Feedback
Since you write that you haven't tested it and want feedback.
I downloaded Q5_K_M high/low noise.
One thing I notice (which probably doesn't have anything to do with this Quantization) is that using a regular workflow, I had a hard time getting various speed lora to work. I tried several different ones, and it always ended up giving error on one of the lora while the other one seemed to work fine.
I then went ahead and built a VACE workflow using WAN Video Wrapper nodes, and in that workflow I had no issues with the loras.
Here the difference between using regular nodes and the WAN video wrapper is very clear.
Here's the workflow I have been using. One note for anyone that want to use this workflow: Torch compile isn't working, even though Triton is installed. Probably because it's now part of Comfy Kitchen, and the nodes are still looking for it at its old installation path. The same goes for Flash Attention (at least if you are using CUDA 13.x and Blackwell GPU), and I have not been able to install Sage Attention for my configuration yet, but setting attention mode to "comfy" works fine.
Workflow:
