Kijai

/

CogVideoX_GGUF

like 26

Can you post a workflow ?
VRAM usage ?
Thanks

EDIT:
Ok, sorry, just found your post at github:
https://github.com/kijai/ComfyUI-CogVideoXWrapper/tree/main/examples
BTW, it works pretty well with GGUF text encoders (half the size)
For those with low VRAM; play with base resolution and height & width in decode.

Thire is only a workflow for CogVideoX-fun in the repository but we can just replace the load model node to use gguf model in other workflows. I set load_device to offload device to run the model because I only have 8g VRAM. It takes 7.5g VRAM to run the model in the settings above and It's about 10% faster than the fp16 model.

Can you post a workflow ?
VRAM usage ?
Thanks

EDIT:
Ok, sorry, just found your post at github:
https://github.com/kijai/ComfyUI-CogVideoXWrapper/tree/main/examples
BTW, it works pretty well with GGUF text encoders (half the size)
For those with low VRAM; play with base resolution and height & width in decode.

There is no way to lower the sampling resolution to lower than 720x480 right? bummer. I am not waiting an hour to get a video result... problem is cogvideoxfun is not that good.

@Kijai is there a way to sort of hack CogvideoX i2v to run at a lower resolution than 720x480?

problem is cogvideoxfun is not that good.

Agree,
Kling and Hailuoai are far better out of the box.
Maybe cogvideox can get that quality but with a lot of tweaking and I'm too lazy.