CivArchive
    Worship It - T2V - High
    NSFW

    An attempt to make a lora focused on worshiping big cocks. It was trained on 69 clips (not on purpose, the 70th clip had a hard cut so it was cut from the dataset last minute). Never trained a motion Lora before.

    T2V is trained on Wan 2.2 T2V fp16 model
    I2V is trained on Wan 2.2 I2V fp16model

    I started training the model as an I2V because those lora's always gave me better results even with T2V. Due to difficulties renting a GPU, I decided to try a different service to compare speed and price so I opted to train the T2V on the second service. I figured it it worked well enough I'd throw them up on Civitai.

    I believe the workflow used to generate the samples was created by a former user - playtime_ai
    It utilizes the MoE ksampler.
    I get best results using the following settings:

    Sampler - er_sde
    Scheduler - beta57
    Shift - 8
    Steps - 8
    Lightning - i2v_lightx2v_4step_1022
    Str - 1 on both high and low.
    (Yes, I use the i2v lightning lora for T2V generations. I don't know why, but it always gives more more dynamic and realistic results).

    Trained with the following trigger words:

    cock_worship
    cock_lick
    cock_slap
    blowjob
    deepthroat
    titfuck
    handjob

    Does somethings better than others. Slapping and titfuck it struggles with. Occasionally does some fun body horror.

    Follow up edit: I've added some clarity to the description and will attach a zip file to the T2V Low noise model for all the captions used for Lora training. They all follow the same structure - start with the trigger word(s) featured in the video trained followed by solely focusing on the movement during the video. All videos were 320x320, 81 frames, 16 FPS.

    Generate responsibly.

    Description

    FAQ

    Comments (35)

    VariousDog9932615May 27, 2026
    CivitAI

    are t2v and i2v models the same but just separated for civitai user convenience, or are these two separately trained and specific ones ?

    alternative_penguin
    Author
    May 27, 2026

    I2V is trained on the fp16 I2v model.

    Edit: I initially started training this as an I2V model because I get better quality using those models in T2V generation. I trained the T2V version on a whim at the same time to test Vast AI out. The two I2V samples were just random images from civitai that I quickly threw two prompts at. I don't typically generate with starting images so I cannot speak to it's performance.

    VariousDog9932615May 27, 2026

    @alternative_penguin Okay thx, so they are two distinct models (often people put just two categories but with same models so I was trying to manage space on my ssd's by removing duplicate if needed)

    MMOFanMay 27, 2026· 3 reactions
    CivitAI

    These are some of the most realistic example videos I've ever seen. Can confirm this generates absolutely amazing videos with text to video workflow.

    I would've liked to see some of his sample prompts but I had success with the following:

    She is giving a naked man's penis { One of the keywords | another keywords | etc }

    The background is (whatever you like)

    alternative_penguin
    Author
    May 27, 2026· 2 reactions

    I've attached the captions from the dataset to the t2v - low noise model in the optional drop down box if you'd like to check them out. Updated the model description, but the thought process was to keep them simple and focus on the movement.

    rawkon222May 28, 2026
    CivitAI

    Thanks for the lora! Could you please post your workflow?

    alternative_penguin
    Author
    May 28, 2026· 1 reaction

    All the metadata is saved in the samples. Just download a video and drag and drop it into ComfyUI.

    worked best with a MoE sampler workflow. I tried a few gens with three samplers, but the motion was a little unhinged.

    TeosKuzenMay 28, 2026
    CivitAI

    Does anyone have a good workflow example with this model for generating i2v?

    theshield964657May 28, 2026
    CivitAI

    how did you caption your dataset? i want to train lora too.

    alternative_penguin
    Author
    May 28, 2026· 2 reactions

    Captions are included in the zip file attached to the T2V low noise model.

    flex25tbMay 28, 2026· 1 reaction
    CivitAI

    Damn. Great job, this is a wonderfully effective LoRA. It does what it says it will, and I didn't have to tack on a bunch of other random stuff to make it do so. Thanks for sharing, best addition I have seen on here in a long time.

    flex25tbMay 28, 2026

    Oh also, thanks for doing a T2V, as a lover of the rng dice roll, it's sad to see so many great LoRAs focus only on I2V.

    bgdyamaMay 28, 2026
    CivitAI

    Amazing Lora! Thanks for this. May I ask what resolution and duration did you use for your dataset? any other tips for the training guide?

    rajdrive123May 28, 2026

    Plus one to this. Fabulous lora curious about how this was trained.

    alternative_penguin
    Author
    May 28, 2026· 4 reactions

    Used my MacBook to trim clips down to six seconds and crop them to perfect squares in the Photo’s app. Converted all the clips to MP4 by rename the extension.

    Ran the clips through FFMPEG to:

    1. Convert the resolution to 320x320

    2. Convert the frame rate to 16 FPS

    3. Trim the videos down to 81 frames (this step may be unnecessary, I believe AI toolkit can do this automatically. But it only takes a minute or two for FFMPEG to do it)

    The commands can be found through google, the AI should spit out the code as the first result. For step three the code needs to be changed from c:v to c:a if I remember correctly.

