I'm actually building a tool using Imgui/SDL3/stablediffusioncpp and video models (especially wan2.1) support a feature I would love to add but I don't know much about 'around models' programming. Does someone knows if it's doable with the actual ggml library? If so I may try in a near future.

I'm actually building a tool using Imgui/SDL3/stablediffusioncpp and video models (especially wan2.1) support a feature I would love to add but I don't know much about 'around models' programming. Does someone knows if it's doable with the actual ggml library? If so I may try in a near future.

It's harder than it should be because most video models use 5D tensors at some point, and GGML only supports up to 4D tensors. I don't think it's undoable, but it's definitely tricky to work around this limitation. I gave up implementing LTX because of that.

Thanks for your insight.
Yeah that seem to be a hard things to work around without patching the lib itself which seems also to be pretty hard.