[Question] Why can the image VL model understand the video with repeated content twice?

[serend1p1ty]

Nice work!

According to my understanding, you pass a video in ImageGrid format to Image VL model.

For example.

Fast path.

Frame-1, Frame-2, Frame-3, Frame-4, Frame-5

Slow path.

Frame-2, Frame-4

Then you concat them, got

Frame-1, Frame-2, Frame-3, Frame-4, Frame-5, Frame-2, Frame-4

Finally, you feed these tokens to VL model to get result.

Why can the model understand a video where the content is repeated twice?