Could you provide more details about how NVILA-HD was trained?

[fanmengbo]

Hi, thank you for releasing AutoGaze and NVILA-HD.

I would like to better understand the training procedure used for NVILA-HD. In particular, after integrating AutoGaze into NVILA, did you:

1. retrain the model through all five training stages using the full set of training datasets, or
2. initialize from an existing NVILA checkpoint and only perform the final adaptation stage?
   Could you also provide more details about the NVILA-HD training setup？

These details would be very helpful for reproducing NVILA-HD and for understanding how much additional training is required when integrating AutoGaze into an existing MLLM.

Thank you!