<h2 id="wan-2.2">Wan 2.2</h2>Recommended LoRa weight: 1 See version details for other parameters.<h3 id="i2v">I2V</h3>This version uses the same datasets as the 2.1 version, with added datasets about penises.I don't think the model is perfect yet, but it seems improved compared to the 2.1 version:<ul><li>Better generalisation (in my opinion)</li></ul><ul><li>Better saliva supports (though not on hardcore scenes, I wish to train a version for this)</li><li>Way better subject creation (adding a penis, a girl, a man to an image). You shouldn't have weird heads holding an half-penis in the mouth appearing now... At least not much.</li><li>Penises behaves now more naturally when handled (they were often too static to my taste)</li></ul>Known issues:<ul><li>Nude woman added in the video will very probably have a penis, reducing Low Noise weight could works sometimes, but I'll need to retrain it to differentiate penises &amp; vajayjay (it's an edge case though, so not that important)</li></ul><h3 id="t2v">T2V</h3>This version have an improved datasets compared to I2V, with added datasets about vulvas so we can infer video without supplementary LoRas (and avoid woman with penises like in the I2V). It's an imperfect model to be honest, side views are great, but other use cases can be picky on the prompt (impossible position won't have good amplitude). I can't spend more time on it for the moment, so let's hope it's enough! Though, most of the concept is here, I'm curious to see how it behaves with your prompts, a v2 will maybe be needed depending on your feedback. Known issues:<ul><li>With Kijai's 4-Step Lightning LoRa, the Low Noise model can be pretty opinionated on the woman look, and difficult to control. I recommend to not use the Lightning LoRa on low noise in this kind of use cases.</li><li>In use cases where the woman position is a bit awkward, where it doesn't seems easy to suck on a penis or the prompt is a bit too precise &amp; cluttered, the amplitude won't be very good. It can be forced with prompt like <code>she leans on the penis</code> and increased weight on the High Noise LoRa, but be careful on prompting.</li><li>Prompting need to be explicit since prompt adherence is pretty good, the LoRa won't do the work for you.</li><li>Penises are not always perfect, and you could need a side LoRa to improve their look.</li><li>Anime doesn't have a great support (well Wan is not good at it either so...)</li></ul>Low Noise ? High Noise ?The High Noise LoRa contains most of the deepthroat movement, and basic details. The Low Noise refines the movement and improves overall details of the action (penis, testicules, penis/mouth interaction, saliva, micro-movements, etc). It means that the Low Noise LoRa is not mandatory, especially in use cases where the deepthroat is not done on a penis. So don't forget to try without it! Lightning LoRa ?This LoRa behaves very well with <a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Wan22-Lightning">Kijai's 4-Step Lightning LoRa</a> (CFG: 1/2, Shift: 7/8) I even consider results better with it... But that's maybe because I'm not so good at inference on raw Wan 2.2 yet.All in all, please let me know of any weird quirks about this version, I've tested it quite a lot, but who knows. Enjoy ! (Don't use this to do illegal sh*ts, I condemn them firmly) Trigger words (at the start of the prompt):<pre><code>blowjob, deepthroat; </code></pre>Helping trigger wordsFor POV (not mandatory but can help)<pre><code>blowjob, deepthroat, pov;</code></pre>For upside down deepthroat (not mandatory but can help):<pre><code>blowjob, deepthroat, reverse; </code></pre>For on lap deepthroat (can help, but this use case is under trained so I can't guarantee a good inference)<pre><code>blowjob, deepthroat, on-lap; </code></pre>For 69 (not mandatory, but can help):<pre><code>blowjob, deepthroat, sixnine;</code></pre>About penises, you can have moderate control on it with those keywords:<pre><code>penis, dick, erect, flaccid, the foreskin is pulled back, the foreskin is pulled up, glans is visible, glans is hidden, testicules, ...</code></pre>Helping phrases:<pre><code>a video on a woman performing a deepthroat blowjob. [...]</code></pre><pre><code>A man is moving his penis back and forth in the woman mouth and throat</code></pre><pre><code>[...] the penis is going deep into her mouth and throat</code></pre><pre><code>The woman is bobbing her head back and forth while sucking the penis</code></pre><pre><code>[...] The woman makes the penis wet with translucent saliva 
