What is the number of visual tokens?

[mu-cai]

Hi Authors,

Thanks for bringing this work! I think from your description, the number of tokens is 64 × (N + 1), which should be 64*7=448 visual tokens?

Thanks

[mu-cai]

I just realize that authors use 144 tokens to represent one sub image, according to the code.

[guozonghao96]

In our implementation, we use 144 tokens to represent a image slice. We also use 64 tokens for one image, but there saw a drop compared with the 144 tokens (the efficiency will improve).