🎯 Anime Cumshot Aesthetics – Precision Load

THANKS^^

Is the Lightning version stable now? The previous one seemed to have lower image quality and less dynamic motion compared to 2.1, which made me switch back. I haven't gotten around to trying the Lightning version for 2.2.

BBBAAA2 THANKS

You can refer to the images I've provided. While it's not a perfect configuration, it's the setup I use to achieve my results.

THANKS^^

THANKS^^

THANKS^^ Every LoRA has something you can learn from. I also started out by learning from the very first WAN CUMSHOT LoRA. It's through his LoRA that I've been able to develop to where I am today.

Thank you for your kind words. I'm glad I could be of help.

Thanks^^

I have a file on the right named 'Training Data.' This is my workflow, and you can try it out. However, I can't guarantee it will work for you, as results can vary depending on your specific setup

I will try to adjust this in the next version, but I can't make any guarantees.

This is because while my source material is all translucent semen, the trained model's output is coming out a color whiter than my source. I suspect this might be due to the mix of anime and real-life imagery. Additionally, I believe the base model, WAN 2.2, has a tendency to perceive it as a color closer to white. To be honest, this current version is already the most natural one I've created so far.

I will do my best to try, but I can't guarantee the outcome, as I also don't know what color the final trained result will ultimately be.

Some people prefer a thick texture, while others prefer realism. My model is a result of trying to blend these two preferences, and honestly, I may not be able to guarantee the color.

Perhaps you could try other people's LORAs to see if you get better results. I've already done my best to create, filter, and maintain a high standard of quality for the source material. Since so much online footage has shaky camera work, which could result in shaky video output, I've tried to manually screenshot scenes myself to create clean datasets for training. Despite all my effort, sometimes the quality of the final result disappoints even me.

If you're willing to provide better facial cumshot source material—ideally with stable camera work and colors you prefer—you're welcome to upload it to your preferred platform and send me a DM. I'll gladly try to use it for training, though I still can't guarantee a good result.

This is because it's hard to find good, uncensored anime footage with the right colors. Honestly, finding anime that includes an uncensored penis in the frame is practically nonexistent. I've been searching for a long time and have only found three, believe it or not! And the quality was terrible, haha.

As for real-life videos, too many have shaky camera work. All of my current source material is self-produced, so it requires a lot more time. Thank you all for the feedback.

@irishpc491 The reason using only low noise still produces good results is that the training steps for high noise only account for 10% of the total, while low noise training makes up the remaining 90%. This ensures stability for general use. I use high noise specifically to train unique actions, expressions, and emotions that WAN itself can't generate. In other words, high noise is essential for me to achieve these distinctive effects. For general purposes, however, the low-noise version is essentially a fine-tuned version of WAN 2.1 and works perfectly fine on its own. I'm still trying to find the perfect balance, as using or not using high noise can sometimes produce very different dynamic effects, making it hard to say which approach is definitively better.

@mickyIammicky Hey! Thanks so much for sharing your experience. First, I’d like to ask if you used my LORA on its own or if you mixed it with others? Also, did you have a chance to look at my example prompts first?

If you're mixing a lot of LORAs, I'd strongly suggest you try using just my LORA by itself first. The effect of LORAs isn’t always better when you mix more of them; sometimes they can interfere with each other.

I haven’t personally run into any issues, but I know everyone's setup is different. If you have a moment, maybe you could try using my Training Data. I’ve found that when people use just my LORA with that specific data, it tends to work perfectly.

If you were already using it by itself and the results still weren’t great, could you tell me what the specific issue was? For example, was it the angle, the amount of fluid, the color, or did it fail to generate anything at all?

If the problem was the angle, I’m happy to say that I’ve already prepared more training data and will be updating the next version to hopefully fix that.

As for the fluid color, I can’t guarantee a better result. The WAN 2.2 base model’s understanding can be a bit strange; it might generate a perfect semi-transparent color one time, then suddenly give you a milky white color the next. I'm not sure what the root of that problem is.

Regarding the amount of fluid, I'm sorry to say my current skills are limited, and I don’t know how to use prompts to precisely control the volume.

If the issue you encountered was something else, please let me know. I'd be incredibly grateful for your feedback, as it's very valuable to me.

How are you using my LORA? Are you using it by itself or mixing it with others? Which base model did you use, and did you try my example prompts? I'll need to know your exact process to figure out where the issue might be.

@KubaSenpaii Do you mean ejaculation from off-screen, or the kind where the man walks into the frame and then ejaculates? If you mean off-screen ejaculation, I've noted that I haven't specifically trained material for that. But if you mean the kind where the man walks into the frame and then ejaculates, I have trained material for that.

Theoretically, I've never had this kind of issue. It's the first time I've heard of the "multicolored" problem. Did this happen with my prompts as well?

@KubaSenpaii Yeah, I specifically mentioned in my description that this version does not include an off-screen ejaculation scene. You probably didn't see it.

