Minimax H3 has a hard time differentiating between anus and vagina, especially from the rear. This lora helps her figure out where to stick her finger(s).
Contains accompanying audio, so moans, etc work. Also works with character loras/ref workflows.
It usually works well, but sometimes you get finger warping in/out of anus. Also, minimax doesn't do well with hands behind a person, so it sometimes incorrectly chooses which hand is attached to the arm (eg, thumb pointing in wrong direction).
You should be able to prompt for which fingers are inserted, although it doesn't always follow directions (prompting for index finger alone may result in index + middle fingers for example).
Took me a little while to figure out how to train motion + anatomy. Basically a crapload of 3 second clips @ 512px and a bunch of images with both finger in and just anatomy @ 1024.
Workflows should be embedded in example videos.
Description
FAQ
Comments (21)
Amazing work! Works really nice with ref2va
Thanks!
Yes! I was happy to discover that even though it wasn't specifically trained for ref2va, it works well. I ran it on i2v with a doggy style lora and results were great for me.
@sugarfreewafers which doggy style lora did you use? Can you please share the result :)
@honryindian Posted a vid of doggy + self finger. Vid should contain workflow.
works amazingly well. t2v at 0.65 with no other loras and vanilla checkpoint is a sweet spot imo
Cool! Glad you are liking it!
Works, but why specific SELF?
@citivai846 This was my first H3 lora. I tried doing a larger range of anal play but wasn't coming out great. Narrowed down my dataset to help me understand why my training wasn't working (single concept is usually easier to train than multi-concept). Now that I've figured out how to train well, I'll go back and review my datasets for other play as well!
H3 has hard time with the holes? Really? To me it looks so big... so nice... so hard... so differentiated...
With the rear view, MM has liked to put a man's anus where his "taint" is (the skin above the testicles and below where the anus actually should be).
Thanks. Can you tell me where the workflow is embedded? The video has nothing on it, unless I missed it?
Hmm. To verify that the workflow is embedded, I re-downloaded a vid from civitai and drag/dropped to comfyui and it loaded the workflow.
Make sure you download the sample vid after clicking on the video, then use the actual download (arrow) button on civitai's single video display page. Otherwise video won't have workflow embedded.
Workflow is basically vanilla t2v: minimax fl2v convrot 8 + this lora + Pussy4nus lora (not required) + larryvh's turbo lora @ 0.8 strength + larryvh turbo sampler node. Simple scheduler + 8 steps.
@sugarfreewafers thank you very much for your message, friend. I am not new to loading up workflows from files. It's really not loading up for me as I did exactly how you said to do. Which video one are you talking about?
If this one does not work, I will load up my basic workflow and just try it from there. Thanks. :)
@jth3gravedigger625 Cool, just wanted to make sure you knew about that. It wasn't too long ago that I discovered downloading from the main page didn't include the metadata.
I tried with all of the videos and they all worked. Only thing I can think is that civitai downloads have been a little flakey lately, especially when a model has been newly uploaded. Try again maybe?
I'll also upload the workflow directly to the model page as an additional file.
@sugarfreewafers thanks, friend. your workflow worked as it intended. I have positive results. I will test more out of this. Good job and thanks for your helps. :)
Great work, thank you!! How did you force Minimax to understand where anal is? please advise. did you use some special captions?
maybe used normalization training? Most people don't do it because it takes longer
TLDR: vids for motion, pics for details. :-)
It was a learning process. I started out with a large and broad dataset (self, other person, dick in vag with anal fingering, etc). That created a bit of a mess. Some of the motion worked but lots of body horror. Too many concepts in one go, really.
I then simplified the concept to a single woman on her hands and knees anal fingering. The motion trained quickly, but no anus. I realized that the 512px/3sec clips i was training on probably weren't high enough resolution for the model to pick up on. I also didn't have many clips of no finger in anus, so the model didn't know what a non-violated anus looked like.
To add that information, I used krea to generate a large variety of 1024px images of varying body types, poses, finger in/finger not in anus. I captioned these because I wanted to make sure that I could differentiate poses and finger in/not in. I trained with both vids and images at the same time (vids@512, images@1024).
@pocketpie What do you mean by normalization training? Are you talking about what Ostris does in this vid to teach the model what the 'not' state looks like? https://www.youtube.com/watch?v=8Ug0dA4jXyY
I guess I kind of did normalization training, although I did it in a single pass by providing both the concept I wanted to train "finger in anus" and the normal state, just "anus".
This is a really great LORA. I've generated about 50 videos with it and DaSiWa 8step Hybrid v2 and most of them come out clean at 0.7. I bet you could go much lower with a turbo checkpoint and still get great anatomy distinction. Basically fixes everything wrong with H3 and anuses.
Hells yes! Thanks for the helpful feedback!