[...] a bit of foamy translucent saliva accumulates on the penis</code></pre>(careful with the foamy saliva, it can get weird quick)To help trigger throat bulging (still unstable, couldn't make it works correctly yet):<pre><code>Her throat is (prominently|slightly) bulging as the man's penis penetrates her mouth and throat</code></pre>Example prompt:<pre><code>blowjob, deepthroat; a video on a woman with fair skin performing a deepthroat blowjob. She kneeling in front of a man. The man is moving his erect penis back and forth in the woman mouth and throat, while she is sucking the man's penis, the penis is going deep into her mouth and throat. Her throat is prominently bulging as the man's penis penetrates her mouth and throat. She makes the penis wet with translucent saliva.</code></pre><h2 id="wan-2.1">Wan 2.1</h2>Recommended LoRa weight (480p): 1The weight can be reduced to limit amplitude, or limit weird behaviours. But 1 should be fine most of the time.Can be used alone. Enjoy! Trigger words (at the start of the prompt):<pre><code>blowjob, deepthroat; </code></pre>Helping trigger wordsFor POV (not mandatory but can help)<pre><code>blowjob, deepthroat, pov;</code></pre>For upside down deepthroat (not mandatory but can help):<pre><code>blowjob, deepthroat, reverse; </code></pre>For on lap deepthroat (can help, but this use case is under trained so I can't guarantee a good inference)<pre><code>blowjob, deepthroat, on-lap; </code></pre>For 69 (not mandatory, but can help):<pre><code>blowjob, deepthroat, sixnine;</code></pre>Helping phrases:<pre><code>a video on a woman performing a deepthroat blowjob. [...]</code></pre><pre><code>A man is moving his penis back and forth in the woman mouth and throat</code></pre><pre><code>[...] the penis is going deep into her mouth and throat</code></pre><pre><code>The woman is bobbing her head back and forth while sucking the penis</code></pre>To help trigger throat bulging (instable at the moment, need more work):<pre><code>Her throat is (prominently|slightly) bulging as the man's penis penetrates her mouth and throat</code></pre>Example prompt:<pre><code>blowjob, deepthroat; a video on a woman with fair skin performing a deepthroat blowjob. She kneeling in front of a man. The man is moving his penis back and forth in the woman mouth and throat, while she is sucking the man's penis, the penis is going deep into her mouth and throat. Her throat is prominently bulging as the man's penis penetrates her mouth and throat.</code></pre>TrainingThis is my first LoRa training, any feedback is welcome. I'm pretty happy with the final result, but there are still details to work on (throat bulging, stability in reverse, handle spit better, ...).Trained on Wan 2.1 480p I2V in fp8 scaled where it triggers the best, it seems to work not so well with 720p I2V (no amplitude, but the concept is still here). Datasets is composed of 60 videos in 480x270 down-sampled from 1920x1080. It was trained over 60 epochs at 2e-4 learning rate, with LoRA+ at 4 on a 5090 (~6h). Network dimension is set to 24. For those that want to known how I trained it, I explain it a bit in this CivitAI <a target="_blank" rel="ugc" href="https://civarchive.com/models/1497390/deepthroat-wan-21-i2v?dialog=commentThread&amp;commentId=784143">discussion</a>. I'm no expert, but it can still help!

Deepthroat, Blowjob - Wan 2.X I2V & T2V

Same training datasets as Wan 2.1 version with add penis datasets. Change log: - Better support for subject creation (penis entering frame, etc) - Better saliva support &amp; related effects - Overall better generalisation with a lot of use cases According to my tests, prefer playing with CFG &amp; Shift than weight to modulate movement.Shift: 5~8CFG: 3~4 (1~2 with Lightning LoRa)

Files with SHA256: d66fde0fc9548a4ee217201c40285d50a3a0a1c32d8d76de86be1337aebac8d1