Agree. This is a holy grail in computer porn visualization. I regret I can only give one upvote.

I agree, this lora found me when I was down and picked me up and is now paying for my college education. We couldn't do any of this without YOU, Cumshot Wan LoRA.

Yeah, it works very well. It might make more sense to use this, and to use a base image for I2V to start videos with, instead of just rolling the dice with T2V.

Can you use the gen to set the pose, then this lora to create the dick out of words and snow, then the doggystyle lora to have them doggystyle? I just got a different one to work like that.

+1 to this request, the detached penis oral sex loras, or the women biting penises loras are all sub par

@psspsspsspssspss yes please +1

I just posted my prompt and video of using the deepthroat lora and this one together, as one, like Voltron, to accomplish this. I did need a gen where the woman was kneeling. We could still use a dedicated lora for it, mixing lora on Wan does normally make you lose quality.

shift value in what node?

@gillyguthrie803 The one I'm using is ModelSamplingSD3. It's one of the core nodes. It should be connected between the Load Diffusion Model node and the LoRA loader, or any other node that needs a Model input. But it should be connected directly to the Load Diffusion Model node, without anything else between it, I think. Not sure, but maybe you could connected at the end of the chain just before it's plugged into the sampler, but that's not how I've done it.

To expand on this, I have the ModelSamplingSD3 plugged in at the start as well. I can dial the strength of the main model up to 1.5 which helps. I get a lot more when I set shift for ModelSamplingSD3 up around 12. But this is just a larger, nearly naturally possible volume. Dialing the model up to 2 causes massive video distortion and setting the Shift on MSD3 to 20 seems to be the same as 12. I cannot recreate the "cum cannon" effect. Keeping the clip to 61 frames and below is critical. I'd say 49 works best. I tested 33 and it doesn't help and that clip is just too short. So a tl;dr is Set core model somewhere between 1 and 1.5, add ModelSamplingSD3 and test it at 5 and dial accordingly (up to 12), 49-61 frames, IKSampler Steps set to 40 lead to more detail than 20 steps. Keep CFG on KSampler low. I found 4-5 is best. 6 is way to strict and just produces the same thing every time.

@EmblematicAi Shift set to 12? That's interesting, I never pushed it that high. I thought it might result in unstable videos. Did you mean you set the LoRA strength to 1 to 1.5? I thought above 1.0 might a bit too high. I'll have to try that.

Obviously, I'd avoid using 'cumshot' or 'cum' in a starting image. But getting a good lora for the male member is pretty much necessary, then you can describe the scene as you want it. The problem is that Flux (which is what I use to generate starting images) doesn't really know how to draw a penis. If you try that without a penis lora you'll get a meat tube attached to a man's abdomen, or something even weirder. Those kind of bad generations are bizarre and sometimes even a bit repulsive.

The lora I'm using is 'The Penis [Flux1Dev, PonyXL]' by DP_COUPLE_SWG.

It works really well. Otherwise, as always, the more descriptive the prompt, the better.

@Agkistrodon sweet thanks. I haven't use flux at all, only different pony models. Maybe this is a reason to try it out finally.

And yeah, if you prompt Wan for male nudity without any NSFW lora you get some pretty disturbing results too!

@logenninefingers888 Well, it might make sense to use Pony first, as it seems to do better with explicit material, depending on the checkpoint. Plus it's a lot faster - then once you get something you like, you can run it through an img2img workflow for Flux, to refine it, with appropriate LoRAs for the male parts, or whatever else. The end result will be a bit better, I think.

@Agkistrodon yeah i2i pony->flux is a cool idea, thanks.

@logenninefingers888 It depends on the situation, as for how useful it is. Sometimes Flux does certain concepts much better than SDXL & Pony, and in those situations, using Pony doesn't work very well. Flux certainly has more detail, but it's a lot slower than either Pony or SDXL.

how i can use this lora, please guide me

Yes, you must go out your way to prompt the description of cum on face in various ways to help.

bro how you use this lora

in which software bro

@vijayverma18 Using newest Comfy on Win 11, 4080, 64gb

I have that too , i managed to minimize the occurence of color shifting with those settings : Euler + Normal, CFG 5.0 , ModelSamplingSD3 Shift 2.0.

