Wan 2.2!
Re-trained the dataset on Wan 2.2, make sure to grab BOTH Loras (High and Low)!
Getting great results with K3NK's workflow - https://civarchive.com/models/1824027/wan-22-t2v-i2v4-stepskijais-wrapper-workflowk3nk
Update 6.17
I realized while trying to improve this LoRA that I never uploaded a T2V version, so... here it is. Enjoy!
-- -- --
My second LoRA... I've made some drastic changes to my captioning approach, and the results show some nice improvement.
Please share anything you make with this, I'd love to see some better prompts!
This was trained on WAN 14B I2V, on 45 videos normalized to 480P / 24FPS and trimmed to 3 seconds using diffusion-pipe. However, it works okay with the T2V model, and I've included some examples from that.
My captioning approach:
45 videos, resized to 480P, 3 seconds, 24FPS.
Ran each one through ComfyUI_Qwen2-VL-Instruct to generate a base video description, but this unfortunately doesn't pick up on any NSFW bits. This usually took a few tries on the same image, since the LLM almost seemed "disgusted" by the suggestion. :D
Grabbed my "favorite" frame and ran that through Joy Caption 2, then I manually combined the Qwen description and the Joy Caption Two caption to make the final .txt file.
Description
FAQ
Comments (33)
Fantastic work! Wan 2.2 version is perfect!
Does this support angles other than from behind? Do specific keywords need to be used to prompt this correctly?
I trained with descriptive captions, so no. And there are a few "from the side" views in the dataset, so you could try "from the side" or "side view." Not sure if it'll help... play with the aspect ratio / resolution, too. The majority of the videos in the dataset are 720x400.
aww yeah time to make a LoRa of ya boy so I can hop up on there
Thanks for the Wan 2.2 update!
Question: Is there a way to prevent the woman from squatting all the time?
(Was struggling with the same problem in the Wan 2.1 version)
Not sure. Try changing your resolution to something 16:9-ish if you're not already doing that. There's a lot more variety in that ar bucket.
I've been trying really hard, best I can do is get her to start in a different position and then move into the squatting position lol Honestly a bit of a testament to how well this works
dngstn32 Thanks - I'll give it a try!
love your work! More futa sex would be amazing!!!
It's a great lora, for once it's not slow motion hell when I use like so many others, so kudos for that!
One thing to enhance : for i2v, maybe cut off the head of the people in the trained videos so it doesn't modify the faces.
Right now it works pretty well in i2v, only issue is the face morphing.
Yeah, I thought about trying that or blurring faces for the I2V training when I get around to it. Thanks for the suggestion!
Wan 2.2 version is perfect!
Works great in I2V!
Nice, good to know... I've had others say that faces would get distorted so I was thinking about queuing up a training run with blurred faces on the Wan 2.2 I2V models. Might still do it for the hell of it, though. :)
dngstn32 it will be very cool, wan2.2 gives more control, I think your model will be used very often in the future together with others.
But there's only 1 I2V model. Do you load it in both High and Low? I'm still very confused why most people post I2V 14 B high+low, and others post only 1 LoRA. I'm even more confused when I read that it works well.
SD_AI_2025 The I2V model is for Wan 2.1. You only needed one Lora for both T2V & I2V with Wan 2.1. For Wan 2.2, they shifted to the high / low pass split, so I trained a Lora on each Wan 2.2 T2V model (high and low).
Let me re-label the older versions to clarify that.
SD_AI_2025 I load the "HIGH" and "LOW" T2V models for I2V generation.
@dngstn32 it would be cool if we can do where the girl facing the viewer because most of people that use i2v want their character face visible and consistent. But thanks for the lora i'm gonna try it tonight and post the result here.
this is the best lora on civit ai. no question
What gives the best results: using two LoRAs simultaneously with both high-noise and low-noise models, or only with low-noise or high-noise models?
这阴茎太长太丑了
trigger phrase plz
There isn't one... but you can check the samples for the prompts used. They're all T2V.
This Lora is absolute gold, will you make an I2V version for wan2.2 any time soon ? Thanks again for the awesome work
I actually made it a while ago... just haven't made any samples worth using for a post. Let me throw some together and I'll upload the Loras tonight.
Aaaand uploaded. Enjoy!
What's the prompt / trigger for i2v? its not a jump cut / scene change? i.e. the starting image needs to make sense?
@dngstn32 thanks a lot you're a legend
@dngstn32 on the first two i2v examples I love how her asshole is well used... please can you tell me which model did you used to gen these two pics?
Could you provide a simplified workflow for I2V?
I can't, for some reason, properly setup the workflow linked, some of the custom nodes refuses to work/be installed etc.
My specific problem is with the lora is that:
1. The guy is facing the wrong way(cock as a face) and the girl just slides up and down akwardly, even when editing the prompt to be more specific.
Details
Files
bbcRide_wan22_T2V_low_e20.safetensors
Mirrors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors
bbcRide_wan22_T2V_low_e20.safetensors