Generalized undressing/appearing penises on any realistic human figure. That's the idea at least. Only needs very high strength on base model (to depict the penis better). It works best in mixed models that can already depict penis, like this one especially which was recently removed from civit, I backed it up here: <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/Libidinous_Omega_Wan22mix-v2_backup/tree/main">https://huggingface.co/TenStrip/Libidinous_Omega_Wan22mix-v2_backup/tree/main</a>Quick read for best use:Base trigger: reveals a [small/large] penis and testicles.Clothing removal/Set-up (followed by base trigger):Pulls the front side of the [underwear/shorts/panties] over andPulls down [pants/underwear/panties/shorts] andPulls up [skirt/dress] andSpreads legs andOpens legs while sitting and reveals a tucked hidden penis and testiclesMotion description (down or upward clothing removal):The stiff penis gets caught during the pulling up and flops out.The stiff penis gets caught during the pull-down and flicks up.Angles:From Below.From Side.From Behind. Turns around.Tilts the camera down.Follow-up motions:Touches the penis and moves it around.Grabs the penis and strokes it.Grabs the erect penis and strokes it.Then shakes hips and jiggles the penis side to side.Penis descriptions besides Large/Small (not entirely reliable):The penis has foreskin.The penis is flaccid and small.downward pointing penis.upward curving penis.ErectFloppySoftStiffI'm inexperienced with wan captioning so I assumed being vague would allow for both he/she/it and other descriptors mixed in. I used only these phrases and descriptors throughout the whole dataset. I think editing them with more phrasing works, but you only seem to get the best results when directly copying these broken English phrases directly and they seem to act like long keywords. The real results is with shift, dimension (bigger is better), LTX lora, model, and starting image.This is my first try at 2.2 motion Loras. This version behaves... oddly sometimes? I found that reducing the normal rank64 i2v LightX loras that you usually have at 2.0 or 2.5 down to 1.0 creates much better results. I got worse results from the 1030/1022 LightX loras, the penis usually seemed quite under-detailed with those. The best results come from the merge/trained wan models like smoothmix and such.I sat lurking thinking this would have been around by now. I went on runpod and got it done myself while checking out ai-toolkit in preparation for Zimage stuff. Already identified some mistakes to fix next time like leaving one 10 second clip in (creates jittery motion sometimes), not having enough POV=low knowledge of the top side of the penis, entirely femboy/trans focused (masculine males should still work though), and tbh the 1024x dataset might be overkill for most users. If you want to, you can gen full 1080p close ups of penis with this.If you want more, or can't wait for v2 any amount on here helps and goes straight to training more. I got many more ideas and datasets:<a target="_blank" rel="ugc" href="https://ko-fi.com/tenstrip">https://ko-fi.com/tenstrip</a>This is technically v1.1, the first try with 512x--256x dataset at rank 32 failed to capture any anatomy detail and only motion. This is a chunky rank 64, but it really requires the right motion and detail.

Files with SHA256: fe81f1c4ed091a701bac37e87d1a824f55e7d7b7c927158d640bb7e1a5c74879