Welcome to DR34MSC4PE's latest WAN LoRA: C0WG1RL + REV3RSE C0WG1RL
9/6/2025: It's finally ready to share - Our WAN 2.2 Reverse Cowgirl Text-to-Video LoRa.
This thing is actually pretty crazy. The usual prompts/triggers still apply: Use r3v3rs3_c0wg1rl for when she's in the reverse cowgirl position. Use c0wg1rl when she's in the c0wg1rl position.
You can also use different camera angles, such as: Close-up detailed view, wide angle view, side angle view from the left/right, low/high angle view looking up/down, etc.
Try adding camera movements as well - "The camera slides slowly to the left, creating a parallax effect", etc.
This lora also works GREAT with a Text-to-Image workflow. Create reverse cowgirl images and add motion with our I2V Lora.
Have fun! It should work with other 2.1/2.2 LoRas.
Please share workflows that work great for you!
Please generate responsibly!
8/27/2025: At long last, v1 of our Cowgirl/Reverse Cowgirl Wan2.2 14b lora is LIVE!
Be sure to use some variation of 'she's positioned in the reverse cowgirl position'.
Sample prompts:
r3v3rs3_c0wg1rl, a nude woman with light brown hair small breasts and a navel piercing straddling a nude man with a visible erect penis penetrating her vagina. The man's hands grip her hips from below as she rides him in the reverse cowgirl position facing forward with a slight devilish smile. His penis is thrusting into her and his testicles are visible below. They are in an urban street setting at night with neon signs in the background including a prominent red and blue "NEON" sign. Cars are parked along the street and buildings with lit windows line the background. The perspective is from a low angle looking up at the woman. The lighting is bright highlighting their bodies against the dimly lit street. The woman's legs are spread wide and the man is lying on his back beneath her.c0wg1rl, a nude woman with light skin long blonde hair and small breasts straddling a man who is lying on his back. The woman's legs are spread wide and she is sitting on top of the man in the cowgirl position with his erect penis penetrating her vagina. She has a navel piercing and a focused expression on her face. The man's lower body and erect penis are visible while his upper body is out of frame. They are in a dimly lit room with a fireplace in the background casting warm light and shadows. Logs are stacked in the fireplace and there is a wooden shelf with various items on the right side of the frame. The perspective is from a slightly low angle looking up at the woman. The lighting emphasizes the contours of their bodies and the texture of their skin.REMEMBER: You can use both r3v3rs3_c0wgirl, or c0wg1rl, depending on her positioning. r3v3rs3 works best and can work for both positions.
This LoRa works GREAT with Kijai's WAN2.2 I2V 14B finetunes found here: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/I2V
8/18/2025: WAN 2.2 TI2V v2 is READY!
After a few weeks in the smoker, she's ready to be served. Introducing v2 of our Cowgirl/Reverse Cowgirl Wan2.2 ti2v LoRa.
This LoRa works BEST with Kijai's WAN2.2 TI2V 5B finetune found here: Wan2_2-TI2V-5B-FastWanFullAttn_bf16
T2V and I2V Workflow's have been included - Sage-Attention and Triton are strongly recommended. An install guide for ComfyUI can be found here.
Sample prompts from the dataset have also been included. Feel free to use these as a base for your prompts.
Yes, penises can be a bit... long. We'll hopefully get that improved in a future iteration.
08/12/25: Introducing our Wan 2.2 TI2V 5B Cowgirl Lora.
This Lora is a work in progress. Epoch 280 seems to have the best overall performance during testing, but updates will be made as we get our training & dataset settings dialed in.
For WAN2.2 The trigger has changed, c0wg1rl now denotes the traditional cowgirl position. r3v3rs3_c0wgirl for, you guessed it, reverse cowgirl positions, which this Lora excels at. The WAN2.1 Loras remain unchanged, c0wg1rl is still the trigger.
Use the same types of prompts:
r3v3rs3_c0wg1rl, A nude woman with light skin small breasts and long blonde hair straddling a nude man with a muscular build and light skin in the reverse cowgirl position. The man is lying on his back on a beige couch with the woman facing forward with her eyes closed and mouth slightly open in an expression of pleasure. His erect penis is visibly thrusting in and out of her vagina. The woman is leaning backward and her hands rest on his chest and her legs are spread apart.
The background includes beige curtains and a white flower-shaped floor lamp. The lighting is bright highlighting their bodies and creating a clear view of their skin tones and textures. The perspective is slightly above and to the side focusing on their upper bodies and the action. A selection of the training captions have been provided to give you more of an idea on prompt structure.
