V2 Version (Wan 2.2)
November 14 2025 -> Added Wan 2.2 High and Low LORAs.
Some notes:
Cum -> While the Lora is able to generate cum, sometimes even surprisingly well, I have noticed that when speedup loras are disabled, the cum is much less impressive. Additionally, dedicated cum loras might work with this. Haven't really tested.
This seems to work if you play around a bit (adjust for where to land): "A briefly visible liquid streak of cum bursts from the tip of the penis in a sudden line landing on the woman's forehead, streaks of translucent splatters of cum on her forehead cheeks down to chest and pants, staining her clothes."
Dynamics -> You can get quite some interesting scenes, but it appears that getting the main character to move their feet is very difficult. Not sure whether due to lora being overtrained or what, but prompts such as "she moves her feet around the penis" or "she removes her feet from the penis" and similar seem to have no effect. If you have any suggestions on how to improve this, please let me know.
V1 Version (Wan 2.1)
Wan i2v footjob, sockjob, shoejob. Trained on 720p, but works on 480p model as well. Samples should contain the relevant workflow for each.
If you have any questions, please let me know in the comments. Please also share your creations. Thank you!
Sample prompting
She is giving a footjob. she moves her feet up and down rapidly, stroking the penis. White translucent cum comes out of the penis
What Works
Footjobs, etc.
Can do with or without cum
Using causvid lora to speed up, however, if CFG is too low, it might not generate enough movement
What Doesn't Work
Moving feet onto penis if they are not already on the penis. The model will instead generate wireless stroking
Socked or bare feet are often "blocky" - without much flexibility. While it does not morph socks into shoes and vice versa, it also seems to not really be able to move toes / similar
Cum will often be unnaturally white, and will come in very large amounts
Getting cum to properly drip down the penis is difficult but somewhat doable. Getting cum to shoot and land somewhere desired is pretty much impossible
Description
Wan 2.2 LOW
FAQ
Comments (42)
Can we please get a Wan 2.2 T2V version? Please
Are you looking for video generation or for image generation? I have a qwen image model trained that I wanted to release next. If there's enough interest, I can look into T2V as well
Generally video, WAN T2Image in NSFW is... frustratingly nasty lol. I was going to try the linked Flux lora on a chroma model to see what itd do. Qwen could be awesome, but its nsfw can be frustrating too. But biggest issues with my qwen workflows is the lack of diversity. All the seeds looked the same
@llhappyll822 fair point. I'll try to train a proper T2V. I trained a model specifically for T2I (with only a few frames during training), but it ended up strongly replicating the dataset faces (overtrained maybe) and making overly thick penis (in some of the sample videos you can probably spot WAN generated images).
For qwen, what I really enjoy is the prompt adherence -> at higher resolutions you can really get it to add little details quite well. The biggest issue for me was the grainy effect it gives almost all images. For NSFW -> qwen really struggles with everything apart from the penis (on the lora I trained at least).
Wow, the Wan 2.2 version is really versatile. Definitely worth making. I love the extra actions that were trained in. Thank you.
Thank you, glad to hear you are liking it! To be honest, I think the versatility is largely from Wan 2.2 itself - the dataset did not change much from Wan 2.1 version. I would be interested in training a smash-cut version next, but it is very difficult to find a large variety of training data for this (where we would see the character's face/full body and then them giving a footjob).
@absampleph The data should be easy to make. You can toss a frame of the training data video into Qwen Edit and tell it to pull out that character into a different situation/clothed, then add that as the first frame to your data. There's also a Cinematic Smash Cut lora for Wan 2.2 that just came out today that would let you input your starting image, and smash cut to them looking normal. Then you can grab that frame and add it to the beginning of the footjob video. OR maybe you (or anyone else) can just combine this lora with the Cinematic Smash Cut lora and get all the smash cut functionality you want.
@Jellai ah so training on synthetic (AI generated) videos. Makes sense. Until now I have been quite unwilling of doing that as I think it can degrade quality, but maybe this is indeed the way to go. Thanks for the suggestion!
