from my understanding and low quality research from a lot of workflows and asking people on reddit, it seems the clip needed to be connected only the higher noise.

vAnN47 Cool. So far I've tried no clip, high, and low with wonky results but that was using the loras meant for 2.1. Giving this one a go now so fingers crossed! ^_^

It isn't apparent from a visual workflow how the internal data structures are working. But as people say, it seems one clip pass can provide data to high and low iterations. And this needs to be HIGH, else the clip data won't be available until the low pass, since high comes first.

You cannot escape the training data! If the LoRA includes image as well as motion/concept, the image will bias its output. Most recent LoRAs on this site seem poisoned, by which I mean elements not connected to purpose come through strongly when used. The answer is better training methods, but maybe fast and easy training- the type available to most- inherently has this problem.

blobby99 Thank you, I thought the training data was to blame as well. That kinda sucks. I need to sort out my loras and find the ones with the least "face data". All this testing and tweaking just slowly makes one lose interest you know. :D i2v at least fixes some of the issues.

The OP's example video's seem more varied than my own. In all my videos the girls look 99% identical. Maybe better prompting helps and not just counting on the random seed.

I just tested with all other cumshot loras on my pc and got all identical faces to this one. Then I disabled the lightx2v lora, ran 20 steps with 3.5 cfg and it helped some but didn't fix it a 100% imo. Something wierd is going on. Maybe i'll try another workflow or go back to Wan 2.1 for the time being.

Maybe it's cfg-zero-star or some other optimization doing this. Could also be comfyui going though many updates atm. and messing with stuff (I'm on all nightly builds). A pity testing takes so long to do. a minimum of around 20 minuttes is normal but just around 5 minuttes with lightx2v. I had some problems with identical face with Wan 2.1 but nothing like this. Well maybe in the future someone adresses this.

I feel dumb lol. :D

Rumaben I finally learned to train loras with kohya_ss with my dual 5090 7985WX system.
Now I got to figure out what tools I can use to do Wan 2.2 Lora's. I figure that with such a system I can feed in more training data with different faces and poses.

Any hints would be appreciated.

stewie69 Sorry I know very little about training. I just do local inference. Only tools that I know of is:

https://github.com/kohya-ss/musubi-tuner

https://github.com/tdrussell/diffusion-pipe

https://github.com/ostris/ai-toolkit

Two Nvidia 5090's is impressive man!! :O Try asking reddit if you can stomach it. :D That or perhaps a great discord channel.

dtwr434 Yeah, it's very weird and I'm not sure if this 2.2 release is gonna take off. They really should have tried to make this one model.

dtwr434 I think despite it being strong at 1, this works really well when you turn it down so its a net positive. Before with 2.1 single model when you turn it down, its a sacrifice between it forgetting its function versus the face being overfitted. But from how the dual models works with each other, it sidesteps the face issue a bit because the low noise model doesn't have a blank slate to go ham on the face, it has to work with what its given.

But it depends on your cfg value and how many steps you're doing on your high noise pass. Did you try turning up the cfg for more flexibility and increasing the number of steps?

dtwr434 Did you give up?

dtwr434 Maybe you could release it as an experimental version, maybe some manages to get good results through experimentation?

dtwr434 thanks for your reply. Its a great Lora

dtwr434 can you use tensorboard on the log files. look for the epochs with the greatest loss, around where the slope has levelled off?

dtwr434 Look at this tutorial, it really helped me. This is how I'm selecting which lora's to test. https://civitai.com/articles/83/using-tensorboard-to-analyze-training-data-and-create-better-models I am only trying character lora's but you could also do 1 frame samples to see how the faces turn out. I use swarm UI, load up wan video model, then use the grid generator option to swap all of the selected loras from tensorboard. Create a 1 frame grid and compare for likeness. I know yours is a bit different but perhaps a similar process.

try turning off the motion lora. Even the new 2.2 one destroys the fluid

Kiefstorm Motion lora? Not sure what you mean.. All I'm using is this lora..

Kiefstorm as always, please show your lora with only this wan2.2 lora implemented, seed, example picture.

I never got any even close to that you got on pictures, using only this lora, without any fasters lora, without everyting, no matter what prompt means.

And they told me - try comfyUI - it's a bit harder, but it flexible.

