Regarding Camera control accuracy.

[KEH0T0]

Hi,

Thanks for the wonderful work.

I am currently trying to reproduce the camera control example provided in the project page (dynamic off-screne memory).
To obtain the camera parameters, I followed your instructions and used the ViPE model to extract pose.npy and intrinsic.npy.

However, the model fails to follow the provided camera trajectories. The generated content does not align with the input poses at all.

Are there any additional preprocessing steps or scaling required to make the camera control functional?
I have strictly followed the documentation, so I'm wondering if I missed any crucial steps. Any guidance would be greatly appreciated.

Also, in my tests (e.g., with a cat), the model tends to keep the subject centered regardless of the input camera trajectory. Is the model’s action-agent training causing it to prioritize object tracking over raw camera poses?

[JingyeChen]

Thanks for your interest. Could you share the source video, pose npz, prompts for us to reproduce the bugs?

[KEH0T0]

Thanks for your quick response.

google_drive_link

[DarcyFan]

Thanks for your quick response.

google_drive_link

Hello, have you found a solution to this problem? Because I also encountered this problem, the generated video doesn‘t match the input’s motion mirror track