CivArchive
    Masturbation Cumshot for Wan I2V 14B - v1.0
    NSFW

    v1.1 UPDATE!!!

    It's been a while! This was my first Wan lora a few months ago and it still gets a lot of use (thank you all for that!) so I thought it's time to make an update.

    This version is again I2V and improved in all areas, longer training at a higher resolution, improved dataset and captions. You can now somewhat control where the cum lands by adding:

    which lands on her {thighs|pelvis|stomach|face|breasts}.

    to the end of the prompt. Other prompts are the same as the previous version. This update seems to have more realistic cum that lands better on the body or other objects.

    Workflow I used has been added to the metadata of the videos.

    Original Notes

    LoRA for the Wan 14B I2V 480p model, untested for T2V and 720p.

    Diffusion-pipe LoRA trained with 20 videos for 3k steps, 65 epochs. A few of the sample videos have my workflow attached, nothing fancy but I highly recommend the new Skip Layer Guidance for wan, either from KJNodes or WanVideoWrapper depending on if you use native or Kijai.

    Samples are made from images generated in SDXL then run through wan, no upscale or interpolation.

    Example prompt:

    A woman holding her large erect penis. She is sitting on a couch naked. She moves her hand quickly up and down the shaft of her penis and she ejaculates white sticky liquid from her penis.

    Description

    Initial version

    FAQ

    Comments (23)

    Degenerator123Mar 18, 2025· 9 reactions
    CivitAI

    Works on T2V too!

    e: After playing around with it for a bit in T2V, it seems this LoRA is very flexible, it can do handjobs, facials, groups, etc. Still figuring out optimal strength, settings and prompting but this works well!

    crombobularMar 18, 2025· 2 reactions

    can you post some of the videos? am curious how well it works in t2v

    cowgirlfanMay 7, 2025

    hi, did you find any optimal settings for this?

    Degenerator123May 9, 2025

    @cowgirlfan Sorry it's been a while and I haven't used it much for T2V after I made the comment, so I don't have optimal settings to share, I mostly use Wan for I2V now. There are some examples in the gallery here of others using it for T2V though. From what I remember it worked well for the concept but had a bit of influence on the bodytype, you could combine it with another lora that affects bodytype. If you want suggestions for I2V you can check my other comment

    koukousss27234Mar 18, 2025· 4 reactions
    CivitAI

    It is amazing that the liquid movement is so real!

    Bojo66Mar 18, 2025
    CivitAI

    Woah! Amazing!

    Your samples are awesome, but I've yet to produce anything as good with it. What is the skip layer doing? Haven't been able to load it, so can't test. I also notice you use low CFG and high number of steps...is that the magic sauce? Curious for any insight. Would really like to get something as good as your examples.

    definitelynotadog
    Author
    Mar 18, 2025· 3 reactions

    It's hard to explain how skip layer guidance works but it has improved the movement/quality in my gens dramatically, you can read about it here: Add skip layer guidance by AmericanPresidentJimmyCarter · Pull Request #61 · deepbeepmeep/Wan2GP I had to update to comfyui nightly for it to work. I tend to run a minimum of 30 samples, I find CFG between 3-4 works well. I also use ModelSamplingSD3 to increase the shift which helps movement in my experience. I'd say I get about 50% hit rate using my LoRA, it has quite a lot of variety so it can take a few tries to get what you want but I get much better results using the skip layer guidance so I highly recommend it!

    @definitelynotadog 30 samples you mean steps? thanks

    definitelynotadog
    Author
    Mar 18, 2025· 1 reaction

    @logenninefingers888 Oh yeah I mean steps sorry

    @definitelynotadog np, for a minute I thought you meant you were generating 30 vids and picking the best one. Would need to be a patient man!

    neuronsactivatedMar 19, 2025· 3 reactions

    If anyone is curious I followed the trail to reddit and the explanation for the skip layers thing is that it basically makes the unconditioned prompt (or more accurately, the negative prompt part of the cfg) even worse, so when that is eventually removed from the conditioned prompt result.

    Basically, positive prompt is ok. Negative prompt is bad. Negative prompt with missing layers is even worse.

    (ok thing) - (even worse thing) = (even better thing).

    Knowing this, I'd guess this is why clip skip -2 on anime models is so ubiquitous.

    crombobularMar 21, 2025· 1 reaction
    CivitAI

    it follows the motion well but often no liquid ever appears and i'm not sure why.

    definitelynotadog
    Author
    Mar 22, 2025· 1 reaction

    You can try adjusting the prompt to get more liquid, like "lots of" or "huge amounts of" can help a bit. Might also be a workflow issue. I had a decent amount of liquid in most of my tests.

    funscripter627Mar 23, 2025· 3 reactions

    Sometimes lowering the strength helps for some reason.

    crombobularMar 23, 2025· 2 reactions

    @funscripter627 that worked perfectly thank you!

    dsljkg797Mar 22, 2025
    CivitAI

    why my result alway move the caemra to other side ,final show a brand

    crombobularMar 22, 2025· 1 reaction

    bad prompt and bad workflow

    strangephotoMar 24, 2025· 5 reactions
    CivitAI

    Your model is absolutely fantastic. Top notch.

    EndlessDreamOnceHumanApr 12, 2025· 8 reactions
    CivitAI

    Works great on 720p

    swanApr 17, 2025
    CivitAI

    In anime-style,while the jerking motions work well , but liquid effects mostly not good. I'm not sure why. However, things worked much better in realistic style, and results also varied significantly depending on the pose and penis size too.

    Degenerator123May 9, 2025· 9 reactions
    CivitAI

    A combination that seems to work for almost any futa input image using the 480p I2V model: this lora (V1) at 0.7 strength, this at 0.7 (Wan cumshot V1 I2V) and this at 0.8 (Female orgasm, wan_V1) with the following positive prompt (combining prompting for these 3 loras):
    "A woman holding her large erect penis. She moves her hand quickly up and down the shaft of her penis and she ejaculates white sticky liquid from her penis. Cum shoots from the penis. She has an orgasm, her body twitches and shakes from the intense orgasm."
    Input images resized to 624x624 or 416x624 (probably works on other resolutions too), also works for shorter 49 frames (3 seconds). I use a slightly modified version of Kijai's I2V example workflow.
    I use GIMM-VFI (F variant) to interpolate to 48 FPS, works well for the flying liquid.

    Edit: v1.1 works great by itself at str 1, still nice to combine with the female orgasm lora v1 at something like 0.8 str

    cowgirlfanMay 11, 2025

    thanks! any suggestion for CFG level, especially on-site? default is 3.5

    izaaksternMay 14, 2025

    @cowgirlfan For these cumshots you need a lot of steps and I havent found a way to manipulate them onsite. So video quality is mostly a mess