Before you comment - yes, movement can be static, or positions can be wonky - please experiment with different phrases in your promts, and different tweaks to your settings.
You may have to add language such as "She bounces up and down", "The man's penis thrusts in and out of her vagina."
I've found the DPM++ Sampler family seems to work pretty well.
Before you ask: Yes, a 14b version is in the works.
Please experiment, and don't forget to share your findings and creations with the class!
===========================================================================
DR34MSC4PE is
c0ur4ge (trainer/qa/inference code) /
ERA5ER (captions/tooling code/data/qa)
Check out other models by DR34MSC4PE
Like our work? Consider donating for coffee: https://ko-fi.com/dr34msc4pe
PLEASE GENERATE RESPONSIBLY!
Description
Trained on WAN2.1 -14B 720p BF16
FAQ
Comments (37)
Nice! I notice your example I2V vids don't have the cowg1rl keyword, were they trained without it?
They were trained with the c0wg1rl keyword, but I was still getting decent results without using it.
Just tried your lora . Must say it works and it's amazing . Great work. I'll definitely be making more using this one.
HOLY Moly. Just saw this...
If this works anything as well as DR34MJOB then quality.
Time to do some testing. ;)
Lemme know what you think! ;)
@ERA5ER It's Quality! I was having issues with it seeming to effect the quality of the face but switched to some new workflows and up'd my output res.
Finally happy enough with my results to post them. Queued up now so will no doubt be up later tonight.
Oh and I randomly noticed the balls have a mind of their own. Might be my fault.
Still testing... ;)
In my case it doesnt seem to output at 720p but way beklow that. Is there a way to mend that? Maybe via prompt?
I'm not too sure. You're using 1280x720 and the 720p i2v model? What are your specs?
@ERA5ER Im using the 720 i2v base model. Resources are "Bouncing Boobs - Wan 14B" (bounceV_01_I2V_720p) and REV3RSE C0WG1RL -- Wan 2.1 T2V & I2V LoRa (I2V-720p). The starter picture has a size of 1216 x 832 px
@x3n0forge When you say it doesn't output, what do you mean? What happens? How many steps?
You try using the workflow from one of the sample images I posted?
@ERA5ER The resulting animations I create have a size of 480 x 688px but they should be more closer to 720 px? I mean, they work, but they are just too small.
@x3n0forge Make sure you're setting your width/height node to at least 720x1280
I m new to this. can you guys help me with workflow?
@ERA5ER Ok, I will try to adjust the settings accordingly
Utterly amazing work again guys! You are special when it comes to LORAS! Really hope you will do a doggy and missionary one as well, as nothing beats your Wan I2V stuff. Thanks soo very much ❤ 👍
Really appreciate it, thank you! Missionary and Doggy are definitely in the pipeline. Still working on putting together the datasets. Stay tuned!
Fantastic! I will be using this a lot more going forward! 🙌
Probably the best and most well-trained Lora on this site. Generates a perfect shot every single time with great prompt adherence
How did you train it? What dataset and which configuration settings?
Because I noticed most WAN lora's aren't half as good as this one. Not sure what the magic is but props to you man. Truly stunning work
Thanks so much, really appreciate it!
I put together the datasets myself, curating about 75-100 3-4 second video clips, as well as 75-100 high resolution images. Both video/images have a mix of the same position from different angles, settings, lighting conditions, and focal lengths (wide, medium, close-up, side-angle, etc.). I auto-caption all the images/still frame from each video using JoyCaption Beta One in TagGui. Then go through and manually make corrections as well as describe motions/movement in the video clips.
Training is with diffusion-pipe on a rig in my garage with two 3090s (gets great temps in the Winter-time). Train on Rank 32/64, AdamW8BitKahan, lr between 2-5e-5, on ONLY the video clips; testing epochs along the way to see where it starts learning the position. Then once it seems to have a pretty good grasp of the position, add in the images for another handful of epochs. The images help add in detail/sharpness. Again, lots of testing at each epoch until you feel like it's "done" to your satisfaction.
I think dataset/caption accuracy is most important. I like to think of it like the saying 'Abs are made in the kitchen', meaning you can train as long and hard as you can to get Brad Pitt abs, but you're not going to get good results with a sloppy diet (dataset).
ERA5ER This is incredible advice! Thank you for the elaborate process and for not gatekeeping it. Wish you all the best
Is there any way to prevent him from closing his eyes? Writing prompt words without closing their eyes is useless. Basically, every time the characters in this Lora video use them, they will close their eyes
You bring up a good point - much of the dataset has the woman closing her eyes in pleasure, or looking down (which captioning may have mistaken for closing her eyes).
