Updated for Wan2.2!
Trained with https://github.com/PGCRT/musubi-tuner_Wan2.2_GUI overnight on a RTX4080. I used 68 clips, ranging from 3 seconds to 6 seconds, for the dataset.
Both HIGH and LOW models are needed for it to work well. All the example videos were generated with Lightx2v LoRA. Workflows are included in the videos.
Recommended Settings:
Steps: 8 (4/4)
CFG: 1
Shift: 8
Scheduler: Euler, Beta57
Lightx2v LoRA at 2.2 strength for High Noise, 0.8 strength for Low Noise
Sample Prompts:
Cowgirl
dp_c0wg1rl. A medium-shot view of a naked woman with fair skin tone, wet, dark hair engaging in a threesome sex with two men on a couch in a brightly lit room with a tropical outdoor view. The woman appears to be in pleasure, with her mouth open and eyes closed. The lighting appears natural, likely from windows.Reverse
dp_reverse. A nude redhead woman with small breasts and fair skin is engaging in a threesome sex with two men on a bed with white sheets. The background shows a blue lake, green trees, and distant mountains. The woman has an expression of pleasure with her mouth slightly open.Gangbang
dp_c0wg1rl. A blonde woman with tattoos is engaging in a threesome sex with two men on a white couch, while performing oral sex on a man with large penis. The men have varying skin colors. The background shows a white couch, and the lighting appears to be artificial, casting a soft glow. gangbang.You can describe the details and the scenery as you like.
Buzzes are appreciated. I didn't put it behind a paywall so you can start playing with it right away.
Enjoy!
Description
Updated for WAN2.2
FAQ
Comments (26)
Wan 2.2... Nice :)
Thanks!
Awesome contribution! If you ever decide to work on this concept some more would you consider adding DAP/DVP?
Maybe one day. Preparing a dataset is some serious work and I'm a lazy person. 😂
@schwadorf understood... just ruminating 😂
DAP request seconded, so I can use it on men. Thank you.
@lucylawful999 Ain't gonna get any kink shaming from me 🤣
Greattt work!!
Finally. Been strugging to get this to work with other LoRA's. Cheers!
Looks really good am going to try on image to video even though its text to video just to see the results.
Please report back your findings! T2V loras generally work with I2V as well but I've not tested it.
@schwadorf It worked...Enough. I will post my results but bearing in mind I used it for animation but I imagine it would work just as well if not better for real life scenes
@Devilday666 Hello, how do I convert a real-life image into a video?
best NFSW lora
Amazing work, thank you.
In general, really good. Pose control is a bit inconsistent with the keywords. I cannot seem to get two people from the side where both men's faces are visible -- it's tending to add a third person if I try to get specific.
As a quirk, I'm getting slow motion unless I drop to 8 fps as the baseline with the workflows I normally use.
Can you try the recommended settings I provided in the description? I use lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16 lora for both High Noise (2.2 strength) and Low Noise (0.8 strength) and I don't get slow motion as you can see in the sample videos.
@schwadorf I made a 50+ or so more and had much better results with further testing, thank you. (I wish tagging data in Civit was less burdensome, or I'd post more.) I get a random percentage of slow motion clips, which is a bit odd, but fine. It can still get confused about legs sometimes in some poses, but seems to perform quite well overall. Generally clips are "hits" rather than "misses" which is par for most LoRA's really.
My only real complaint now is that it seems to want extremely similar camera placements and similar bodies without other interventions. It does what you show, for sure, but if you do take it further, a bit more variety would be welcome. (Or if you have suggestions about how to get it to switch things up, I'd welcome them! I've tried a bunch of stuff, and other LoRA's can bully it, but.)
I will keep experimenting tho, as I really like this LoRA. Thank you again.
@boobkake22 Glad to hear you sorted the slow-mo issues you were having. And you are right about the lack of variety, that is something I want to fix if I ever do a v2.
This was my first Wan2.2 lora and I wanted something that just works after having issues with the motion in the Wan2.1 version. But I think I overtrained it a bit and the pose got baked in.
What you can do to spice things up is, you can find a photo of your favorite pose (or grab a frame from a clip) and run it through Joycaption to have it generate a prompt for you. You can also add some lines for the camera movement like "Camera dollies in/out" or "Camera orbits around them" etc.
@schwadorf That's fair! Well I'd certainly love to see another version. One of my favorite concepts.
Joycaption is new to me, so thanks for the pointer. Curious how it does.
Works well with character LoRA.
Dumb question, how can I see the workflow included in the videos?
Click the download button on upper right corner of the video, then drag & drop it into ComfyUI.
@schwadorf Dumb question too，I only got a video player， when I drop the video into comfyUI ？
thank you , you are the best
Details
Files
Wan2.2_dp_v2_HighNoise-000020.safetensors
Mirrors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan22_DP_HighNoise.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
DoublePenetration_HIGH.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_Double-Penetration-High.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
DoublePenetration_HIGH.safetensors
Wan2.2_dp_v2_HighNoise-000020.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.