Frame token budget

[davidluciolu]

Hello!

Thanks for your work!

Could you share the setting of per-frame token for QA MLLM? In my opinion, this influence the performance a lot.

Thanks again!

[exiawsh]

Hello!
@davidluciolu
Thank you so much for your interest in our work!

Regarding the setting of the per-frame tokens for the QA MLLM, you can refer to the specific implementation of each script at this address: VideoITG/scripts/eval_lmms_eval at the main branch under the NVlabs/VideoITG repository.

We didn't make any modifications to these and simply adopted the default parameters provided there! I actually haven't done much exploration on the per-frame token settings myself, so I sincerely apologize that I can't share more insights on how much it influences the performance.

Thanks again for reaching out!