This LoRA generates women in the face-down-ass-up position, seen from the back. Trained on AnimeFullFinal with 100+ images. This is a difficult LoRA to use - as are most position-focused LoRAs. If you enjoy this LoRA you could buy me a coffee: https://chownyou.gumroad.com/l/coffee
This LoRA works best from 0.6-0.8. you MUST use a model that understands hardcore - revAnimated will generate monstrosities. Lower weights sometimes generate a woman in the position by herself, and sometimes floating penises - or attached penises. It seems to vastly prefer anal sex, but was trained on vaginal images as well.
The activation word is fdauposition, but there are a host of related words you can use to nudge SD in the right direction: 1girls,all_fours,anus,ass,ass_focus,ass_up,balls,bent_over,big_ass,big_penis,bottomless,curvy,dat_ass,doggy_style,duo,erection,feet,female_focus,female_penetrated,female,femsub,from_behind_position,from_behind,huge_ass,huge_cock,kneeling,large_penis,looking_back,male_penetrating,male/female,male,maledom,penetration,penis,pussy,sex_from_behind,sex,spread_legs,testicles,thighs,vagina,vaginal_penetration,veiny_penis,face_down_ass_up, bulldog_position, backsack and top_down_bottom_up were all cleaned from the tag lists; their combined influence should be represented in fdauposition.
I can't get this position to show up normally, and the openpose preprocessor doesn't like to work on similar images, so this LoRA is the best way I've found to generate them.
Description
V1.0
FAQ
Comments (5)
This works great but seems to keep generating pics with watermarks or logos even when they are in the negative prompts. Are they properly tagged in the dataset?
The dataset is scraped from R34, so they are only properly tagged when people tag them correctly. I usually fix logos with img2img by cranking denoise to 100%. I can see about releasing an updated version with manually edited tags
I cant get it to work at all?
what do you get, do you get no influence on the output image at all or does it change it in unexpected ways?
If you get no influence make sure you're on stable-diffusion-webui >1.1 with the extra networks extension. Run webui-user.bat from the command line so you can see the output.
If the output is surprising, work backwards from an example image. The model and prompt have a huge impact on the lora working - use the same model to start. You can crank up the lora weight to get a result as well, it'll just over-bake the image and put balls everywhere
ty, i think this does things well enough considering how hard poses can be!
do you have any plans in the pipeline for this pose, but facing the opposite way, i.e, facing the viewer top-down_bottom-up?