@mrrocko101269 today installed comfyui in lighetning ai still learning, will update later bro

Okay so...this is going to seem complicated, but it's easy, I promise. If you're not already in the world of ComfyUI, the easiest way for you to do it by far would be using a free program called Pinokio. Pinokio is kind of like...Steam for open source AI applications? You get Pinokio, it serves as the hub, then you go through and find the applications you want to try, and it automatically downloads everything you need and gives you a Web UI to use to run the tool. So normally the way to run Wan is via tons of nodes in Comfy, this does all that in the background for you and gives you a SD type Web UI, it's amazing. Please donate a few bucks to the creator of Pinokio if you can, he keeps it super updated and fixes any issues that pop up.

So download Pinokio. Download Wan 2.1 from the "Discover" page. You might have to play around with settings, etc to get it set up and to a point where you can easily create videos. Also what GPU you have matters here, you need to adjust the configuration based on your VRAM. Pinokio will download the models for you automatically but you can have it fetch different ones if you want to use ones for weaker hardware. But all the settings are in a really easy set of a few dropdown menus. Also if you are having trouble with settings or knowing what specific settings do, ChatGPT understands all of this and can walk you through which settings you should tweak for your specific hardware.

Get it up and running so that you can put an image into the Image to Video mode, at 480p (going into higher resolutions isn't necessary unless you're publicly presenting your work), give it a prompt, and it spits out a 5 second video (81 frames, this is standard) starting with your initial image that at least sort of follows your prompt.

Once you have done that, THEN I'd start adding in Loras. Loras work basically the same way they do with SD, you download them into the "i2v Loras" folder in your Wan folder in your Pinokio folder (Pinokio gives you a nice easy link to open the file location for each application also). So download a Lora from here, reboot your Wan page in Pinokio, click "Advanced Mode" at the bottom of the main screen. Once you have Loras added, you get a section that says "Activated Loras." Just select the Lora you want from the drop down list and it will automatically add it to your prompt behind the scenes, no need to do anything to your actual prompt. Below the Activated Loras dropdown there is a section that says "Loras Multipliers," and there you put the strength you want from the Lora, in decimal form (so 50% strength is 0.5, 85% strength is 0.85, etc.). I usually do around 70-85% depending on the Lora or what I'm going for.

It will take some trial and error, but I know with Pinokio anyone who can figure out how to use SD on their computer can figure out how to work Wan also. And if you run into errors (you can see specific errors in the Terminal section of the sidebar in Pinokio) just punch them into ChatGPT and 99% of the time it can tell you exactly how to fix them.

Good luck!

@jacw5566414 Ty for info! it is much appreciated!
I have one more question tho, since I can see you are very helpful :
how to import model checkpoints like Kijai-Wan2.1-I2V-14B-480P
into WAN 2.1 installed on Pinokio.
Cant find how to select it anywhere in the UI. Sorry if its completely obvious.

try also using this lora at low strength (i've been doing between 0.3-0.5, not that i'm an expert): https://civitai.com/models/1358184/wan-21-penis-cock-dick-lora-t2v-i2v

I don't know if it's sageattention per se, but I've generated maybe 10+ videos, none of them look anything like what is present in the examples.

I have sageattention running and I'm getting outstanding results (will publish soon). If you haven't got sage yet, you're dying. It's so fast, nothing compares. Look at the one-click installer from SECourses - it's pain-free.

im sure lea would love you

I haven't tried this one yet, but in my experience, LoRAs are very dependent on resolution.

You should try quantized models instead of lowering resolution! I can generate a 5s video in 5mn on RTX4080 12Gb with the Wan2.1-i2v-14B-480-Q4_K_S.gguf model.
This workflow is optimized and will automatically download the correct model you select: https://civitai.com/models/1385056?modelVersionId=1670750

@MetaGen Thanks so much! Finally got it to workk now with the GGUF models. All loras work now perfectly

This lora works really well. Share the specific error you're getting, maybe someone can help you.

832x1216 works fine for me in onsite-gen. I think that's 2:3 or 3:2 aspect ratio. Assuming it's maintaining ratio for the output.

How do I know if I have a 2 step workflow?