@absampleph No, I'm not saying that completely. Just for the starting image, to get the concept across. For the main action, still use the non-generated videos. You're just adding a concept to the very front (that concept being a smash cut), and it's the part that the lora never has to generate itself. That part is always provided by the user via I2V. All the actual actions will be learning from real footage.
But it is important that the starting image is one where you can judge it as being close to the footage, in terms of likeness of the face, and body shape/weight/proportions/etc... Because the thing it will learn beyond the smash cut, is matching the original body/likeness. If your source video is trained at 256x256 or something, that can be fudged a bit more. Also, the footage you're using, I can imagine you have access to a real screenshot of them not performing the action at some point in the video, that you can edit on. I'm just giving suggestions for what to do if that doesn't exist.
Great work!! TV2 VERSION, PLEASE! Thanks!
Thank you! Seems there is quite some demand for a T2V, so will look into it. Out of curiosity - do you find it easier to prompt straight into T2V without making an image first? I typically find this approach too time consuming to get the result I am looking for.
@absampleph jumping in on this convo. I personally prefer to use T2V workflow so I would also very much love a T2V version as well. I realize it's probably not the most efficient/effective way to do it, but it's the way that I have things set up and working well with my system. I probably need to take a different approach but if it's not too much trouble I'd love a T2V version. Thanks!
i wonder how you all making this images. thats my only issue i'm struggle with 🤔
this lora is so great
@p_p Appreciate the feedback! With regards to how I generate images - for these examples I mostly used qwen. I will try to get the qwen footjob model released today, will let you know once it is available.
I tested it and it works like this, but the details still need adjustment. It doesn't need to be in the correct position from the beginning, and the sudden appearance of the pe nis shouldn't be deformed.
scene cuts to low-angle close-up .
The woman lifts her leg, revealing her foot clad in white sneakers, to pose as a shove job for the er ect pe nis of a na ked male at the bottom.
The woman's face is shown throughout the entire video.
try
face from original, same person, identical face
The view zooms out to reveal The woman lifts her leg, revealing her foots clad in white sneakers, Using both hands to support the soles of the feet, perform a penis shose job on the lower half of the naked man's penis at the bottom of the screen.
She moves her feet up and down, stroking the penis. (A squirt of cum bursts from the tip of the penis in a sudden squirt:1.1) landing on the woman's legs and forming a pool of translucent cum on her legs and feet. As that happens, the woman closes her eyes - she appears happily surpised, and smiles with a surprised look.
The woman's face is shown throughout the entire video.
Thanks! Please share your generations in the gallery, would be much appreciated!
@absampleph I've already shared it. Please see my post.
@fctq9169 Thank you! Your samples look good!
Very nice, please post some of the image use in this lora with their prompt
I used qwen for approx 1/2 samples for image generation. I will try to get the model released along with examples.
@diffjj23835 FYI - I have uploaded a qwen lora (see suggested resources on this lora page)
Are all of these using the Flux lora you trained to make the input images? The starting images look a lot higher quality for some reason.
OP specifies some of the images are WAN and Qwen. Honestly some pretty great work
@llhappyll822 Holy shit. Qwen did some of this? I wonder if there's a lora I missed... Thanks.
Yes sorry for the confusion - none of the starting images were generated using the Flux lora. I added the suggested resources when I released the Wan 2.1 Lora. I have trained a Qwen (which works really well) and Wan 2.2 T2V model (which seems to be overtrained), and used those to generate the images. I will try to get the Qwen model released soon and then will look into T2V model retraining.
@absampleph I knew this had to be Qwen if anything. The Flux lora isn't able to do what I'm seeing here. Qwen is a beast with certain subjects, and it's doing really well with this. I can't wait for your lora!
@llhappyll822 @Jellai Qwen lora is released. Hope you enjoy!
@absampleph Legendary results. Thanks for sharing.
