Added a high lora for 1.1 as well, which is 50 epochs, does not need to be used at all times but can be helpful in certain scenarios, suggest weight of .5 to 1.0. Details of high lora impact down in the testing section.

Updated to version 1.1 which is 60 epochs(double the previous one) and should function well at in a weight range of 0.75 to 1.25, examples below in testing section.

This is a fairly simple low-noise only anus lora. The goal was to generate realistic looking anuses in the right spot, that aren't massively gapped or abused, when one is not visible in the starting image. In particular this was aimed to work in conjunction with my pov face sitting lora.

Prompt is simply needs to include the word anus somewhere for this to kick in. However, WAN seems to often have no idea where to put buttholes so the following helps a lot:

A woman presents her anus. Her anus is directly above her vulva.

Obviously if the subject is on their back describe the anus as being below the vulva. The lora should be able to add some specifics around what the anus looks like such as:

a small round anus

The anus is pinkish in color

Her anus is centered, slightly puckered

Given the training material for this often had both a anus and vulva, you can use it to control the shape and color of the vulva too but I've not tested that extensively. Critical keywords are vulva and labia in terms of prompts.

Testing

I tested this using base WAN 2.2 I2V Q8 and the lightning loras. Nothing else in the mix so i could get as good an idea as possible of what WAN knew vs was introduced by the lora. I decided to test both high and low loras at different weights with each other. The results are below in the grids(final frame and 4 second frame). You can see that while low lora alone does ok, it does produce a lot more detail with the high lora in even at a low weight. I've tried to minimize the motion in the videos and i'm not sure if that'll have a negative effect yet, so please do shout if it does. That said, high lora alone makes for some very odd bits so i'd avoid that combo! Good luck and happy butt testing.

4 Second Frame Grid

5 Second Frame Grid(final frame)

Dataset and Training info.

Input data set was 167 videos all pulled directly from reddit (you can guess the name of the subreddit). Batch trimmed to just the first 3 seconds and capped to 16 fps.

Data set was auto-captioned using JoyCaption Beta and DarkAges 70b.

Low Lora training was done at 512 for 60 epochs, 1 repeat and batch size 2 using diffusion pipe. LR was 2e-4.

High Lora training was done at 256 for 50 epochs, 1 repeat and batch size 6 using diffusion pipe. LR was 2e-5.

As always still not sure i know what i'm doing yet and open to feedback!