What's the difference between the HighNoise and LowNoise versions?
#2
by
Phaserblast
- opened
Thanks for posting these. Would you mind explaining the difference between these two noise variants?
Well… from what I have readed. I understand that the workflow is divided into two ksampler:
- High noise would be for the 1st KSampler.
- Low noise would be for the 2nd KSampler.
You can run the first one and generate video, but the models do the following:
- The high-noise focuses on generating overall layout and motion.
- The low-noise refines textures and details.
Here you can see workflow and examples to use it on ComfyUI
https://docs.comfy.org/tutorials/video/wan/wan2_2
Are these choices a result of the quantization? In other words, more bits allocated to the exponent means more precision to the layout and motion whereas more mantissa means finer details (or vice-versa)?
Edit:
I see, both models are needed. The two samplers are connected in series.