FUN-Control models have poor quality with zero conditioning

[codeysun]

Thank you for your amazing work!

I noticed that when using the Wan FUN-Control models without control video (i.e., setting control latents to zero tensor), the generation quality is very poor. Even worse than the base Wan model. Here are examples with Wan2.1 1.3b. Top is FUN-Control, bottom is base Wan.

Do you have any insights on why this is the case? When using sparse control videos like openpose, the uncontrolled parts of the video have similar poor quality.

video_1.3b_fun_nocond_tokyo.mp4

1.3b_base_tokyo.mp4