<h2 id="facial-concept-lora-(video-or-wan2.2-5b)">💦 Facial Concept LoRA (Video | WAN2.2 5B)</h2>Hey everyone!This is my fifth video LoRA, trained mostly for facial cumshot animations, but it can also produce full body shots when prompted. This time, I trained on anime-style data, and I was honestly surprised at how well it works on realistic models too.Huge thanks to <a target="_new" rel="ugc" href="https://civarchive.com/user/FLOW0308/videos">FLOW0308</a> and <a target="_new" rel="ugc" href="https://civarchive.com/user/Bonz_ai/videos">Bonz_ai</a> for their amazing videos on Civitai — they made this possible. I didn't want to use realistic cum with realistic dataset because what is the point of it when you can find that on any adult website. <a target="_blank" rel="ugc" href="https://civarchive.com/user/Rvage">Rvage</a> have released amazing WAN2.2 5B workflow with torch compile, block swap and other goodies, check it out: <a target="_blank" rel="ugc" href="https://civarchive.com/models/1676197/wan-video-stdgguf?modelVersionId=2095534">https://civarchive.com/models/1676197/wan-video-stdgguf?modelVersionId=2095534</a><hr /><h3 id="training-details:">🔧 Training Details:</h3><ul><li>Model Target: WAN2.2 5B</li><li>Dataset: 3-second clips (72 frames), anime style, focused on facial shots, occasional body framing</li><li>Angles: Front-facing, POV, slight side</li><li>Cross-compatibility: Works well with realistic styles despite anime dataset</li><li>Resolution: 512x512</li><li>Training Steps: 5000 total (concept was harder to learn than usual)</li><li>Training Time: ~7 hours on an RTX 5090</li><li>VRAM Usage: ~14GB peak</li><li>Workflow: Default ComfyUI video training setup with LoRA loader</li><li>Trigger Word: <code>facial</code></li></ul><hr /><h3 id="notes:">💡 Notes:</h3><ul><li>The concept took more steps to learn compared to other positions or actions, but the extra training paid off.</li><li>Works in both anime and realistic models — better than expected.</li><li>Still room for fine-tuning the fluid physics and face-to-camera alignment.</li><li>Dataset was mostly male masturbating so if your picture doesn't contain a male with own penis grab it can do weird stuff.</li></ul><hr />Hope this helps anyone looking to make specialized erotic video LoRAs! Feedback is welcome — enjoy 🎬💦

Files with SHA256: aecbbf8f323b4fa609359ed274dd3a592eb57a131fa7dd0af7b0c72053bc745a