After applying this Lora and generating several videos, we found that the beginning of the video was prone to noise.

Delaying or not using the TeaCache Start step can result in noise recovery as the video progresses.

The training videos include several videos generated from Wan2.1.

The sample videos were generated with Wan2_1-I2V-14B-480P_fp8_e4m3fn.

Although noted in the training captions, the direction and speed of the rope movement seem to be random.

The samples were generated with a strength of 1.0, but the correlation between strength and movement is unclear.

It is unclear how this will work when applied to t2v or realistic scenes.