It will be open one day. I'm just unhappy that SeaArt AI took it without asking. But to be honest, I can't really stop people from mixing, copying, and rewriting it, right? I do understand what you're saying, though. It's just that I spent a lot of time on this LORA. It's not a famous LORA or anything, but it's a piece of work I trained very seriously. At the time, I just wanted to share it on Civitai. Maybe in the future, I'll choose not to make it exclusive. But honestly, I don't think calling it "exclusive" really has an exclusive effect, lmao. Thanks for the chat.

@LocalOptima I appreciate your agreement. By the way, I really love your Oral Insertion LORA. It's incredibly realistic, a level of detail I just can't seem to achieve on my own. I have a particularly difficult time getting the penis to look so lifelike.

I'd like to ask, are you using the official model combined with my LORA on its own?

Also, have you tried adding the "LIGHTING acceleration LORA" as an extra? I'd also like to confirm whether you've successfully gotten results when testing my prompts.

The weight for this LORA (both high-noise and low-noise) is set to 1. I'm asking this with the assumption that you understand how to use WAN2.2, as I've encountered many people who are still using the old WAN2.1 method, only loading the high-noise LORA, which leads to compatibility issues. I've also heard that using only the low-noise LORA can work, but I haven't tested it myself. In theory, as long as the image angle isn't too unusual, it should work successfully.

Please avoid using heavily modified "All-in-One" mega models for testing. I've discussed this with others before, and those who used these unofficial models with my LORA ended up with incompatible results.

I'm not sure about your usage or prompts, so it's difficult for me to determine what the problem might be. I just wanted to say thank you for the feedback.

This is a great question. Unfortunately, this is likely a problem I won't be able to fix at the moment. It stems from my very old source material and LORA, and honestly, I'm not sure how to improve it.

The biggest issue is that the WAN model has always had a problem of confusing fingers with the penis, which often results in fingers appearing to ejaculate. I suspect this is part of the problem. To get the results you want, you might need a specialized model like WAN2.1 Masturbation Cumshot for Wan I2V 14B. This isn't a problem I can solve myself, so you could try using someone else's LORA to see if it works better.

Another possibility is that my LORA models were not trained on any masturbation-related content, which is why they might not have a strong impact on your request. My LORA models are all trained for facial cumshots.

But I just looked at how other people are using it, and it seems they can get successful results as well. You might need to improve your prompting.

To be honest, I just had a chance to look at everyone's examples, and their usage methods are all very unique. A lot of the successful outputs aren't even in the direction I originally trained for. My guess is that the prompt itself accounts for a huge percentage of the result. You might need to try experimenting with your prompts a bit more.

I just took a look at your examples. It seems like your penis shots are facing the camera and are fully exposed. My material wasn't trained in that area, so I'm guessing that might be one of the reasons for the issue.

But I'm not entirely sure. I saw that HumanAd228 seems to handle this successfully as well. Maybe you could check out his work for reference, or even ask him how he wrote his prompts.

WAN2.2 requires the simultaneous use of both a high-noise model and a low-noise model. It's difficult to explain the principle directly, but in simple terms, both models must be loaded at the same time. The high-noise LoRA is loaded into the high-noise model, and the low-noise LoRA is loaded into the low-noise model.

If you're using the parameters I recommended, things should generally work fine. Just make sure not to exceed 81 frames—that’s the upper limit. As for any blurriness, I haven’t encountered it myself, and from what I’ve seen in other people’s demos, everything looks good.

Your description is a bit vague, so I can’t really pinpoint what’s going wrong on your end. You might need to go through some trial and error.

As long as you're using my setup—with the official version of the base model and not one of those all-in-one variants—there shouldn’t be major issues. If you’re planning to mix multiple LoRAs, you’ll need to experiment and test things yourself.

I haven’t trained on materials related to masturbation, so you might want to try mixing in approaches from others to see what works. Sorry I can’t offer much help in that area.

The main reason is that introducing too many different behaviors and labeling styles can affect training stability and cause fluctuations in the loss. That’s why I try to keep the behavior patterns consistent during training to maintain reliable and stable results.

Most of my training data has been focused on scenarios involving ejaculation toward others. So when switching to a masturbation context, there's a chance the model might force the target onto the subject’s own face—since the default target is typically the face.

You might want to try omitting the face as a target in your annotations to see if that helps reduce the issue.

So far, when I use my own images, I haven’t encountered this issue. I also haven’t heard similar feedback from others, so I honestly don’t know where the problem is, and I can’t really fix it. My image and character materials are quite diverse—they’re not all from the same person—so theoretically, it shouldn’t be getting locked into a fixed style.

@Refink I think I roughly understand what you mean. If the character’s gaze is inherently hollow, vacant, or like a dead fish stare, using my LoRA might end up replacing the original expression or gaze with something different. I’ve had cases where the character in the original image had a vacant, lifeless expression, and I wanted to preserve it—but couldn’t. My guess is that the WAN model itself doesn’t include such expressions, which might explain it. Unless I can expand my dataset in the future to include these kinds of expressions, I probably won’t be able to solve this. I’ve also tried describing “preserve the original expression,” but it still changed the gaze from hollow to lively. This part might be beyond my control—unless I get lucky and manage to generate a similar expression. Otherwise, I don’t think I can fix it. I’m not sure if your issue is similar to what I’ve experienced before.

