Possibility to accept video input for vision models on llama.cpp?

[arcseedai]

Is there a possibility ( or interest) in adding video input for vision models on llama.cpp?
I see some amazing models like Qwen 2.5vl that accepts video inputs but I can't try it out, I wonder if implementing this is on the scope of the llama.cpp developers or not

[CISC]

I think there is definite interest, the main issue is dependencies, see the first point of Coding Guidelines. There is some leeway here, we are f.ex. vendoring stb_image for image support, however something like ffmpeg which would be required for video is almost certainly out of the question.

Something like ffmpeg.wasm in the WebUI could be acceptable, but certainly not optimal. @ngxson?

I think this probably boils down to something best left to third parties and not something that should be integrated in llama.cpp.