This demo is based on Wan 2.1 T2V model and uses Value Sign Flip technique to generate videos with different guidance scales and biases. More on GitHub
Positive prompt should be at least 2 sentence long or the results will be weird.