I’ve tried training this type of LoRA, but the results were poor—mainly due to a lack of suitable data. I couldn’t match the quality of my current version. To get body-focused results, I’d probably need to train a separate LoRA just for that, but honestly, it’s too time-consuming. Even my current version took many iterations to get right.

To be honest, I’m feeling a bit stuck. There are already lots of similar LoRAs on the site, and many people are getting decent results by mixing models. I’m starting to question whether it’s worth continuing. Maybe I’ll revisit it when I have more time or better training material. For now, the newer versions I’ve trained just don’t match the original, so I haven’t released any updates.

THANKS

You can use the Oral Insertion LoRA with a strength of 0.3 to 0.35. That’s how I usually make the penis look realistic—you can give it a try.

@tinderizza I couldn’t see your image—it showed a 404 error. Maybe it’s because my dataset isn’t diverse enough. If you have materials like this, feel free to share them. I don’t lack training capability, just the raw materials. Otherwise, you might want to try other LoRAs. I probably can’t help much with this part. Have you tried other related LoRAs? How did they perform? Did you find any LoRA that seems to solve your issue? If so, I’d love to check it out or study it further.

One possibility is that you locked the seed and didn’t change it. Another reason could be that my dataset wasn’t diverse enough, since this is just a beta version. I’ll see if I feel like updating to a second version later. For now, the versions I’ve trained aren’t quite good enough.

@dagabagool Are you sure it was my LoRA that changed the girl's face? This is the first time I've heard someone mention that. I've created tons of anime girl faces using my own LoRA, with materials I made myself, and I've never had any face-changing issues. I've also used other people's anime face LoRAs and never encountered this problem. Did you only use my LoRA, or did you combine it with other models? Was your base model an official one or a merged model from someone else?

@tinderizza You can try combining my LoRA with others to solve the penis issue, since I don’t have much time right now. There are plenty of LoRAs available these days—you can experiment with mixing them. Also, Civitai basically doesn’t give any Buzz anymore. Honestly, I used to be motivated to keep updating, but now I only update materials when I feel like it. Otherwise, feel free to try other LoRAs—there are probably even better ones out there now.

@dagabagool I’ve seen a lot of comments, and many people who’ve used my LoRA say the results are great. While they’ve mentioned some drawbacks, no one has ever said it changes the girl’s appearance. That’s why I’m skeptical—you're the first person to bring this up.

@tinderizza @dagabagool I couldn’t solve the penis appearance issue back then because I didn’t have enough training material. But now there are many LoRAs focused on appearance—you can try mixing them. Personally, I always use a combination of LoRAs, and so far I haven’t had any problems. I use the Oral Insertion LoRA with a strength of 0.3 to 0.5, and it can change the appearance of the penis. You might want to give it a try.

All in all, just make do with it for now—this is only a beta version. Back when I first started learning AI, I was super motivated. But now I work full-time and only play around with it when I have time after work, so my drive has slowed down. I can’t keep updating regularly. So if this version doesn’t solve your issue, try using other LoRAs for now. I can only say that, at this stage, I’m sorry I can’t help you guys fix the problem.

I didn’t train it that way, because my original material is only 3–5 seconds long and it always starts with the shot right away. I didn’t include any data that controls the timing of the shot. Since the training material is limited to 3–5 seconds, it’s impossible for me to present too much content within that timeframe. You can try using someone else’s LoRA, or you could combine two different stages of video yourself—one before the shot, and another using my LoRA afterwards. But I don’t know how to do that. At the very least, my LoRA cannot control the timing.

I haven’t run into this issue myself—using the official model has always worked fine. Are you using the official model or one of those AIO mixed ones? I think most people having problems are probably using AIO models.

You can try using the LocalOptima Oral Insertion [WAN 2.2] LoRA. It’s effective for generating an "off-screen male entry" even if the original image doesn't feature a penis. I recommend setting the LoRA weight to 0.7 or 0.5; just combine it with your existing prompts and it should work fine.

Regarding the issue you mentioned, I haven’t run into it myself—using the official model has always worked fine. Are you using the official model or one of those AIO (All-In-One) mixed ones? I think most people having problems are probably using AIO models.

@gman_md My workflow is quite straightforward:

Base on the official model

For high-noise cases, I use Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_v1030_rank_64_bf16

For low-noise cases, I use lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16

If I want more variation, I simply apply other people’s LoRA to adjust the weights

I’ve never encountered the issues you mentioned. From what I’ve seen, other examples also work fine. My suggestion is to start with the simplest workflow examples and test step by step to locate the problem.

There was indeed a case about six months ago where someone had strange body distortions. Later it turned out she was using mixed models (like AIO or SMOOTH versions). Since those mixes already embed multiple LoRA, they can easily cause unexpected results. The only way to handle it is to test and adjust gradually.