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
Trained with Wan I2V 14B.
FAQ
Comments (36)
Yes - we need a T2V version!
Gotta clear some HD space for my linux partition so I can clone the T2V repo, but I'll try and get to that ASAP.
@dngstn32 That's great! Thank you very much for your kindness! Your work is amazing!
@rafaaguiar990170 Sure thing. Just FYI, both the bride and black fishnets in my examples were done with the T2V model / workflow, they are NOT I2V. So I'd say grab it and give it a try... there's definitely room for improvement but it still works. :)
@dngstn32 That's great to know, the results are very good! I'll try it, thanks!
None needed by design, and it seems to have worked. I didn't put any in the training data. Just roll with the prompts in my examples and you should be good to go.
@dngstn32 Ok, thanks! And congratulations once again for the excellent work!
This one is powerful, I accidentally said reverse cowgirl in a cowgirl i2v and she went through a warp zone, but by the end she was beautiful reverse cowgirl.
Lol, nice... I haven't played with the weight much, just kept it at 1.0 after trying to lower it and ending up with lovely Stretch Armstrong dicks. Face consistency on I2V could be better too, I'm not sure how to accomplish that.
Very nice!
Please make a mating press next time :)
"My second LoRA... I've made some drastic changes to my captioning approach, and the results show some nice improvement."
Would love to know your approach?
Sure... let me update the description of the LoRA so I can include hyperlinks.
Not sure if this also works with BWC. I'll have to test it. If not, could you consider adding one in the future? Thanks.
Just tried a few times I2V... it's not perfect. The dataset focused primarily on dark ones, for... size reasons. I haven't tried adding images to the training data, but I'd imagine that could help.
@dngstn32 bwc is very possible with wan and xp checkpoints/loras (flux is still in limbo) but i can see how it's much harder in terms of fining real content to train off (which is how i imagine you made this) since real data creates far better loras than AI data. That being said, if you ever find a way to create convincing bwc at these kind of sizes that would be awesome since that's definitely a missing segment right now in terms of loras. Thanks for the effort if you ever decide to take on the challenge. Either way i appreciate your efforts/talent in loras in this space.
Are you using I2V? If so, it'll work with literally any skin colour; i've used it on white, black, and green with no issues. You can see this in most of the top-rated videos on this page.
How do you train? Locally on a PC? What software do you use?
Yep, locally... I have a 4090. I use diffusion-pipe running via WSL, in Ubuntu.
Yes, same as me – same graphics card and also WSL. I’m wondering how you managed to train on video? Every time I try, I get some errors.
I prepare video clips of 3 seconds, 16 fps, and it still doesn’t work.
Could you please send me an example of a config.toml file from your video training?
i described my issue here:
https://github.com/tdrussell/diffusion-pipe/issues/207
@CyberAImania I used the training data .tomls that @dtwr434 provided here: Wan POV Missionary - T2V v1.0 | Wan Video LoRA | Civitai. I don't think I modified any of the settings, except to point at my checkoued out Wan directories etc.
@dngstn32 Do you blur the faces in the training videos?
@BuffaloSymphony I didn't, no. This particular dataset doesn't exactly show a lot of faces, though. ;)
You might want to change the loras name, I feel alot of people have been using your lora just for how good it is even with the intention for it only being BBC
Nice work. Check out Joycaption if you want something that can do NSFW. It still needs fixing up manually though. I've also heard there's an uncensored version of Gemma.
Thanks, I've been using Joy Caption on single frames from the video and then adding manually, which is pretty time consuming. If only there were something like Joy Caption for porn videos...
good stuff. Very versatile too, moreso than you would think based on the title.
Your LoRA works very well! 👍
One issue is that in the second half of the video, a hand suddenly appears to touch the butt. How can I avoid this sudden extra hand?
Did you figure out something that mitigates this problem?
@lebu0009203 I haven’t found a one-size-fits-all solution yet, but adding "hand" and "hands" to the negative prompt words can sometimes work.
@SAY_AI yeah same. I put stuff like "ass grab, hands, hands appearing from offscreen" in the negative and that also helps mitigate the issue
Sorry that was my hand
Details
Files
bbcRideWan_e30.safetensors
Mirrors
bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
148_bbcRideWan.safetensors
bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
WAN_bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
bbcRideWan_e30.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.