A question about the training processes for the naflex versions v.s. fixed-resolution versions
#4
by
qijimrc
- opened
Thank you for your great work! After reading the paper, I have a question that confuses me. Did only the NaFlex versions (e.g., siglip2-so400m-patch16-naflex) undergo training with the distillation loss and masked prediction loss as described in Section 2.3? And were the other fixed-resolution versions (like siglip2-so400m-patch16-384, ...) only further trained from the checkpoint of the training described in Section 2.2?
Thank you very much!
The NaFlex checkpoints did not undergo the training stage described in Section 2.3 (we mention it at the end of Section 2.4.2., but it's not very obvious I agree).