Diffusers Sana Image-to-Video Bad Result

I use the code given in the documentation I got a bad generation result of image-to-video when using custom dataset, is there any bug in the pipeline, and how to fix this?

https://github.com/user-attachments/assets/27dc84dc-b7f7-4dcf-bb0c-e4b1806e45d1

this is the input and prompt that i use: image: prompt: A stylish young woman gracefully spins while crossing a city street. She twirls elegantly with smooth, natural body movement, and her wide white trousers and loose gray blazer flow softly with the motion. Her hair gently sways as she turns, and she keeps a relaxed, confident expression. The camera captures her in a cinematic fashion-editorial style, with the urban street background softly blurred. Natural daylight, realistic movement, elegant spinning motion, stable face, consistent body proportions, smooth motion, high detail.

I also have run it with your sample dataset and it seems not to be as bad as my custom dataset

https://github.com/user-attachments/assets/0ac05295-d684-46e7-876d-19cbd9d400fb

https://github.com/user-attachments/assets/bc5c39eb-d1fa-44ba-898c-ef0e38b04b43

Thankyou