    I included the captions in the T2V as an optional download. Start with trigger word or multiple words. Focus on the movement and describe the trigger word with natural language. Tried to keep it simple. Blonde woman, sultry brunette… etc to vaguely identity the subject. Tried to include multiple synonyms like penis and cock, tits and boobs, etc in the same caption so the Lora should respond to natural language. Sometimes included clothing in the captions, but I don’t think wan needs that information for motion.

    Used AI toolkit. Used a PRO 6000 from runpod.

    Transformer Quant: none

    Low vram: off

    Save every: dataset x 10

    Rank: 32

    Steps: dataset x 100

    Timestep: linear

    Use EMA

    Decay default

    Everything else is default. Turned off sampling.

    Edit the advanced config resolution from 256 to 320

    I believe it took 10 hours.

    Follow up edit: I had made one Lora based on a 256, 81 frame dataset that was a little bigger and few more triggers but at rank 16. It would do the movements well but things would morph together. The second time around I made sure each clip started with everything important was in focus and visible in the initial frame. If the clip was focused on the blowjob trigger it would start with it in the woman’s mouth. Handjob would clearly show the hands at the base or middle with the tip clearly sticking out. So on and so on.. some clips feature multiple triggers. The first trigger is the initial movement, followed by a second (or even third) if applicable.

    rajdrive123May 28, 2026· 1 reaction

    @alternative_penguin Thanks for the detailed response.

    silvereqMay 28, 2026
    CivitAI

    LTX2.3 version please! :) (I can help you train it if u can share your dataset)

    RedditUser981May 29, 2026
    CivitAI

    can i request you to make armpits licking for wan 2.2


    lt_warfMay 29, 2026· 1 reaction
    CivitAI

    this is probably the best t2v BJ Lora I've used (Ive tried many)

    great work, hope you make more!

    AnakondasMay 30, 2026
    CivitAI

    Please show your settings to get the same quality as yours.

    AnakondasMay 30, 2026
    CivitAI

    And how can you get a penis slap on the tongue? What prompt do you use? I can't do it.

    ron01468May 30, 2026· 1 reaction
    CivitAI

    Amazing Lora! Adds exactly the kind of spice to BJ videos I was looking for but didn't quite get from others. Haven't tried t2V but the I2V seems to work pretty well.

    RedditUser981Jun 1, 2026· 1 reaction
    CivitAI

    bro can i request a armpits licking lora there is whole community is waiting for this kind of lora


    ss9999Jun 2, 2026

    One already exists.

    alternative_penguin
    Author
    Jun 3, 2026· 1 reaction

    No interest. To make a Lora like this would cost about 2 grand in my time.

    RedditUser981Jun 3, 2026

    @alternative_penguin ok bro thanks for replying

    KiefstormJun 1, 2026
    CivitAI

    I noticed you're using the I2V light2x LoRa in your T2V workflow. Is there a reason behind that? It seems to be working fine, I'll do a comparison with the t2v version of the light2x lora

    alternative_penguin
    Author
    Jun 1, 2026

    So when generating with dr34mlay and other similar Lora’s I noticed I got better better results using the I2V versions with the 1022 version of the lightx2v Lora. No idea why it works, but the generations always looked more realistic without the AI skin. The 260412 version posted on Kijai’s hugging face would give the best handle motion even better. 1022 would still give good results, but worked better with cumshot Lora’s. So eventually it just became my go-to.

    alternative_penguin
    Author
    Jun 1, 2026

    https://limewire.com/d/VHBiF#erDMPjTBM8 here's is 6 comparison videos. three use dr34mlay v2 t2v with the three different lightx2v loras and three use dr34mlay v2 i2v with three different lightx2v loras. Same settings for everything else.

    Bl1nder5Jun 5, 2026· 1 reaction
    CivitAI

    Awesome (I2V).

    TeosKuzenJun 11, 2026
    CivitAI

    @All

    Guys, what checkpoints or diffusion models do you use for the best and most accurate result? Which one gives you a good balance of consistency, dynamics, speed, and accuracy? Or is this achieved only by using crutches in ComfyUI?

    BlackForestGypsyJun 14, 2026

    try try try and try again, try different iterations with same seed and observe the effect. a good gen is barely anything short of magic.

    whitespider9999Jun 18, 2026· 3 reactions
    CivitAI

    I realize the skill it takes to train these and get a lora from "good" to great. however if you ever feel inspired, a ltx 2.3 version of this would be amazing.

    the_zenithsunJul 1, 2026· 1 reaction
    CivitAI

    Look, the output from this model are the craziest things I over generated. The img2img is INSANE

    613222Jul 6, 2026
    CivitAI

    amazing lora. the preview videos don't do it justice.