An ultimate ComfyUI workflow for Wan2.2
4 steps
GGUF
T2V + I2V
Image Caption
Description
FAQ
Comments (37)
i am having touble with CGControllerNode , its comes up missing when importing the workflow?
same here. updated comfy + all custom nodes, even tried installing 'cg-controller' custom node
Yes, please, @fictiverse - where do we get this missing node?! XD
It is just a ghost node, delete it, i uploaded a fixe, tell me if it is better
fictiverse OK, but FYI I did download the "v1.0 Fix" version and that node is still shown as Missing. I will delete it.
altoiddealer Ok, thanks for the feedback, my ComfyUI must be a bit broken, that must be an old extension i uninstalled.
The T2V workflow works just fine, but when I try to use the I2V workflow with the same models and loras I keep getting this error:
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 21, 104, 60] to have 36 channels, but got 32 channels instead
For I2V video you need I2V models and I2V lightx2v lora
fictiverse Yes I'm using the I2V models and the I2V Lightx2v lora
the wan_2.2_vae.safetensors is not working for the i2v you should use the old wan2.1 vae. for me it solved this error
arman947 I am already using the old wan2.1 vae as well, still getting the error.
arman947 Do you mind sharing your workflow?
If nothing else works, make sure to update your ComfyUI from the ComfyUI manager. I had the same issue and updating ComfyUI itself worked for me.
im having the same issue, updated comfy, and updated via requirments.txt.
BinaryBottleBake try this VAE, https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan2_1_VAE_bf16.safetensors
I've tried everything for the last 24 hours... I can't get it to work, I keep getting the same error message.
dominic1336756 It happens if you have installed sage attention
carlesferrero379 Yes i have sage attention...since many months...Should i desinstall just for wan2.2?
dominic1336756 I just have 2 copies of comfyui one with it and another one without just for trying wan2.2 until someone gives a better solution
same here :(
KSamplerAdvanced
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 7, 90, 64] to have 36 channels, but got 32 channels instead
Update comfyUI to the latest version, I had the same problem and fixed
dominic1336756 I know I hate youtube tutorials as well but trust me on this one. Just watch the installation portion of this. https://www.youtube.com/watch?v=CgLL5aoEX-s
To summarize. It will install an additional seperate comfyui in a new folder and you can retain all your models loras etc and keep doing stuff like you used to. The installation process is also fully automated through a .bat file so it just takes like 30mins or so. Idk what was wrong with my installation either but this is the only thing that worked for me (even using same wan 2.2 files and vae). It works with and without sageattention for me now. If anything fails you can just delete the new comfy ui installation and go back to the old one.
In case you are still facing the issue:
Some nodes will cause this. Some users removed a not much used node flow2-wan-video which was causing the issue and it worked
I FIGURED IT OUT, IT WAS AS SIMLE AS UPDATING COMFY. I thought I had already updated it and didn't realize I was a bit behind.
If I disable image caption, replace the Prompt Strength node with a different float node, and disable both CGControllerNodes, I can get it to run through both KSamplers, but ultimately it never creates a video. I'm assuming the CGControllerNodes are vital for something, but I can't find any package that makes them functional. The image caption just crashes complaining about the cache size being zero if I leave it on.
I'm also noticing that in the Video Final node, there's a red circle around the images input.
I was able to get it to output by manually linking the second ksampler to the vae decode, and then manually linking the Clear Cache All node to the video combine. Not sure why that worked.
I looked in the workflow JSON and didn't see any input or outputs set for the CGControllerNode, so I deleted those nodes to get around the missing node error.
But yes there is something that needs to happen towards the end, around the VRAM clearing. I disabled the film grain and no images were passed to make the video. Enabling it fixed that, but idk why... More tinkering needed.
Why is this workflow so needy and weird acting? invalid prompt: {'type': 'invalid_prompt', 'message': 'Cannot execute because a node is missing the class_type property.', 'details': "Node ID '#12'", 'extra_info': {}} yet I have zero idea what that means.
Apparently, edit the workflow in a text editor and find "CGController" and remove them making sure you keep the JSON format intact. Not sure why they uploaded a workflow in a poor state.
PlushyKitten this worked for me
It say it cant use Loras with the scaled versions. Does somebody know how to deal with that?
ngl my brain bugged because i am new to this. maybe make the workflow more simple to understand xd
i2v very blurred (RTX3060) :/
See a comment I just posted - maybe it will help resolve this for you
Until @fictiverse fixes it: There are 3 CGController nodes that must be deleted. One is a hidden behind the "Seed" node (it's between Power LoRA Loader / Load Image). Un-pin the seed node, move it to the side, delete the CGController node, and put the Seed node back.
This error was causing blurry output:
I noticed this error in CMD: "Return type mismatch between linked nodes: scheduler, ...". Apparently, the scheduler is not relayed correctly from "Diffusion Params" to "Essential Params Expand" and finally to each KSampler.
I resolved this by disconnecting the scheduler input on each KSampler, and just selecting it in each one.
TBH I don't trust the Essential Params Expand node at all