Truly doing the Lord's work

Porn is the cornerstone of all modern tech advancements. lol. seriously though. Cameras, lighting, compression formats, streaming formats, most of these are advanced most rapidly by the porn industry. (seriously)

@BreezyHeezy Is the porn industry training their own models though? I would have to believe they are, but maybe there is some legal problem there where they don't want to be sued.

@ToxicBot Well, I'm sure that the porn industry is working with A.I. to some degree, and my guess is, since they have terabytes of data in each studio, image and video training without legal issue would be quite easy, since they own the rights to the training data.

@dtwr434 Oh interesting regarding the anatomy issue.

I do wonder if having the male anatomy in the training data is the main helper, assuming that your training set contained genitals in every video.

In fact, I'm almost certain of this. I just tried the same prompts I used for the Lora and with both 1.3b and 14b it returned something that didn't have much if any knowledge of male genitals.

I'll look into training a bit more since I've got a 4090. Did you train this one locally?

@dtwr434 Thank you! I'll give it a shot and hopefully my "report back" is a 1.3b version of this ;)

You get it to stay with prompt descriptions. Otherwise, it'll treat it like typical fluid.

@BreezyHeezy The prompt descriptions provided didn't get it to stay for me. Suggestions?

@sdfaceswap I didn't have to do anything special to get it to stay beyond the given prompts in the examples, though saying it's thick helped a tad. It's still not perfect, but this is is easily the best I've seen so far.

@Fresnelz Yeah, "thick" definitely helps but "thick sticky goopy" or other similar prompting will help even more.

@sdfaceswap Yeah, trigger words like "thick sticky goopy" work really well, as well as descriptors like "the thick sticky fluid sticks to the skin", or something like "the cum is thick like grease/glue", or "the cum stays in place on face/hair/body/clothes" the best way to test your prompts on something like this is to do like a 2 or 3 second video and just render a few versions to see what's consistently working for that particular scene before rendering a longer video. With that said, a prompt like "the thick sticky cum sticks to their skin because it is thick and sticky like grease(or glue, or icing, or mucus, or snot, whatever else you can think of)

@sdfaceswap I tried commenting before but it didn't save. Yeah you can use words like "thick" "sticky" "goopy" "slimy" "slime" "jelly" "petroleum jelly" "lotion" "icing" "snot" "viscous" "syrup" "molasses" "axle grease" "lithium grease" in your prompts, that will change the way the fluid dynamics are perceived in the model. You can also be more descriptive in general. I.E. "the thick sticky goopy cum is thick like grease(or glue, or slime, or whatever works in your prompt) and sticks to their face/hair/body making a goopy mess." or "a sticky mess" or "a goopy mess like slime" or "slimy mess" any other type of descriptor set such as "the cum has the consistency of hand lotion mixed with petroleum jelly" or "the cum is thick and frothy" or anything else that would work. descriptors like "thick viscous fluid" work well. These base models are trained on absurd amounts of data so it can pull from non sexual sources to quite convincingly approximate something like the fluid dynamics of semen, or spit, or urine, or snot, you just have to tell it what you want. Think out of the box if your prompts don't work well. Basically, the model is squirting a fluid out of something and then it will react like the average fluid that its trained on unless you describe the fluid's specific properties, "sticks to their skin" "stays in place" "is thick and goopy like molasses" If you overdo it, it'll cause it to go in the opposite direction, making clumpy clusters instead. So you gotta find the balance for how to word it based on your own prompting style to get it to work. But you can use these ideas to get SD or flux or recraft to generate completely realistic looking semen on a photograph without ever mentioning cum or semen by just describing what you want and using various types of fluid descriptors, and the end result would rival a well-trained LORA.

@BreezyHeezy Haha, Civitai, gotta love it. I think any time you comment on a post it never shows until you refresh the page. Thanks for the additional info!

A few additional terms I find work quite well is "...that glistens in the light". or "...lands on her face in clearly defined semi-straight lines" I find this returns a better definition on the liquid. Basically, don't be afraid to over-describe in this case.

