This Lora is trained on the WAN 2.2 I2V 14B low and high noise models and can be used to generate facial cumshots. It works with or without a man present in the initial image.
Update 20/07 - LTX 2.3 Experimental version
On request of some members here, I took my Wan dataset and tried training a LTX 2.3 lora with it.
To be honest LTX 2.3 doesn't seem to understand fluids as well as Wan and there are alot of different workflows and loras and configurations you can tinker with in LTX, that I really don't have the patience for at the moment.
So the Lora is here in case anyone wants to try it out but yeah....
This was trained without audio by the way.
Update 09/07 - V3 out
This is not a big upgrade here but it brings some small improvements. The twin jets can now be seen more often and the cum stream themselves should be move visible in flight. Previously sometimes it was just too fast.
You can try this to control things a bit:
the cum comes out in (thick|thin) jets
the cum comes out in (thick|thin) spurts
the cum comes out in varying thin jets and thick spurts
Update 07/01 - Alt V2 out
Happy new year everyone! Hope you are all having a nice start of the year!
After some more testing, I decided to also put here a High/Low alt version of the V2, which seems to give better results in some situations. I think it's mostly in more zoomed in shots but could help in other cases as well. If there's something you are trying with the normal V2 and you think it could work a bit better, try out the alt version then.
Update 01/12 - New version out
Not everything I was going for worked out with this version, I need to tweak the dataset a bit a train again but I was anyway happy with the results and decided to share it.
Some concepts are not quite there and there's quite often alot of cum but... Enjoy!
Please check the Training Data zip attached to the High version for all the captions used in the training. Videos have the workflow attached.
Update 02/10 - Attached workflow file (to the high version)
Trigger Phrases (some work better than others, the least tested ones are highlighted below):
a woman is kneeling in front of a man
a woman is lying down on her back and the man is on top of her
a woman is kneeling in front of two men
cum shoots out of their penises and lands on her face and eyes
(cum shoots out of the penis and lands on her tongue)
cum shoots out of the penis and lands on her face and her eyes and her mouth
(cum shoots out of the penis and lands on her face and her mouth)
cum shoots out of the penis and into her mouth and tongue
big cum load
(high speed cumshot)
(thick cum load)
(cum strings)
she has her mouth open
she has her mouth open and her tongue out
she has her mouth and eyes closed
(she nervously smiles)
(she smiles)
(she flinches)
she frowns, she hates cum
she hates cum
the man strokes his penis back and forth with his hand
the men are stroking their penises back and forth with their hands
the woman strokes his penis back and forth with her hand
the view is from above
the view is from the side
big close-up
close-up
Works pretty ok with 6 steps and Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors at strength 3 (high) and 1.5 (low).
CFG 1 - Euler Ancestral
Description
Check notes of High version
FAQ
Comments (55)
Thank you I'll test to see what it's going to give thank you for your share ha I have a question with what software you do this lora wan 2.2?
I trained it with the Musubi tuner
@Civitaiwolfxx thank you for your return again you have done a good job you are the best currently on the las style with ejaculation the rendering is really amazing
thanks man, appreciate it
no files?
Sorry, was splitting it into two files instead of a single zip and ran into some issues. should be fine now
i downloaded this early today and has been working well for me. has there been any changes made? do i need to re-download?
Nope, I just changed the way it was uploaded. its exactly the same. Glad to hear its working well!
Damn, these are so solid. Wan 2.2 14B really shines when it comes to keeping track of things like facial expressions during other actions.
this is going to be lots of fun 😁 ty!
The CUM LoRA models are getting ridiculously strong—everyone's results are amazing, and I'm still refining my materials for the next update
Now this lora gets the texture right!
Fantastic Lora. The only downside, is, impossible to use it for titfuck. Even at 0.5 there is cum that goes out of her mouth. Even when not mentioned in the prompt:
cum shoots out of the penis
If anyone know how to make the guy cum without having cum that land on her face or goes out of her mouth, when doing something else that facial, titfuck, jerking off, etc...When the penis is not directed to her face.
Let me know, please.
Yeah, I guess every video in the training is a facial.. sorry
@Civitaiwolfxx Your Lora is the only one that can produce realistic cum. I really hope you will make another one not for facial :)
I can try to look into that :)
Yes, when I prompt women to stick out their tongues, they already have semen on their tongues before ejaculate
@linkzero7535999 They've been busy ;D
Can you by any chance provide a workload file, I tried the settings you recommended, i also tried other workloads that work with other loras, but the results are far away from those you get
Workflow you mean ? let me check
Attached a workflow (as training data)
@Civitaiwolfxx Yeah, workflow, my bad. Thank you!
@agent_agent794 did it work for you ?
@Civitaiwolfxx I'm waiting for my current Que to finish generating. After that I will get the nodes I need, restart comfy and test the workflow. But sure as hell I hope it works considering that from the pomo videos it's clear you have done an amazing job training it, a lot more realistic than other cum loras i saw.
@Civitaiwolfxx I have a question. I have to download the unet models.
you say: Download the required 14B_Q4_K_M.gguf models and save them in your unet path (download the exact same filnames as you see in the "Unet Loaders"):
You mentioned Q4, but the file names in the unet loaders are Q8. Which one should i get considering I'm running a 4090 and 64GB of ram.