@Ropehead If you have two samplers (t

either ksampler, or customsampleradvanced) and they're linked up in some way, and you also see a split sigmas as well as a node which is cutting the steps in half, its a two sampler workflow

you're the man ty

Great job for letting people know this @DudeManDudebro

You can prompt something like she opens her mouth and extends her tongue

@Kiefstorm i,m very grateful , thanks!

Try the Ahegao Lora as well.

Same, did you ever find a fix?

@mrrocko101269 I just see your comments and no I don’t know how to fix it💔😔

Try adding meta data such as prompt, technique, any info about your workflow etc. Then the images vids will be accepted and posted civitai.

You can make your prompt more specific, e.g. use something like "big, high-pressure blasts of cum" or "sudden copious spurts of sticky cum" or similar. Just be creative. However, even with specific prompting, at the end of the day, you're still relying on a lucky noise seed. From my experience, you can count yourself lucky, if you get one good generation per 20 tries. Just be patient and give it lots of tries, if you have the means and patience.

Water-themed environments can help : "swimsuit", "beach", "swimming pool", etc.

+ keywords like "excessive amounts of cum", "splashing" etc.

@xnapx Great. I'll try it and thanks for the advice.

@qdr1en Thanks a lot, I'll test it... at least I know how to start. Thank you.

@Sh3pard I am getting quite consistent results with those prompts :
- positive :
"Dolly In, focus on the girl face.a high-quality hentai video clip of a (exciting irresistible:1.05) libertine slut receiving (massive cumshots:1.1) on her face (from out of frame:1.05).

You can clearly see (very excessive amounts of cum splashing:1.15) on her face (in slow motion:0.7). she ends up very wet and (charming:1.1).

uncensored uncut reportage. Natural , human. (facial art, bukkake).

Allegresse, allowed, impunity, unrestricted, admirable, wonderful."
- negative :
"(erotic, glamour), therapy, censored, restricted, intolerance, (recalcitrante),

lowres, static, captions, worst quality, low quality, JPEG compression residue, ugly, mutilated, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, malformed limbs, fused fingers, still image, cluttered background".

Try at your own risk ;-)

@qdr1en thank you !!!

@qdr1en I am pretty new to this. what do the numbers do in the prompt when you add them like that?

Use the Comfy Templates. -> Video -> WAN 2.1 Image2video Then Add a (load LORA) node (with this lora) . Connect it between (Load diffusion model) and the (model sampling SD3) nodes.

Then Connect the clip nodes to the lora. like this... ( Load Clip) -> (load LORA-thislora) -> (Clip text encode (POSITIVE and NEGATIVE) ) Document page ComfyUI Wan2.1 Video Examples - ComfyUI

If you still can't figure it out, ask me and I'll send my workflow. Make sure you are using the latest version of Comfy or it will not work.

@newJason can I get your workflow? Having a bit of trouble myself.

@dtwr434 Awesome, much appreciated!

I will set the CFG scale to 6.5 to make the generation results follow my prompt more closely.

@FLOW0308 why not like 4? because i heard text to video is different than text to image

@dominik1996d840 Oh, sorry. I'm using image-to-video (i2V). You can see an example in my files.

@FLOW0308 and for t2v? what i saw with cfg4 was i2v but i use text to video ^^ how can make the woman stroke the mans penis with just one hand?

@dtwr434 i tried with one of her hands and it worked

@FLOW0308  how many steps i should use for t2v? are 20 steps enough? i use wan 2.1 in pinokio

Thank you, the same goes for me

yes and for me

Drag the upscale factor node all the way on the bottom left and connect it to the one on the top right. That's all you have to do. If it broke just like mine did, you'll also have to connect the positive and negative prompts as well.

it would be nice to have something similar to this for hunyuan

@dtwr434 Ah that's fine. In the two days since posting I've dropped FramePack and moved on to COMFY :P

It always happens to me too. I don't know why.

Same for me. Let me know if any of you find a solution.

A little late, but I just want to second that I've struggled with this too. I suspect the training data didn't include enough wider shots. I've not found a better solution than closer framing of the start image for i2v, it usually works okay anyway, but I've not found a way to encourage the man to end up closer to the character when they are further away.