@BreezyHeezy Yeah that's the direction I started heading with some better (but still inconsistent) results. Thanks for the tips!

i2v works for me. At least the movement does, but I haven't been able to get the actual liquid to appear. Though that may have been because I had the lora set to double blocks and that also messes up the generation in t2v (censored genitals and no liquids)

Comparing with and without the lora the concept is vastly improved though results are still a gamble. What seems to help is putting some cum in the reference image to get things going.

To generate cum, decrease the LoRA strength or increase CFG. You can get it to generate literal bucket loads of exploding cum. It's a lot of fun, and by a long cumshot the best LoRA I've ever used.

@mentallydifused haha it's so funny. I used the workflow available for download here and edited my prompt a little to include stuff like 'massive' and it came out like a firehose lol

@mentallydifused Can I know what your data is? LoRA strength?increase CFG.?

@funscripter627 ah thought it was just me having that issue on i2v. I still haven't been able to get the liquid to appear either

@Aeralynaia Try lowering the lora strength and alter your prompt a bit with stuff like huge, enormous, etc.

@funscripter627 Nice I will try that thank you !

Wan really has an incredible understanding of liquids and its physics. Even without any lora, just describing cum as "thick white liquid sticking to her skin" in i2v gives very realistic results.

there just wasnt decent cum loras for hunyuan. doing it without loras using prompt like "white sticky viscious liquid" was way better than with loras.

Yeah same, really hard to get it to work properly with img2vid, would love some kind of update for that.
Edit: I think I improved it, I'm not sure but I think putting all the loras on double block or at least this one helps.

I tested on 0.5 strength, works perfect on wan2.1-i2v-14b-480p-Q5_K_M, probably on other models, too.

try with a logical setup if source image woman looks to the right or it would logical come from right side.
Add "from the right side a man enters the view" or if its close up "from the right side a man lower body enters the view".

I was actually amazed how well it worked though there can be funny accidents like fully floating penises floating into the scene.

@dtwr434 , are loras able to be extended (like additions ? )
was so impressed by this one able to start with image of woman that i was wondering if we can't extend the lora with like for the cowgirl one with a trigger word like gets_on / gets_off type of deal ?

Ak ok! I'm using musubi tuner and I think you can only set one resolution per training run. So one run with the videos and after that another run with the images?

@dtwr434 looked up the documentation again. Resolution can be set per Dataset in Musubi, too =) . Seems like I have to expand my Datasets now... good thing is that when the next big Video model comes out we all have pretty good datasets prepared right at day one ;P

Look in the files section above, you will see a workflow, it's in the 3kb file, you might also be able to load it by drag n dropping one of the images (if PNG)

possibly could add an image2vid node on that same workflow. I haven't tried yet. BUT the text2vid workflow works amazing!

Tried it with I2V and the cum is pretty minimal, which could be my prompt since as i'm just a beginner and trying stuff out. Previous attempts with only prompt and no LORA would produce a DELUGE of milk liquid coming out lol

@Gekko78 Interesting. I kinda get this same result. I am quite sure my prompting is fine. But also I noticed with another workflow I use that uses Upscaling the first pass seems to have a lot of liquid ( This is text2vid though I was using) but after a couple of upscale passes it just becomes barely visible. I guess that is something to do with how it's upscaling maybe. But I get great results just from a basic WF without upscaling!

Going to try now more i2v and see if I can get anything out of it!

@dtwr434 Thanks for the info. I was using Kijai's WanVideoWrapper nodes though, and I'm limited to 16GB on my 4080 Super. I'm surprised, but your node setup does work on this 16GB card. I may have to experiment with sampler nodes, as it looks like WAN can use many of them, if not necessarily all of them. I'm going to try and replace some of these nodes with Kijai's and see if that works. I just finished a generation, and the quality is much better with this setup, than the one I was using. It was pretty obvious even with the TAEW2_1 preview in the Ksampler node. I'm kind of puzzled as to exactly why, though. I'll have to experiment and see if I can isolate the issue. Anyway, thanks for the quick response. I'll upload the video here...

@dtwr434 BTW, I was curious, how many video clips & still images were in your training dataset?