@agent_agent794 well, I didn't make the workflow or add those notes there. I just gave you the workflow i'm using. with a 4090 you can use the Q8 of course or probably even the non-quantized version. I have a 5070 and use the Q8 without any issues
@Civitaiwolfxx Well, it didn't work. The output is very hazy, the chick is opening her mouth but no cum/cock/man to be seen. I even tried with an image that has cock already close to her face. The res is 720, 5 seconds long. I tried 3 different images, all have the same result. Any idea what should i check/adjust?
@agent_agent794 do you have all the components from that workflow, the lightx lora etc ? that workflow was taken from other generations 1-1 so it should definitely work
@Civitaiwolfxx yep that was it, i had them but in the workflow they had 00005 or something at the end that i didn't have in my names, once i removed it it works perfectly. Thank you!
@agent_agent794 glad to hear !
@javbukkake i renamed the files before submitting them here but the workflow I posted had my original file names. should be fine
To be honest, the semen texture of this lora is not as good as the other LORAhttps://civitai.com/models/1922973/f4c3spl4sh-cumshot-i2v-wan-22-video-lora-k3nk,But this LORA action seems to be easier to control, and it can be achieved with a simple prompt.
Well, maybe its up to personal taste in the end. The one you mentioned comes across as not that great to my eyes but anyway, thanks for trying it out :)
@Civitaiwolfxx Forgive me, I'm just trying something new, like coping with various styles of images with some fixed-format prompts, I found that your LORA allows the character to maintain facial consistency instead of adding something unwanted like eyelashes or something, thanks for your work, it's still a good LORA, thanks for your hard work.
@linkzero7535999 no worries and thanks :)
both works great together. This has more control over the prompt than f4c3 and with f4c3 you control the amount
@jevus Can you share the weights of the two in H/L, thank you
This LoRA alone with the recommended settings and prompts does not give me good results. In fact the cum is thick as paste (not in a good way) and the amount of cum is just enormous (again, not in a good way). Don't know what's wrong here. In combination with other cumshot LoRAs on the other hand this LoRA works pretty well. I love the "she hates cum" reaction. ☺️
Hey, did you try the workflow I attached here?
Tried your workflow but cant get past to the Low Pass, runs out of memory. I tried increasing the blocks to swap but no luck. on a 5070 and still extremely new to this
Go to file -> Export -> this exports your workfile to a save location you picked. you can now feed this into your LLM of choice and troubleshoot. i learned comfy this way in a couple weeks of intensive genning and gooning
I'm using a 5070 as well and it works for me. How much RAM do you have ?
how do you guys achieve high quality video? i'm new to comfyui and wondering to use the img2vid. used the template workflow but the result kept getting blurred. any workaround or i'm missing something? workflow share would much appreciated. thanks.
Did you have the lightx lora as well there or just mine ?
If you're not using lightx speed up lora, then you have to increase the iteration steps quite a lot to not get blur.
If using the lightx speed up loras and get blur or grain, I suggest trying to raise the low-noise lightx weight.
@sarcasm83334 @Civitaiwolfxx thanks for the answer, i managed to fix the issue beforehand and yes it was the lightx lora which i didn't use it before. also i switched to gguf from diffusion model since i'm on low vram.
another question, so i have generated 720p quality but apparently it was not as close as most of the works that people posted here.
1) are you guys generating 1080p?
2) how many steps and cfg you guys are using? and is it on High or Low or Both KSampler?
3) do you guys start off with low resolution and then upscale the video afterwards? if that's so any recommendation node for upscaling(ignore this if upscale is unnecessary)?
@semiramisxxv I can say on my end I'm just following the workflow attached, so 720 pixels on the longest side. I guess i can sometimes increase it a bit if the source image for i2v is more rectangular than square. I don't really like how things look upscaled but that's just me.
Don't know about the other generations
Sometimes I forget to remove or bypass this when I'm trying to generate normal stuff, and semen will randomly be flying at the people from off-camera. It's pretty hilarious, lol. Thank you for your work on this, it's awesome. Good work
thanks man
This is working really well but for some reason as soon as the cum lands on her face it disappears. Not sure why that's happening.
Prompt that its thick, goopy and sticky. play around with that concept until you start getting results. Should only take a sentence or two.
I sometimes noticed this, and I added "the semen sticks to her face and accumulates"
@heylelheosphoros666 @idreezy Thanks for the tips it's working great now.
I notice the same regarding any cum that goes in the mouth.. really struggling to get any to actually go in the mouth and remain visible there... it's like the tongues have a magic cleaning ability lol.
I'll try the tips above about accumulating in the mouth or something. I really hope someone's cooking up a "cum in mouth" lora for that specific thing. :) (Not to say I'm unhappy at all with this Lora or any others, the Lora creators are an absolute Godsend!)
@sarcasm83334 There is a Premium Bukkake styled Lora, if you're familiar with their content. It focuses on cumshots in the mouth, but I'm not sure how good it is for the semen staying in the mouth. Maybe you could mess with it and test it?
Details
Files
Wan22_Cum_low_noise_1.V1.safetensors
Mirrors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
f_low.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
yan射L.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_C_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
wan2.2-i2v-low-cumhelperV1.safetensors
Wan22_Cum_low_射精颜射.V1.safetensors
Wan2.2-I2V-Cumshot-V2-LOW.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors
Wan22_Cum_low_noise_1.V1.safetensors