@dtwr434 I see :-) I hope they will fix that soon

Wondering this too

I saw this on the civitai help thread. This could be why it is down:

"Looks like they broke all new WAN LoRAs because the generators don't work with the new WAN model options."

@dtwr434 thanks for the heads up

@dtwr434 Tested and got good results. Thanks for providing an alternative.

@dtwr434 AMAZING thank you!

I had worse ones. I had one attempt in which the penis poking out from girl's pussy and start spilling ketchup lol. I guess I’m just too bad at prompting XD

I've seen blood come out, traumatizing. And many times a guy shows up with a penis in his mouth pointing outward

I have a couple videos like this too. No idea what that's about.

Having this problem as well frequently in i2v. Havent figured out a solution yet. Cum will randomly shoot out of peoples mouthes too

same+1, tried so many times, didn't get at least one better result

CausVid .30, cfg 3.0 - 3.7, euler_ancenteral / beta, Seems to work ok still experimenting. cfg to high causes artifacts.

2 KSampler??

@dnbfiend thx! it was the cfg i rised it to 3.0 and worked.

Same. Also having cum shooting out of hands and random objects. I keep trying prompts like "cum shoots from offscreen and hits the womans face" But even telling it I dont want a penis on screen I still get sprouting penises too.

@anonameguyman1234252 Yeah, just tried an off screen one and got literally nothing. Another one of a sex scene and I CANNOT stop it from making her mouth pour out cum. Very weird behavior.

its inconsistent too. Sometimes it will actually just shoot the cum from offscreen... other times with the same prompt and photo or style of pose it will fuck it up most of the time. I am finding pictures that are narrower vertically seem to work a bit better, like if Crop the chick super close vertically, maybe it realized theres no room to insert penises/fake penisese, whenever i leave extra empty space in the background on the sides it seems like weirdness happens a lot more. I am using i2v for all my stuff.

@anonameguyman1234252 I've also found that adding actions for her pussy in the prompt helps. If her hand happens to be close, you can add a "she uses her fingers to rub her pussy", and it will usually prevent the new dick.
When you do this though, the cum doesn't originate from the existing penis, it just kind of appears. So I think what's happening is that it doesn't recognize the penis in the image as a penis, which is why it keeps trying to create one its self.

@mobdik17378 So in the scenario where I insert a man already in an image next to the woman, a prompt that specifies the placement of the man and is descriptive worked better and prevented a rogue demonic dick from sprouting: "the man to the right on screen strokes his penis. Then cum shoots out of the tip of his penis in constant steady spurts as they hit the woman's stomach and legs" I tried this prompt before the better one (The man on screen strokes his penis. Then cum s hots out of the tip of his penis in constant steady spurts as they hit the woman's legs and body) and the womans pussy grew a dick and both dicks shot cum back and forth at each other... lmao. This seems to be helping the rogue dick problem, but I havent fixed the cum often shooting out of the mans mouth or hands randomly problem.

I am maybe going to try negative prompts like: cum dripping out of men's mouthes, liquids appearing or dripping out of mens mouths. Will see if that has any effect to stop the random cum shooting out of hands/mouthes stuff. If you find a solution or any more tips lemme know.

@mobdik17378 I think you are onto something. In images with decent resolution with an erect penis in frame already I get things to work properly already. If the cock is flacid/ semi flacid it gives me problems way more. I think it can't detect the tip of the penis in these situations or when the photo is too blurry. It seems to not like having a guy in frame with the woman too where it shows his face. Cus half the time it will shoot the cum out of his face. I dont know if it thinks the nose is a penis or something lol.

Add Penis LORA

Wan 2.1 Penis - Cock, Dick, , Lora, T2V, I2V

Wan 2.1 - Erect Penis, Cock, Dick, Lora, I2V

Just pick one and use it.

@FLOW0308 Thanks for the response I will try that. Just need to figure out how to add a LORA loader to the workflow I currently have, very much a beginner at this. Appreciate you taking the time to give some advice!

@FLOW0308 Hi, I tried this and tested with both LORAs, now I don't get as much with the guy entering with a plastic hot dog in his own mouth, but I do get cocks with sort of plastic coverings/unrealistic still/sometimes looks like a fleshlight on his cock lol. Will keep experimenting, think my prompting must be ruining it somehow

@Davemanblaze201 

I only use weights of 0.5 or 0.7; I don't really go for complete realism. I think I made something randomly yesterday, you can check it out. For that one, I only used a 0.5 weight. The two LORAs also have different ways of being used.

