VidTok Input

[Joe-XuZq]

I have just started researching on the video tokenizers and read your paper, I am confusing about the input of VidTok, as in the paper, the standard input is 256*256 with 17 frames, so does this means no matter how many frames the input video has, we just sample 17 frames from the original video? If so, how we know which 17 frames we need to take?
Looking forward for your reply.

[annitang1997]

Hello, in the released version, we are indeed training and testing with a unit of 17 frames. We are currently fine-tuning all causal models to support tokenization and reconstruction of videos of arbitrary length. The relevant models are expected to be released in a few days.

[annitang1997]

Hi, VidTok v1.1 has been released, which supports tokenization and reconstruction of videos of arbitrary length with fine temporal smoothness.

[B0B8K1ng]

Can I use VidTok to tokenize images? Does it support 2 frames input?