I'm in the process of piecing together an updated dataset for a v2 release to try to improve upon a few issues. I'll see about getting more content in there where she keeping her eyes open. No eta on that.
In the meantime, you may just have to experiment more. Perhaps there's another face/body LoRa you could pair with this one that could help?
@ERA5ER v2?!?! :😀
Sign me up. Though how you can improve this I don't know.
Excellent LoRA! Strongly recommended.
Version 2 of the lightx2v lora is making this lora generate videos that are in slow motion. Version 1 seems okay. V1 of lightx2v is not great with motion though. To remedy this try combining it with the Accvid lora.
Is that using both lora at a lower strength?
Hmm I tried with accvid lora but wasn't very satisfied.
A little bit more motion but still very stiff
Try addin the Wan general nsfw helper, that one always brings tons of motion even with the lightx v2 but set it at 0.65 - 0.85 along with this reverse cowgirl on 1
ERA5ER The slow motion issue is with both your lora as well as the Lightx2v v2 lora at 1.0. I think I tried lowering your lora all the way down to 0.6 but it didn't remove the slow mo effect. This is no knock at your lora. It's the best. I just wondered, maybe others were having this problem as well. :) Strangely though I don't get slow mo with other loras only yours. I also tried setting the shift at 7 and 3 (orig. 5), no change. Could be my wf though, i'm using a few tweaks like slg, enhance-a-video and the new radial attention together with the flowmatch_causvid sampler at 6 steps.
damiang Thank you for the tip, I'll try that. :) I did a lot of testing and the best strengths I could find was with the Lightx2v v1. lora at ≈0.80 and the Accvid lora at ≈0.50-0.60.
Rumaben The same phenomenon occurs when I increase or decrease the Lora intensity, but it still feels like slow motion. I think it was probably learned with a 30 FPS difference during training.
Changing out my previous used Lightx2v v2. rank64 for the rank32 and adding the pusa lora at a strength of 1.2 seem to fix my slow motion issues. I'll need to do some more testing. Accvid never really gave me perfect motion.
Final comment about this then i'll stop, promise. lol. :D So I tested with the pusa lora. Best strength I found with it was around 1.0 maybe a little lower. Still not satisfied so i'm going back to causvid v1. and just accepting the little less movement from it. No slow motion with your lora when used together with the v1. causvid lora, but a more aggressive prompt is properly needed like adding the word "jumping" into the prompt and not just bouncing, that helps a little. :)
IMPORTANT: INFORM ALL WAN USERS! I DISCOVERED HOW TO USE LORAS 'NSFW/SFW' IN WAN 2.2 T2V, GETTING THE FULL POTENTIAL OF "HIGH NOISE" AND "LOW NOISE." THE RESULTS WERE IMPRESSIVE!! I ran it on an RTX 3060 6GB, 32GB RAM, 480x480, length: 41, generation time: 5 minutes to generate each video. The quality is fantastic! For those of you with more powerful computers, the results will be even more incredible. I used the models: WAN 2.2 - Q4_1.gguf FOLLOW THE STEPS BELOW: 1: Don't use acceleration/enhancement LoRas (such as Lightx2v or other) on "high noise"; use only on "low noise" with strength 1. (I'm using this version of Lightx2v LoRa): lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16 2: Use 'NSFW/SFW' LoRas on both "high noise" and "low noise" models. On "high noise," use strength 2, and on "low noise," use half the strength, with strength 1. You can vary the strengths as you like, but always keep the "low noise" strength at half the strength of the "high noise." 3: Use these settings in KSamplers: "high noise": steps: 9 / cfg: 3.5 / euler/simple / start_at_step: 0 / end_at_step: 4 "low noise": steps: 9 / cfg: 1.0 / euler/simple / start_at_step: 4 / end_at_step: 10000 Note: To improve speed, I set the Windows system to "best performance," disable unnecessary background programs and applications, and lower the desktop resolution. This significantly speeds up video generation. I hope this helps.
workflow? just upload a video gen with meta and we can just use your workflow. i found something similar but still tweaking, ill share when done. 3060 12g and im doing 640x640 in about 3 min.
side note on my last comment mine is i2v
5min slow for 41
It's actually the other way around, you put high noise LoRa on half of what the low noise counterpart is set to. High noise isn't supposed to have all structure, it's the noisy-base and then low noise pass adds details. Add the structure all on the first pass (high noise) and you'll get horrible results.