For Wan2.1 Image to Video 480P. Trained only off a small dataset of only pony stuff but happens to work quite well for any sort of 2D/2.5D character. Keyword is PONYBJ. A decent prompt is something like: PONYBJ, a penis is prominently displayed. a (woman / pony / anthropomorphic dragon, ect...) is bobbing her head up and down sucking on the penis, the penis is going deep into her mouth. Words like "fast, energetically, vigorously" and such can increase the amount of motion I found. 384 x 384, 512 x 512, 480 x 832, 832 x 480 all seem to work well. Simple workflow: <a target="_blank" rel="ugc" href="https://files.catbox.moe/l57eeu.json">https://files.catbox.moe/l57eeu.json</a> You will need kijai's nodes: <a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-WanVideoWrapper">https://github.com/kijai/ComfyUI-WanVideoWrapper</a> Its highly recommended to use layer skip with wan for best results as shown: (<a target="_blank" rel="ugc" href="https://files.catbox.moe/6ne3cq.png">https://files.catbox.moe/6ne3cq.png</a>) <a target="_blank" rel="ugc" href="https://www.reddit.com/r/StableDiffusion/comments/1jac3wm/dramatically_enhance_the_quality_of_wan_21_using/">https://www.reddit.com/r/StableDiffusion/comments/1jac3wm/dramatically_enhance_the_quality_of_wan_21_using/</a> Also, using the wan lora block edit node to disable blocks 21-29 seem to improve results not just for this lora but for every wan lora I have tried. (<a target="_blank" rel="ugc" href="https://files.catbox.moe/ro3kqp.png">https://files.catbox.moe/ro3kqp.png</a>) I'll work on longer clips with a more diverse dataset next. Sticking to 2/2.5D though, plenty of 'realistic' loras already. Note: Sometimes it will mix up what is supposed to be a penis and a tongue and therefore create nightmares. You can sort of avoid it by putting tongue in negative prompt. Does not seem that common a issue though, just the bad rng roll here and there.

Files with SHA256: 0ce34cb27db9cf3d32cc033a8dd60254415cccebad67763cac3a34526e89f45e