That video of mine seemed to use four LORAs:

Wan Cumshot: 0.7

penis lora: 0.5

Detailz-Wan: Detail Enhancer for WAN Videos: 1

Yesterday's new Wan Thiccum: 0.5

Prompt words are very important. I only use: "He strokes his penis back and forth" or "The man strokes his penis back and forth with his hand."

@FLOW0308 Thanks again for the response, appreciate your time. Going to try those exact LORAs/weights with a workflow (that i'm not sure is great) shortly. Is there any way to download the specific workflow you use?

@Davemanblaze201 

I've uploaded an original video file. Please try downloading it and putting it into your ComfyUI to see if it displays the workflow.

https://civitai.com/posts/17820676

If you can't see the link, it means it hasn't been approved yet and will take some time.

First, I can't guarantee you'll be able to use it because I've added many nodes for acceleration, so you might not have many of them unless you understand those specific nodes. However, you can try to understand it, as I also started by using other people's workflows and then modified them to my liking. But again, there's no guarantee you can use it because everyone has different nodes installed. I also just learned and used things haphazardly, even reinstalling the main program many times. This is because when I tried downloading other people's workflows, I couldn't understand many of them either. I just disassembled other people's nodes and recombined them; if it ran, I used it.

You can see if it works for you, but I won't be able to help solve any problems because troubleshooting node issues is too complex.

@FLOW0308 don't worry about that you've been much more helpful than anyone could expect, it's hugely appreciated. That's exactly how I've tried to get workflows working previously but i'm just at a much earlier stage than you haha. The link says Not Found so probably still being approved i'll check back tomorrow. Thanks again mate.

@FLOW0308 just wanted to say thanks for your help, big improvement now after following your advice and LORA tips (though now the penises are not in the guy's own mouths, but they do constantly detach and it's just a guy holding a floating penis in front of the woman HAHA) - sure I can figure out my prompting errors!

bahahaha I know your having a problem but a “guy walking into the image with a plastic hot dog in his own mouth” sounds hilarious. post a pic so we can see how funny/bad it is? I have just started videos as well and have been getting a lot of support. Check mine out. The workflows are embedded as well. Cheers. I look forward to you making something awesome and sharing it. (other than hot dog videos lol)

@b0bsm1th Haha glad it gave you a chuckle mate, it gave me one the first 6 times too! mostly got it sorted now but i'll absolutely check out your workflows, thanks for the response!

Do you mean like light flashes etc? Try to lower their strength, especially when using multiple lora's. Sometimes even lowering one by 0.5 can make a big difference if you use a lot of them together.

Cool will check it out. This wan cumshot lorra works ok in specific scenarios, but does its own thing so frequently that I rarely get the result I want. I feel like I am constantly fighting it haha.

Try "the man's penis shoots cum" 😃

Chances are your configuration is causing issues. It's a really strong effect, so experiment with other workflows and you might find a way to get better results?

@boobkake22 Thanks for the tip, but I still don't really even undrstand what a workflow is.

@yellowdog1969 Depends what you're using. I mean it more colloqually here: What tools are you using and how they are configured. If you're using ComfyUI, it means the literal files for it uses configure work. How are you making videos?

@boobkake22 I'm just using WAN through CivitAI. Using this lora only, since the others I like don't seem compatible. Img2vid

@yellowdog1969 Wish I had a better tip, but I've not messed with Civit's generation. My only real suggestion generically is to try different resolutions and original image frameings. I get really different results. I love this LoRa more than I can say, but it's wildly unpredictable with "directing the male actors" and the volume and consistency of the shot. It seems like closer images tender to look thicker. I suspect lower resolutions also tend to look more substantial (if at worse quality). Still learning the curves of it myself.

Well, I appreciate the advice anyway

@boobkake22 I use this for i2v too and I find it really goes nuts if the male's face is in the frame. But that is what I want to get working male full body and face in frame with the female. Half the time the dude starts shooting cum out of his nose face or hands and the lorra ignores the dick. If you have any solution for that let me know.

