This is my personal workflow for the JibMixWan 2.2 model used as Text to Image model.
Model: https://civarchive.com/models/1813931/jib-mix-wan
I dislike cluttered workflows and tested quite some time to find the settings that produce the best image quality for my use cases.
In the workflow I use a gguf variant of JibMixWan - you can swap this out for the fp8 or fp16 model which is available to download (using a 'load checkpoint' node instead of the gguf loader). A gguf version is not available for download, but it is simple to create it yourself from the fp16 using the tool and manual in city96s GGUF node package: https://github.com/city96/ComfyUI-GGUF/tree/main/tools
(creating ggufs yourself is a simple skill worth picking up).
Suggestions and links for loras and upscale models are in the notes in the workflow. All missing nodes should be easily installable via the comfyui manager.
Edit: if the Ultimate SD Upscale Node produces images with visible seams or different coloured eyes on a single subject (depends very much on the image) you can reduce the denoise from 0.28 to a safer 0.23 or (with tricky images) to 0.15
Description
You can now easily adjust the steps for all samplers with one node, additional explanations added, USDU denoise set to 0.23 which is a more stable option for different use cases
FAQ
Comments (8)
Definitely working and generating nice results, however I do find it's painfully slow compared to other text 2 image methods. I've used the checkpoint file and the lightx2 lora, 4070rtx on sage attention
are you using the fp16 model? Because that is too big to run fully in your vram on a 4070. Q8 or fp8 should fit (with text encoder on cpu/ram)
omg why is it extremely slow even on 5090?
The upscale part takes time with USDU (that's normal with that node), the rest shouldn't be slow - unless for some reason part of the model is not running on vram, but spills into cpu/ram or shared. I run the text encoder on the cpu (should be set that way in the workflow) so the WAN model can run fully in vram. What model are you using? I use Q8, to make sure nothing spills out of the vram. Don't know if the fp16 might spill over.
On a 5090, at 2 MP resolution, with sageattention, the first two sampler steps should take ~40 seconds each, the Ultimate SD Upscale should be ~3 minutes 40 seconds
4,7It per second 4070tisuper, 29it per second 6000pro. workf;ow need edit for good job.
do you use wan22 high or low lora?
low lora