<h2 id="concept">Concept</h2><p>A minimax version of my SPH(small penis humiliation) JOI hand gesture lora. Minimax almost gets this right but not quite. This also extends out the capability to objects as well.</p><h2 id="usage">Usage</h2><p>Should work in both T2V and I2V. I would recommend a weight of 0.6-0.7. At higher strengths you might find the audio starts to get very whispery, given much of the source content involved women talking in hushed tones. Minimax requires detailed prompts, but here are a couple of good starter prompts that work well out of the box. You can chain them together to get the effect you want, or let the model figure it out with just the base prompt. Many of the prompt tips from my regular JOI lora will work here too, such as locations and moving between anatomy.</p><p></p><h3 id="one-handed-base-prompt">One handed base prompt</h3><pre><code>she moves her right hand up, thumb and index finger extended and nearly touching, and begins a fast, deliberate up-and-down motion, mimicking the act of stroking a small object.</code></pre><p></p><h3 id="two-handed-base-prompt">Two handed base prompt</h3><pre><code>She positions her right hand’s thumb and index finger gripping the little finger of her left hand. She moves the right hand’s thumb and index fingers up and down along the left little finger in a fast, rhythmic motion, mimicking a handjob gesture.</code></pre><p></p><p>You can also adjust it to something other than a pinky finger. For example:</p><pre><code>She positions her right hand’s thumb and index finger gripping a small plastic dildo she is holding in her left hand. She moves the right hand’s thumb and index fingers up and down along the dildo in a slow, rhythmic motion, mimicking a handjob gesture.</code></pre><p></p><h2 id="training">Training</h2><p>The model was trained on 45 videos, 23 one handed and 22 two handed, all 5 seconds long and 24 fps with audio. It's trained at about 0.6mp and 3 frame buckets for about 2720 steps. Captioning was automated using Qwen3 VL 32B heretic. Training was run using diffusion pipe on 3x3090s.</p>

Files with SHA256: 4852330e8f8caa69ec070ee469ce845929f9d5ed3e379dcb2e784e311811e5a6