One thing I recently learned is that this kind of lora doesn't work well with multi-stage samplers. If you do use a workflow with two stages of samplers, make sure the guidance of the second one is set to the same as the first. This makes a huge difference to the cum simulation and quantity.

If none of this makes sense to you, don't sweat it. It makes no difference on Civitai but could be useful to those running local workflows.

you still having trouble with this? if you are stealing prompts and using the same starting images and still having issues, then yeah its probably configuration. you can get configuration as close as possible by making sure you
1) pick filter and choose "Made On-Site"
2) click any of the results and click remix and start from there

in general, even though i use an external generator, my prompts deliver roughly the same results on civitai, so you can use mine as reference if you wish. (if there are none on my profile turn off that filter in 1)

@cero I don't use the same images usually, often I'm using stuff I generated on CivitAI. I kinda got discouraged and so far mothings really helped. I'll look at your stuff though. Thanks.

@cero Yeah, saw your stuff. Same old story. You're getting pretty much what I'm going for but if I tried it, I'd just get tepid results and waste buzz.

Yeaaaa that would be great. I have the same issues with it working upredictably.

What you're describing is indeed the version I'm currently training, but I'm doing so for an anime style. I have two main categories:

'Off-screen' scenarios: This refers to situations where either the action or one of the characters is implied to be outside the main frame.

'On-screen' scenarios': This includes both characters being visible within the frame, and scenarios where characters enter the frame from off-screen. (I currently have a test version for this, but it still lacks emotional reactions and sometimes the screen tends to darken. I'm currently working on testing the second version.)

I've done some preliminary testing on the 'off-screen' scenarios, and the results are promising, though I might still need more source material. I plan to improve it further once I gather more data.

However, all my training is specifically for anime style outputs.

@dtwr434 This lora does a decent job of having guys enter the screen in my limited testing of that scenario. I'm more concerned about preventing that scenario. Using negatives helps, but the lora's influence is very strong and often overrides this. What would help me the most would be to have (in addition to this one) a lora that only focuses on cum shooting from off screen, and one lora that only focuses on cum shooting from a penis already in the frame. No overlap between the two. The existing lora is fully functional as it is, even if two new specialized loras are introduced.

There are worse things in the world than a lora being good at multiple things, but sometimes you want one that's amazing at just that one single thing you actually need. :)

@RiseKujikawa Yes. I've used this lora a lot and it makes dudes walk in from offscreen way too much when I am trying to just use a start image with a guy/dick already on screen. Would be amazing to have a version of the lorra just trained on that scenario so I could reliably get it to work. 

It works ultra fast and good with Lightx2v (Self forcing).

I struggle with this constantly too. One thing that I have noticed that helps is if you crop your image so its a vertical image with next to no space horizontally left on either side of the woman. Seems like not giving any extra horizontal space prevents the random plastic dildos and soda cans from appearing in frame.

I go into the same issue.
Adding "cock, penis" in negative prompt + adding weight to "(off screen:1.1)" gave me better results.
Also it is much more generous when the image takes place in an outdoor location.

@anonameguyman1234252 great, thanks for the response. I'll test it.

@qdr1en Awesome. Thanks... at least I'm making some progress.

Awesome ty. I primarily am interested in i2v realistic style but I will try it out.

Metadata missing or bad metadata with censored words.

pincopallinoAI How do I solve that? In practice?

janosgreka287 Use an image you generated on civitai.

Did you find a way?

@dmodmaster142 no.

Yeah this model is very strong, it's going to have it's way whether you like it or not

It always works on the I2V one for me, what CFG are you using?

same problem, what cfg or prompt do i need to use?

I had that issue too, solved prompting "cum IS SHOOTING" instead of "cum shoots". (with that phrase the cum actually started shooting from off screen, but the penis sneaked in in the middle of action. I also wrote "penis, hands, cock, dick, extraneous unrelated objects in frame". (that stupid dumbass of a robot sometimes put a sauce bottle in front of the bitch ^^).

Try to reduce the strength of this lora.

discribe faces and hair

You can try using the fusionx face Lora on the low noise model. Just a thought. I haven't actually tested it.

Use "less Karen" , "more Heather" :)