Ran Wan2.1 VACE1.3B lora finetuning but have weird result

Hi,

I used lora to finetune the Wan2.1 VACE1.3B model with the script examples/wanvideo/model_training/lora/Wan2.1-VACE-1.3B.sh. I used only one data pair and trained with all default parameters in the training script. The inferred result has a weird flickering effect, could someone give me an insight into what could go wrong?

Input data:

Metadata:

video,prompt,vace_video,vace_video_mask
vid01.mp4,"a photorealistic, cinematic, high-fashion commercial, 360° video of a [V] toy rubber duck as the main product and other toys around it",vid01_vace_video.mp4,vid01_vace_video_mask.mp4

https://github.com/user-attachments/assets/995fb1a7-7905-491b-a24f-b95cb97727b1

https://github.com/user-attachments/assets/eb69efac-84a4-485c-9eba-8ae9e92451f6

https://github.com/user-attachments/assets/3fd453e6-305e-4113-841b-dd1233056698

Inferred result:

https://github.com/user-attachments/assets/e142a0b7-d341-4533-858c-8fca0d85c192

Start by reading examples/wanvideo/model_training/lora/Wan2.1-VACE-1.3B.sh and checking how its defaults interact with the single metadata row and the supplied VACE files. Reproduce the training and inference setup, then document the cause of the flickering result and the smallest verified correction.