This is a high-noise LORA model I trained based on Wan Video 2.2 I2V-A14B. You can connect it to both high-noise and low-noise pipelines simultaneously, with the model weight set to 1.0.
Its function is to realize this specific dance move. You can use it solely for entertainment or in your film and television works.
I trained this model using a 4090 24G graphics card for approximately 6 hours, and this is the version from the 2400th training step. A total of 8 vertical-resolution materials were used in this training, so the effect will be better when using vertical resolution during application. In my tests, it performs well with full-body shots and clothing such as short-sleeved tops and shorts.
The video displayed on the page was generated by me using this LORA on the basis of the quantized model Q8 with 8 steps. If you use a better base model and a higher number of steps, I believe you can generate higher-quality videos.
It should be noted that the performance of the model is not very stable at present, and multiple card draws are required. If you all like it, I will continue to optimize it and launch a better version.
You can use the following text as its trigger word: (HUPOYZM，画面中的女性，以单脚稳稳站立，另一条腿向上抬起并伸直，身体向后倾斜，呈现出富有张力的动态姿势。)
Local use of video models has high requirements for computer configuration. If you want to have a try, you can click the link below—I have turned this LORA into an AI application and hosted it there. By registering through my link, you can get 1,000 points, and you will also receive 100 points every day when you log in. With these points, you can generate many pictures and videos for free!
https://www.runninghub.cn/ai-detail/1960617834315472898/?inviteCode=rh-v1158