@dtwr434 Maybe so, but KJ's wrapper has the option to offload all of the models, even Wan itself, into RAM to save memory. Unfortunately those nodes aren't compatible with most native nodes and I have no apparent way to connect the memory-saving model loader to any of the nodes in your default workflow. Unfortunately the default workflow for KJ's nodes results in poor quality videos. But, I suppose it might be moot, if going past 81 frames per generation causes problems. I am able to generate 832x480 videos with your workflow. Have you tried extending videos by using the last frame as a starting point to continue with? For that matter, I wonder if I can link a Load image node (with the last frame) to VAE encode link with the t2v setup, and set the denoise factor to 0... Haven't tried that yet. Not sure if the t2v model can handle that?

As for the LoRA, is it possible to train a great deal more videos & images (of the same quality) without it overloading the VRAM? I figured you were using diffusion-pipe - wasn't sure if the number of vids/images directly affected VRAM usage, though. Also do you mind if I send a message?

Instead of even 81, you should (personal experience) generate about 65 "only" to get the best results (don't expect 100% good generations no matter what). Once you have 65 frames that are good, you decide of one of those frames as being the probably best one to continue from. You extract this image (I use Avidemux but there are nodes in ComfyUI to do that). And continue the animation from it, 65 frames again. And so on. Good "long videos" = video composited smaller ones.

Generating less frames will save you shit loads of computing time. And you'll end up with much better (more control over what's happening) results.

@SD_AI_2025 I haven't had any luck with continuing an animation without shifts in color, where there's a slight change in coherence, or other minor issues. Have you avoided that issue?

@SD_AI_2025 Yeah, that's what I did for these recent 'cum blast' clips, that I uploaded yesterday - I stuck to 61 frames. But for a better, longer cumshot clip, it might make sense to do about 20-30 frames at a time, and have one 'spurt' per segment. That would allow for a lot more control as well.

@Agkistrodon great ideas here. Ive started doing this and using losslesscut to combine the short videos. I also realized if the faces matter to you using the start and end image feature really helps to keep the face constant throughout video extentions, Like I tried havinge the same face image for both and that is working well. At least in some other lorras like spanking ones ive tried. Going to try that same process out with the cumshot stuff too. Without doing ths, sometimes the subject turns their face sideways, so your end frame will be a side shot and the next gen will lose complete track of what the face looks like.

My prompts get it to work maybe 30-50% of the time. "a load of cum shoots onto her face from off-screen. Cum squirting onto her face." Sometimes theres like milk or a container they pour out of, but I have gotten believable outputs with that. I also used the negatives "Bad penis, detached penis, dildo." cause sometimes it would just make a hand that floated in and applied cum from a dildo lol

Would love to see a lora like this as well

Same It rarely works consistently for me in i2v when I prompt with cum shoots from offscreen. Ive tried getting more specific but it doesn't seem to help. I get randomly popping up items that cum or cum starts shooting out of hands our mouthes.

I saw in another thread you trained locally. Was this using musubi-tuner? If you wanted to share the dataset, I have a pod in Runpod set up with diffusion pipe I can try. Would be interesting to see an A to B comparison.

@dtwr434 still pondering this haha. What images do you think we're missing

Alfredogomezpomm

Same happens for me with a certain workflow I use it's no longer listed here. But I believe in my case it was something to do with the upscaling that was used. It had like 4 preview windows the first pass you could see the liquid was there but after the final pass it was almost gone. So it had to be the upscaling method .. Other than that I am really not sure. However I am using the workflow from this page and it works great. I get good results I am happy with. Definitely not a lora issue but in my case it was absolutely the workflow issue

Since I'm not really competent enough to build my own, I tried several workflows (At the moment I need t2v, gguf, multiple loras) and I was underwhelmed most of the time. What works best, are the simple ones (as the one you provide here). Now there is liquid. :)

There are hundreds of videos posted everyday. Most of them HAVE the workflow. Drag and drop a video in ComfyUI. Or open it with Notepad. All you're asking, like hundreds of people on Civitai everyday, is already there ;)

+ OP already shared the very workflow you're asking for, just gotta read a bit: "[...] I've included the workflow I used for the preview images in the "Training Images" download."