lora:kaopan_v2:1.0>
or
<subject>, kaopan, with <description> panties covering her face
Description
FAQ
Comments (6)
Where's the humiliation? They look like they're enjoying it. :)
There are solutions for that too :p.
(for reals though, there's virtually no way those two play together, and the outcomes are probably horrific. That LORA with the broken keyword is a lot of fun though)
@thegipper Yes, I commented on that LoRA too. :D
Combing the two LoRAs ... I'll give it some thought. :D
AI seems to recognize the [crotch] part of [panties] as [bangs]. It generates [panties bangs] ignoring [forehead]. It seems to be possible to avoid this problem, but it requires some ingenuity and a number of attempts.
Also, [frill] of [panties] is effectively used to make it a fashionable item. Sometimes it even becomes a bra. So called [SDGs].
This has been the biggest issue for me in training. It's especially problematic in contexts with more specific subjects, where it seems like the model is competing between bringing in information about the subject forehead with information from the LORA.
I think it's partially because SD has an absolutely bonkers bias towards surgical masks due to recency bias when the LAION dataset was catalogued during COVID. It pattern matches anything you train that has similar mouth coverage to be a subtype of those masks, and you have to train the model away from that to create new associations. (I've run across this before in other LORA's as well).
“No one cared who I was ‘til I put on the mask...”
\;p