So, it was painful to watch, that many checkpoints had problems with tags like 2girls or lesbians. Usually, missing limbs, feet, hands, arms fused together, crazy positions etc. So I created a LoRA to try fix it. I've generated about 2,4k images using different checkpoints, with different prompts, including various styles, angles etc. Then I've selected about 240 images that pass my filter of "good lesbians, good 2girls", and used this data set to train a LoRA usind different setting.

This is just step 1, I will try to improve this LoRA over time. For example, right now, dataset include images tagged with CivitAI autotag feature. On step 2, I'll try to train lora with manually fixed tags.

But hey! I want to see a LoRA that can pernamently fix problem of multiple characters generating. So if you want, use this dataset, tweak it, create LoRA merge or something. Do what you want, just do it right. Maybe together we can achieve success on lesbians generating on Pony models :D