Since this is the first time using Wan Lora, the model may not have learned the movements accurately.It’s difficult to reproduce stable movements, and since the success rate is low, the model often relies on random movements. Please refer to the parameters below.<pre><code>High noise Strength 1.3 - 1.5
Low noise Strength 0.2 - 0.5</code></pre>If it doesn’t work well, consider adding the following to the auxiliary prompt.<pre><code>Two beautiful girls move their hips up and down and perform a cooperative pussyjob,
two girls grinding together,
rhythmic hip movement,
looping motion,
continuous movement,</code></pre>Since I only have one source video, I’ve prepared and used 10 videos with staggered start times.The number of training steps was 800, so it might have been a bit too few for training. ChatGPT says that’s sufficient, but the official guidelines suggest 3,000–5,000 is better, so I’m not sure…<hr />初めてのwan loraのため正確に学習できていないかもしれません。安定した動作の再現は難しく、確率が低いのでランダムに頼る場合が多いです。下記パラメータを参考にしてみて下さい。<pre><code>high noise Strength 1.3 - 1.5
low noise Strength 0.2 - 0.5</code></pre>上手く行かない場合、補助のプロンプトに下記を追加することを検討してみてください。<pre><code>Two beautiful girls move their hips up and down and do cooperative pussyjob,
two girls grinding together,
rhythmic hip movement,
looping motion,
continuous movement,</code></pre>学習元動画の手持ちが1個しかないため、開始タイミングをずらした動画を10個用意して使用しています。学習ステップ数は800なので、トレーニングとしては少なかったかもしれません。ChatGPTは十分だと回答しますが、公式では3000 - 5000が良いような言い方もしていて、私にはわかりません…

Cooperative Pussyjob

cooperative_grinding_wan22_v1.1_high_noise.safetensors

Files with SHA256: b4198cb470f23c21e3daa365f8541cc365f73b775b66946b586c52f613114125