A WAN 2.1 lora to make breast bounce more natural and very realistic. Also a very nice effect for slow motion (slomo) effects or just add fast or snappy speed for faster movement animation. If you like my work and want to support me you can do so by donating Buzz or by paypal https://www.paypal.com/donate/?hosted_button_id=L87JAGTHDNBRA
Note: Please if you see any of my models somewhere else report or let me know! Thnx
Support creators:
by liking their models (many forget to do it)
by publishing your photos directly in the model's gallery.
by getting involved in their Early Access programs
Thank you for your support.
Description
Ver.1 WAN 2.1
FAQ
Comments (20)
I think you might have accidentally set the base model to "Wan Video 14B t2v", instead of 2.2?
I think you might be right about this. I'm gonna start some testing to be sure, thanks for the feedback my friend.
What does this mean? I can't use it with 2.2?
@dillion1920 No, just that the "model" section on the page was set to WAN 2.1 by mistake, so it was mislabeled and wouldn't show up under the right filters when searching the site. It's been fixed now though
@dillion1920 Is compatible with both but for some reason some people are having issues here on civitai Im using them just fine on Mage.space but im sure they might have another set up. then even I have the base model to the correct one WAN 2.2 that is sometimes I see it goes WAN 2.1 instead that's why I think maybe is a issue here on the platform.
@ThrowawayAccoun42414 did it work for you? or have try it yet? on Mage is working fine but here i have bad results and sometimes i see it bounce back to 2.1 I feel bad because I try to unpublish the model until I make some more test or give the platform more time just to refresh the cache or whatever setup civitai use. I will check in the next 24 hours if not I think I will take this one down temporarily until I manage to make it work here.
If it's wan2.2, do you have any recommended weights for the high and low sampling?
I can tell because when I train it the trainer dint had an option to train it low or high.
As far as I understand, there should be 2 loras for wan 2.2 - for high noise and low noise models. Uh?
now that you mention it, I think your right. Now im a bit confuse because when I train this lora it dint gave me any option to chose high or low and this might be the issue Gosh at the end I think I will remove this lora. Thanks for the feedback :)
@Artikuz_Ai I just confirmed by checking the safetensor's metadata headers: it's a WAN 2.1 T2V 14B Lora. You had it right the first time in your info section, but you might want to update your description section to change it to 2.1 :)
No need to delist it!
@ThrowawayAccoun42414 I see, Thank you very much for your assistance :)
@Artikuz_Ai I'm testing it now, and with WAN2.1 T2V I'm getting a lot of tensor shape errors, so despite the metadata value I think it might be a WAN2.2 model.... but it might be the TI2V 5B version of 2.2
I'll try out each and let you know which one(s) it matches with.
Edit: Nope, definitely not the 5B one. Still checking the others.
Edit 2: I'm an idiot and forgot that comfy defaults the WAN 2.1 T2V template to the 1.3B model, that's why there were errors with 2.1.
The metadata does have the training model file path as
main%2Fsplit_files%2Fdiffusion_models%2Fwan2.1_t2v_14B_bf16.safetensors
Which definitely does imply it's 2.1 T2V... but it absolutely doesn't match the 2.1 T2V 14B fp16 model (and has some horrific results tbh lmao)
I used it as a high-pass-only model on 2.2 T2V and the result actually looks pretty decent... regardless of what the heck happened, it might not be a bad idea to market it as a high-pass only LORA!
Yeah I thinks that's happening with the few error feedbacks from users that use Comfy. well I test them this time on mage but with wan 2.1 and works better somehow even when I train it on wan 2.2 crazy right. Thanks for testing and once you have tested the other I be reading your results feedback and thanks once again.
@Artikuz_Ai It's definitely a WAN 2.1 T2V 14B lora, but the good news is that most 2.1 loras work as either high-pass-only or low-pass-only loras for 2.2 too.
I'm waiting for a few non-LX2V-sped-up versions of my 2.1 testing videos to finish generating, and then I'll post some of the videos I ended up with from both my 2.2 and 2.1 tests
@ThrowawayAccoun42414 nice! thank you so much.
@Artikuz_Ai Hey sorry, work got me distracted with that whole "life" thing, so I was busy.
So I did generate some side-by-side comparison videos last week for you... but after looking at them I sort of realized that I was mostly just seeing that your Lora added big tits, but not a lot of actual increase in jiggle or bounciness.
I can still upload them if you want, but they probably don't demonstrate it as a physics/jiggling lora.
(Granted this was all under pretty lazy prompting, so I have no doubt that there is a way to make this lora good, but I wasn't able to focus enough on it to give a good example)
Edit: That all said: this lora did cause a heavy slant towards boob-bouncing behavior, plus an obvious bias towards sexy-superhero outfits with bouncing cleavage. I wouldn't call this a failure or anything like that as an attempt, it was a pretty solid shot, I would just recommend going for something more obscure than "jiggling boobs" as your next subject. It will be hard to isolate the boobs-jiggling part of your dataset specifically, so going for the holy grail of NSFW loras probably isn't an ideal learning step. Just my personal thoughts!
It's still pretty good overally if you're looking for some good titties™️, it's just a little rough if you want to add bouncing to an existing scene!
@ThrowawayAccoun42414 Thanks for the testing and the feedback, I think my mistake was not enough videos in the training dataset I just use like five 5s videos to se what I get and from there see what i just discover with you and your assistance, so yes it definitely need some finetuning is not bad like you say but it needs better prompting for a more refining output. I also dint work out much the prompt just to see how much the video influence over t2i data training a thing I have take the time yet. I never use jingling only bounce and big breast so I will add more keywords for sure on the next training. Once again thanks for your assistance and feedbacks.
Now is fixed and working just fine 😍
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.