Third person support

[theamusing]

Tried both the 5b base model and the distilled model, but they can't generate third-person view videos.
The technical report shows qualitative results of the base model, and I'm wondering how these results can be generated using the base model?

[hiSandog]

For third-person results, it would help to clarify whether the released base model expects a different action/camera conditioning format than the demo shown in the technical report. If the report examples were generated with a specific prompt, camera trajectory, or evaluation wrapper, adding that exact config would make this reproducible. Otherwise users may assume the checkpoint cannot do third-person view when the missing piece is actually the control signal.