Wan 2.2
V1: Initial Version fast movement less bounce
V2: More bounce, a little less speed, more WAN 2.1 like
Works good with my sample workflow:
https://civarchive.com/models/1598638?modelVersionId=2212578
This lora is for bounce walk videos. It's quite good in the continuation of the original image.
Used I2V-14B-720P as base.
Since ppl were asking about my training Workflow:
I have used 7 videos with about 4 second length, with faces cut off. I have described what happens e.g.
Here some captions:
A brunette girl, is sitting on the bed of her bedroom and standing up, she wears white bikini, her breasts are bouncing, her face is cut off, she is walking into the direction of the moving camera, noisy image quality
A brunette girl, is walking through the floor of her office, she wears a white top and and a black skirt, her breasts are bouncing, her face is cut off, she is walking into the direction of the turning camera
Then I used Musubi-Tuner with default parameters and trained about 32 epochs and then 16 epochs with larger training set, with my 4090 it was done in about 1h 35 min.
That's all :)
Description
FAQ
Comments (22)
Added a V2 with more bounce effect, a little less speed than V1, but more WAN 2.1 like! let me know what you think!
Thanks for the upload on V2 for Wan2.2. I did some tests and I recommend you put a note on the description. Some people may know this but others might not. Using a dual-sampler setup with lightning loras on the high and the low cause almost no bounce and it is as if you weren't using the loras at all. Taking out the lightning lora on the high sampler and using a CFG of 2-3.5 makes the lora behave properly.
After doing 8 steps (4 high, 4 low) with no lightning loras on the high, sampler of euler, scheduler or beta57, and shift of 8, the lora behaved properly. Remind people that lightning loras on the high sampler kill motion.
It is way better than v1! Good job man!!
@Arcane200 Thank you for your comment! That's a interesting observation. I agree that the lightning loras harm a lot in the movement although I did all my examples with high/low 0.8 lightning and 2 high and 4 low steps with dpm++ sde. So i think it kindof also works! Would be cool if you could post your before after results!
@ai_build_art Same things happens to me. With lighting lora added there is no bounce, it just animates without bounce lora. Once removing the lighting lora and increasing to 30 step and 3.5 cfg you can get the bounce effect. Check my last post. https://civitai.com/images/110104190
By the way this problem for T2V. May be it is a problem only for T2V, it maybe Ok with I2V I didnt try yet.
this is the best version so far! nice work!
Used together with The Walk model, it can effectively solve the problem of slow motion.
and cheaper than buying weed!
Great model, but unfortunately it seems to change the face on anime gens. :(
I am trying to use this in my wan 2.2 i2V but she isn't walking but acts like she is about to walk and keeps bouncing in the same place. Strength is at 1 for both high and low. What is wrong? My prompt is pretty much same as the examples here.
best check my reference workflow
https://civitai.com/models/1598638?modelVersionId=2212578
@ai_build_art One of the custom nodes have been removed from github due to some violation it says. So I am unable to install it.
@maaade you can remove the reactor node completely it just optional
this happens often when combining wrong loras with prompting.
youre probably running another lora which is trained on women standing still, you can try to decrease that loras strength or completely remove it if its not really necessary.
make sure to add a general NSFW model such as https://civitai.com/models/1307155/wan-22-experimental-wan-general-nsfw-model?modelVersionId=2073605.
general NSFW models are these are trained on lots of data, such as women actually walking, with the jiggle physics inherited automatically.
furthermore, and this is most important, sometimes you have to get a little lucky with the prompting, what works for me is not providing any more information than is needed, so that the clip and t5xxl dont get confused. if you put a lot of emphasis on boobs bouncing, they can interpret this as only bouncing on its own, instead of the bounce/jiggle being the result of the walking movement.
what works for me is keeping the prompt very simple and limited: "the woman is walking away" and consciously not giving the model any more info. if you use the right loras, the model will generate the desired bounce and jiggle movements automatically with it.
if instead you tell the model: "the woman is walking away and her boobs are bouncing. her boobs are jiggling intensely" etc -> this is understandable, as you want the right jiggle movement, however because you put a lot of emphasis on the bounce/jiggle, the model sort of dedicates 75% of its capacity to the bouncing movement instead of the actual walking movement.
hope this helps. for anyone running into issues, hmu
@10sor I was only running this lora. Strange thing is it works on other wan 2.2 workflows with the same prompt and lora. The issue I am having is in this wan 2.2 loop workflow -> https://civitai.com/models/1772470. Maybe you can give it a try in this workflow and see if you also get the same issue.
@maaade are you running all updated loras for wan2.2?
@10sor yea. Its not the lora problem but only with that particular workflow I linked.
@maaade i looked at the workflow, its very messy but i dont see anything weird, how many seeds are you generating? because i had this problem too and for me i just had to generate like 7 videos, its not really efficient but i think thats just part of the game with ai
@10sor Oh so bouncy walk was working for you with this WF? I tried dozens of times but the girl would not walk but bounce in the same place. I didn't have this issue with other workflows.
The physics is absolutely phenomenal but I do run into the same issue in all videos: The face keeps changing when I use it for wan 2.2 i2v. Is there a setting to keep the face intact?
Great Lora but when i test it, my woman is walking but his head go out objectif of camera, i have only her body !!
This is the perfect Lora. (I like the full speed one, as I can always slow things down in post.). Any hope of an LTX 2.2 version?
Details
Files
BouncyWalk02_LowWan2_2-000048.safetensors
Mirrors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
Bouncy-Walk-LOW-2.2.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
WAN 2.2 - Walk - Bouncy Walk - Wan I2V 14B v2 LOW.safetensors
走路胸弹跳L.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
LORA_I2V_Bouncy_Walk_L.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors
BouncyWalk02_LowWan2_2-000048(her breasts are bouncing).safetensors
BouncyWalk02_LowWan2_2-000048 (1).safetensors
BouncyWalk02_LowWan2_2-000048 (1).safetensors
Bouncywalk-low.safetensors
BouncyWalk02_LowWan2_2-000048.safetensors