Attempting to use a more diverse dataset with this one, so the results are a little less "one trick pony" like my Missionary LoRA. Unfortunately you get the occasional contortionist show, but it's not very frequent.
As always, I've included the ComfyUI metadata in my example videos, so Save them and drag / drop into ComfyUI for the workflow.
For training, I followed dtwr434's very helpful Training Data guides for training WAN loras with diffusion-pipe, in fact I don't think I changed any settings at all for this one.
For captioning, I just drop in one of the first frames of the video into Joy Caption 2 to get a baseline. From there, I add specific things that you'd only see watching full video, like speed, movements, jiggles, shakes, etc.
Please share your generations! Your prompts and settings have helped me learn what makes a better and more capable LoRA.
Description
Initial version, trained on Wan T2V 14B
FAQ
Comments (17)
This is really great. I think as you found out in one of your showcase videos, it doesn't seem to care if you prompt for a slim/slender woman. The end result is great though, even if it can't be customized in that way. Heavier women aren't bad!
Lol... yyyyeah, that might be my bias when I'm selecting source videos. I did deliberately grab a few with more slender bodies, though. Let me share all of the prompts from my dataset, they should have some good examples.
EDIT: Okay, you can find a .zip with my training data .txt files that I used attached to the model now.
@dngstn32 What did you use to train them? Runpod or a local? I've been messing with diffpipe but no ability to train wan loras on a 4090 afaik
@jj43797771 Locally, diffusion-pipe installed in Ubuntu WSL. It's how I've trained all my LoRAs. I've got a 4090.
@dngstn32 Mind if I shoot you a pm? Nothing crazy annoying just curious about wan training a 4090
Can confirm that the training for this LoRa goes beyond motion and also forces certain body types. Works well for what it can do.
Dude theres never enough anal loras because so many of them don't work. This one does so good job.
For T2V probably. Sadly for I2V, above ~7 weight the style goes away and image is all smoothed/washed out. And with about ~6 weight, the style sticks, but then nothing moves much and/or the penis pulls the asshole like it was gum.
After 5 x 5sec attempts on different images, it doesn't seem to work much if at all with I2V :/ Too bad samples looked promising ;) ¯\_(ツ)_/¯
@SD_AI_2025 Ah, but an I2V flavor is training as we speak. I'll pull at 60 epochs (which has shown to be plenty in my previous I2V LoRAs) and see how it looks.
@SD_AI_2025 Ah ok it's not just me. Sad as all my attempts ended up super grainey and bad with i2v. I look forward to the i2v version!
@SD_AI_2025 I wonder if combining this with the detail enhancer lorra or another nsfw lorra would help. If I find something that works I'll re comment.
Do you have a link to that guide?
Yeah, it's in the original version of his Missionary LoRA - Wan POV Missionary - T2V v1.0 | Wan Video LoRA | Civitai
download the training data .zip
I swear to god this LoRA can brick your system. It crashed my system twice in a row and I had to reset the CMOS battery to get my system back and running after the second one.
Hi there...this looks amazing, i was wondering if there is a I2v coming....thanks for sharing
Details
Files
analFromBehind_T2V_e100.safetensors
Mirrors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors
analFromBehind_T2V_e100.safetensors