Is there any way to speed up the inference

[Coronal-Halo]

Is there any way to speed up the inference except lowering the number of frames? Does reducing video resolution speed up the inference?

[cm-xcju]

I also have the same problem. I used one 4090 and it needs about 3s for only one sample

[Fritskee]

@Coronal-Halo the videos are internally reduced to 224x224 in dimension, so don't expect that to have a (noticeable) impact.

@cm-xcju 3 seconds for 1 video? That's actually not bad at all, since you have to imagine that you're essentially doing 8 times a LLaVA call, except that it's now batched.

[Coronal-Halo]

Is there a way to perform batch inference on a batch of videos?

[Fritskee]

Is there a way to perform batch inference on a batch of videos?

If you use the search bar, you can find this issue: https://github.com/PKU-YuanGroup/Video-LLaVA/issues/40#issuecomment-1982277136