Quantized model

[cocktailpeanut]

It takes 30 minutes to generate a 6 second video on a 4090. And I heard 60 minutes on 3090s.

I think the project will gain significantly more traction if there's an OFFICIAL effort from the authors to provide a quantized version AND code to run those quantized models.

[nitinmukesh]

Quantized model will not help. I am using it and it is very very very slow.

Fits well on 8GB VRAM

but inference is so slow (this is with 20 steps and it need 100 steps to generate decent video)
15%|████████▊ | 3/20 [19:25<1:49:36, 386.84s/it]

With 100 steps it shows 11 hours for 6 sec video