Not into footjob but your samples are so damn good and creative that I definitely need to give your LoRA a try. Thanks for this naughty niche training ;)
Thank you, appreciate it! The sample inputs were generated using qwen and wan t2v. Honestly, if you haven't yet tried - I suggest giving qwen a go as the prompt adherence is very strong. The biggest issue though - very poor nudity from what I have seen. Maybe there are loras to fix that though.
This can be used on photos where the feet or body are not in the frame. In a squatting pose, the body will be suspended in mid-air. If you want shoes to be visible, the shoes must be visible in the very first frame. Otherwise, simply following the prompt will result in the hands holding the shoes.
If the starting screen doesn't show feet, depicting someone wearing shoes might show them holding the shoes in their hands. Therefore, to accurately represent feet, use the "foot job" function.
The camera zooms out, revealing a woman raising her legs, exposing her feet. She rubs her feet against the lower half of a naked man's pe nis below the screen, performing a foot job.
She moves her feet up and down, stroking the pe nis. (A squirt of cu m bursts from the tip of the pe nis in a sudden squirt:1.1) landing on the woman's legs and forming a pool of translucent c um on her legs and feet. As that happens, the woman closes her eyes—she appears happily surprised and smiles with a surprised look. The woman's face is shown throughout the entire video.
This can be used on photos where the feet or body are not in the frame. In a squatting pose, the body will be suspended in mid-air. If you want shoes to be visible, the shoes must be visible in the very first frame. Otherwise, simply following the prompt will result in the hands holding the shoes.
i get that this is for foot stuff but the quality of all of the rest is so high that I would love a cumshot/handjob/... lora, you got some magic sauce
Thank you, appreciate it very much! I will look into making other loras as well. My key focus is on foot stuff as I feel like it is underappreciated (very few other creators making such loras).
Great work! May i ask how can i generate the first image that contains a penis below the screen? I wish to input my regular and non-nsfw image. Do I have to rely on another image generation model for optimal results? If so, do you have any suggestions? Thanks a lot!
Thank you!
With regards to first image generation - I use qwen image to generate the images. The lora for qwen is linked as a suggested resource on this page.
I am not entirely sure what exactly you want to achieve by combining 2 images, but if you would want to make a non-footjob picture into a footjob one, you can try either what another commenter suggested here with prompting or alternatively you could also try something like this (but use the qwen fj lora and add relevant prompt to the desired change): https://civitai.com/models/2167203/nsfw-qwen-image-edit-rapid-aio
@absampleph tried it, works like miracle, you are absolutely a lifesavior
Hello! I know this is a different thing but is there any way to modify this such that the woman gives a legjob with the dick between her calves/knees? Sort of the same mechanical motion but a legjob instead? I guess the dumber question is where do we put model requests for the community? Sorry and thanks!
Hi!
The legjob you describe - do you reckon you could find videos of this? The big problem with such positions is that I don't think there's enough material online to get a reasonably large dataset to train. I am training primarily on non-animated videos. If you are looking for animated, there is this creator @icelouse who seems to be doing lots of animated variations. I do not see specifically what you are saying, but might be worth messaging/commenting him.
If you are looking for realism, and have some sample videos (20+ ideally), feel free to message me privately and I can look into captioning + training.
yeah that's a good point - part of the problem is that it's not super common but there definitely are some things out there. i'm definitely looking for non-animated. 20+ would be hard. i could maybe find like...5 good ones. if this is a stretch it's no big deal. when i get some time i can send some links if you want or if it's not worth it then no worries. thanks for responding though!
@james_langon837 Feel free to send me privately a link or the name of the concept (thigh job?) but yeah, 5 videos is probably not enough really.
Details
Files
wan22_i2v_footjob_v2_ab_low.safetensors
Mirrors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
footjob_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan2.2-i2v-low-footjobV2.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
I2V_footjob_LOW.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors
wan22_i2v_footjob_v2_ab_low.safetensors