Fast 5-minute-ish video generation workflow for GPU with 12GB VRAM (WAN 2.2 14B GGUF Q4)

Описание: I never bothered to try local video AI, but after seeing all the fuss about WAN 2.2, I decided to give it a try this week, and I certainly having fun with it.

I see other people with 12GB of VRAM or lower struggling with the WAN 2.2 14B model, and I notice they don't use GGUF, other model type is not fit on our VRAM as simple as that.

I found that GGUF for both the model and CLIP, plus the lightning lora from Kijay, and some *unload node*, resulting a fast 5 minute generation time for 4-5 seconds video (49 length), at ~640 pixel, 5 steps in total (2+3).

For your sanity, please try GGUF. Waiting that long without GGUF is not worth it, also GGUF is not that bad imho.

Hardware I use :
RTX 3060 12GB VRAM
32 GB RAM
AMD Ryzen 3600

Workflow :
https://pastebin.com/WNqLtFYe
https://pastebin.com/ykx8f5KU

Models used for this simple potato workflow :
WAN 2.2 High GGUF Q4 - 8.5 GB \models\diffusion_models\
WAN 2.2 Low GGUF Q4 - 8.3 GB \models\diffusion_models\
UMT5 XXL CLIP GGUF Q5 - 4 GB \models\text_encoders\
Kijai's Lightning LoRA for WAN 2.2 High - 600 MB \models\loras\
Kijai's Lightning LoRA for WAN 2.2 Low - 600 MB \models\loras\