The inference of the initial frame relies too heavily on reference images.

[Zhenghao97]

Hello, author, thank you for your outstanding work.

I would like to raise a question: during local deployment, I noticed that in the inference phase, the world model is excessively dependent on the initial reference image. Without this image, the reconstruction results are considerably poor. Doesn't this limitation restrict the applicability of the algorithm itself?

Additionally, given that reference images are finite while layouts are infinite, there may inevitably be cases where the reference images cannot align perfectly with the layouts—such as in road structures, for example. Could this potentially introduce spatial inconsistency issues?

Looking forward to your response!