Issues about the global_ctx and local_ctx in the ContextNet

[ShileiCao]

Thanks for your great work! I am wondering what the difference between global_ctx and local_ctx is and why to distinguish them, especially why the condition is local_ctx when frag_idx = 0 otherwise global_ctx during inference.

[DeminYu98]

Thank you for your interest in our work!

In the diffusion stage, we implemented a segmented generation strategy. The global_ctx includes the temporal features of all $\mu$ seq from the first stage, whereas the local_ctx only contains the temporal features from the current stage. Therefore, when constructing each residual segment $r$, we aimed to consider both the current local features and the global information.