ComfyUI is pretty helpful, cause you will be searching and installing every freaking unnecessary node to achive same result which you already got.

And now I'm about cumshots. I tried many loras, many seeds, many workflows, and everything I knew from my experience - that's a real luck to achive facials like on example pictures.
HOW?

(Show me your workflow - no, show me yours which is so good to produce everything from pictures with different seeds (maybe not really perfect, but, I'm really not achive anything even close to this examples, with whatever lora you suggests, weights and everything else)).

Or nothing, or stuff like FuckedUpFacials stuff.

Agreed, refuses to create any cum

forfreelsd368 alright mr sassy, I just re-created the sample video and uploaded it. I tagged you both so you can grab the workflow

Kiefstorm very very thanks. It actually works.

forfreelsd368 I uploaded a new workflow, this one is I2V, that incorporates a bunch of helper loras to significantly boost speed. When I get time I will build the same for T2V https://civitai.com/models/1847579

Kiefstorm thanks, already find it :)

Lora matters, applied them on different wf - works great.

forfreelsd368 I am finding out new information every hour. Just play around with lora strength. Sometimes 1.0 HIGH & 1.0 LOW is better. Sometimes 2.0 HIGH & 1.0 LOW is better. Depending on the LoRa and depending on what other action LoRas are involved. I'm still thinking having High Noise Lightning disabled is the best for most situations

Kiefstorm Agreed, I am getting good results with high at .6 and low at .5

try supplementing with magnesium pills.

2.2 jerks a lot, that why the less cum

In the past, I would put that the penis is connected to his crotch in the prompt. It fixed it in the times I tried.

Wi-fi Penis LMAO.. thats a first.

@jeyjeysig Sure beats Cellular Penis. Talk about blue balls.

you probably connected a wrong node into Ksampler as input. it looks for a one dimensional tensor(array) but what it gets is 2D. Double check seed, steps, denoise and cfg. They all expect 1-D values

Google AI Studio. Upload your json. Explain your problem. Enjoy in seconds the solution.

dtwr434 appreciate all your efforts!

The 2.2 T2V works great in I2V for me.

I use "static camera" in the positive prompt and it works. Might also try "zoom in" in the negative prompt if it still doesn't work.

I usually just put the prompts for this one, and ended with "the breasts are bouncing".
Works out fine that way.

I have quite a few different Lora's combined but it seems to deliver a very persistent and realistic looking cum, although results in a "one man bukkake" kind of amount on her face... but if you wish to try:

Lora: Weight:
facials60 0.7
wan_cumshot_i2v 0.4
wan-cumshot-i2v-480p-22epo-k3nk 0.4
wan-thiccum-v3 0.7

Confused as it says 2.2 in the post? Hmm!

@dtwr434 I think I put the I2V in both fields and one with high strength and one with lower strength in the respective areas. Did a bunch of trials and got it working on some. I’m pretty new to this so navigating it has been confusing. 🫤 thanks for the reply!

Been struggling with this too, limited success with "POV" and mentioning the viewer.

@dtwr434 It's for GGUF.

@dtwr434 I can't even get the 2.1 version to add a cumshot in i2v with wan2.2, so sad

Hmm, maybe it was just my imagination, sorry

@dtwr434 When you say 2.1 always ended up better is that your 2.1 Lora running in Wan 2.1 or Wan 2.2 (low noise only?)

@dtwr434 Thank you for that crucial info. I concur from spending weeks trying to get the best cum as Wan 2.1. did get excellent results I was not able to replicate with Wan 2.1. Some examples in below (futa vids at the bottom)
https://www.deviantart.com/stash/221hjvfy6qov
https://www.deviantart.com/stash/22biiby8qg4j
On Wan 2.2 however I was able to get good results with Dasiva's template: see example at https://x.com/BigJuggsAI/status/2019779614810276008

@dtwr434 Thanks for the reply. It definitely struggles with 2girls, If you ever re-do this lora please add more. Since my original comment I picked up another cum lora that I've been using with this one with good 2girl results, this one just seems to handle the fluid better.

My two cents on something like this. In the description theres the type of triggers words. Some times its a plain copy paste. But for me I mainly use a uncensored LLM to describe my image and motion im looking for with the triggers words. Mix it all together and i get fairly very good results.